id stringlengths 40 40 | pid stringlengths 42 42 | input stringlengths 12.1k 56k | output stringlengths 2 232 | document_start_index int64 328 328 | document_end_index int64 12k 55.8k | query_start_index int64 12k 55.8k | query_end_index int64 12.1k 55.9k | truncation_seperator stringclasses 1 value |
|---|---|---|---|---|---|---|---|---|
3fad42be0fb2052bb404b989cc7d58b440cd23a0 | 3fad42be0fb2052bb404b989cc7d58b440cd23a0_0 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
Suppose a user wants to write a sentence “I will be 10 minutes late.” Ideally, she would type just a few keywords such as “10 minutes late” and an autocomplete system would be able to infer the intended sentence (Figure FIGREF1). Existing left-to-right autocomplete systems BIBREF0, BIBREF1 can often be inefficient, as the prefix of a sentence (e.g. “I will be”) fails to capture the core meaning of the sentence. Besides the practical goal of building a better autocomplete system, we are interested in exploring the tradeoffs inherent to such communication schemes between the efficiency of typing keywords, accuracy of reconstruction, and interpretability of keywords.
One approach to learn such schemes is to collect a supervised dataset of keywords-sentence pairs as a training set, but (i) it would be expensive to collect such data from users, and (ii) a static dataset would not capture a real user's natural predilection to adapt to the system BIBREF2. Another approach is to avoid supervision and jointly learn a user-system communication scheme to directly optimize the combination of efficiency and accuracy. However, learning in this way can lead to communication schemes that are uninterpretable to humans BIBREF3, BIBREF4 (see Appendix for additional related work).
In this work, we propose a simple, unsupervised approach to an autocomplete system that is efficient, accurate, and interpretable. For interpretability, we restrict keywords to be subsequences of their source sentences based on the intuition that humans can infer most of the original meaning from a few keywords. We then apply multi-objective optimization approaches to directly control and achieve desirable tradeoffs between efficiency and accuracy.
We observe that naively optimizing a linear combination of efficiency and accuracy terms is unstable and leads to suboptimal schemes. Thus, we propose a new objective which optimizes for communication efficiency under an accuracy constraint. We show this new objective is more stable and efficient than the linear objective at all accuracy levels.
As a proof-of-concept, we build an autocomplete system within this framework which allows a user to write sentences by specifying keywords. We empirically show that our framework produces communication schemes that are 52.16% more accurate than rule-based baselines when specifying 77.37% of sentences, and 11.73% more accurate than a naive, weighted optimization approach when specifying 53.38% of sentences. Finally, we demonstrate that humans can easily adapt to the keyword-based autocomplete system and save nearly 50% of time compared to typing a full sentence in our user study.
Approach
Consider a communication game in which the goal is for a user to communicate a target sequence $x= (x_1, ..., x_m)$ to a system by passing a sequence of keywords $z= (z_1, ..., z_n)$. The user generates keywords $z$ using an encoding strategy $q_{\alpha }(z\mid x)$, and the system attempts to guess the target sequence $x$ via a decoding strategy $p_{\beta }(x\mid z)$.
A good communication scheme $(q_{\alpha }, p_{\beta })$ should be both efficient and accurate. Specifically, we prefer schemes that use fewer keywords (cost), and the target sentence $x$ to be reconstructed with high probability (loss) where
Based on our assumption that humans have an intuitive sense of retaining important keywords, we restrict the set of schemes to be a (potentially noncontiguous) subsequence of the target sentence. Our hypothesis is that such subsequence schemes naturally ensure interpretability, as efficient human and machine communication schemes are both likely to involve keeping important content words.
Approach ::: Modeling with autoencoders.
To learn communication schemes without supervision, we model the cooperative communication between a user and system through an encoder-decoder framework. Concretely, we model the user's encoding strategy $q_{\alpha }(z\mid x)$ with an encoder which encodes the target sentence $x$ into the keywords $z$ by keeping a subset of the tokens. This stochastic encoder $q_{\alpha }(z\mid x)$ is defined by a model which returns the probability of each token retained in the final subsequence $z$. Then, we sample from Bernoulli distributions according to these probabilities to either keep or drop the tokens independently (see Appendix for an example).
We model the autocomplete system's decoding strategy $p_{\beta }(x\mid z)$ as a probabilistic model which conditions on the keywords $z$ and returns a distribution over predictions $x$. We use a standard sequence-to-sequence model with attention and copying for the decoder, but any model architecture can be used (see Appendix for details).
Approach ::: Multi-objective optimization.
Our goal now is to learn encoder-decoder pairs which optimally balance the communication cost and reconstruction loss. The simplest approach to balancing efficiency and accuracy is to weight $\mathrm {cost}(x, \alpha )$ and $\mathrm {loss}(x, \alpha , \beta )$ linearly using a weight $\lambda $ as follows,
where the expectation is taken over the population distribution of source sentences $x$, which is omitted to simplify notation. However, we observe that naively weighting and searching over $\lambda $ is suboptimal and highly unstable—even slight changes to the weighting results in degenerate schemes which keep all or none of its tokens. This instability motivates us to develop a new stable objective.
Our main technical contribution is to draw inspiration from the multi-objective optimization literature and view the tradeoff as a sequence of constrained optimization problems, where we minimize the expected cost subject to varying expected reconstruction error constraints $\epsilon $,
This greatly improves the stability of the training procedure. We empirically observe that the model initially keeps most of the tokens to meet the constraints, and slowly learns to drop uninformative words from the keywords to minimize the cost. Furthermore, $\epsilon $ in Eq (DISPLAY_FORM6) allows us to directly control the maximum reconstruction error of resulting schemes, whereas $\lambda $ in Eq (DISPLAY_FORM5) is not directly related to any of our desiderata.
To optimize the constrained objective, we consider the Lagrangian of Eq (DISPLAY_FORM6),
Much like the objective in Eq (DISPLAY_FORM5) we can compute unbiased gradients by replacing the expectations with their averages over random minibatches. Although gradient descent guarantees convergence on Eq (DISPLAY_FORM7) only when the objective is convex, we find that not only is the optimization stable, the resulting solution achieves better performance than the weighting approach in Eq (DISPLAY_FORM5).
Approach ::: Optimization.
Optimization with respect to $q_{\alpha }(z\mid x)$ is challenging as $z$ is discrete, and thus, we cannot differentiate $\alpha $ through $z$ via the chain rule. Because of this, we use the stochastic REINFORCE estimate BIBREF5 as follows:
We perform joint updates on $(\alpha , \beta , \lambda )$, where $\beta $ and $\lambda $ are updated via standard gradient computations, while $\alpha $ uses an unbiased, stochastic gradient estimate where we approximate the expectation in Eq (DISPLAY_FORM9). We use a single sample from $q_{\alpha }(z\mid x)$ and moving-average of rewards as a baseline to reduce variance.
Experiments
We evaluate our approach by training an autocomplete system on 500K randomly sampled sentences from Yelp reviews BIBREF6 (see Appendix for details). We quantify the efficiency of a communication scheme $(q_{\alpha },p_{\beta })$ by the retention rate of tokens, which is measured as the fraction of tokens that are kept in the keywords. The accuracy of a scheme is measured as the fraction of sentences generated by greedily decoding the model that exactly matches the target sentence.
Experiments ::: Effectiveness of constrained objective.
We first show that the linear objective in Eq (DISPLAY_FORM5) is suboptimal compared to the constrained objective in Eq (DISPLAY_FORM6). Figure FIGREF10 compares the achievable accuracy and efficiency tradeoffs for the two objectives, which shows that the constrained objective results in more efficient schemes than the linear objective at every accuracy level (e.g. 11.73% more accurate at a 53.38% retention rate).
We also observe that the linear objective is highly unstable as a function of the tradeoff parameter $\lambda $ and requires careful tuning. Even slight changes to $\lambda $ results in degenerate schemes that keep all or none of the tokens (e.g. $\lambda \le 4.2$ and $\lambda \ge 4.4$). On the other hand, the constrained objective is substantially more stable as a function of $\epsilon $ (e.g. points for $\epsilon $ are more evenly spaced than $\lambda $).
Experiments ::: Efficiency-accuracy tradeoff.
We quantify the efficiency-accuracy tradeoff compared to two rule-based baselines: Unif and Stopword. The Unif encoder randomly keeps tokens to generate keywords with the probability $\delta $. The Stopword encoder keeps all tokens but drops stop words (e.g. `the', `a', `or') all the time ($\delta =0$) or half of the time ($\delta =0.5$). The corresponding decoders for these encoders are optimized using gradient descent to minimize the reconstruction error (i.e. $\mathrm {loss}(x, \alpha , \beta )$).
Figure FIGREF10 shows that two baselines achieve similar tradeoff curves, while the constrained model achieves a substantial 52.16% improvement in accuracy at a 77.37% retention rate compared to Unif, thereby showing the benefits of jointly training the encoder and decoder.
Experiments ::: Robustness and analysis.
We provide additional experimental results on the robustness of learned communication schemes as well as in-depth analysis on the correlation between the retention rates of tokens and their properties, which we defer to Appendix and for space.
Experiments ::: User study.
We recruited 100 crowdworkers on Amazon Mechanical Turk (AMT) and measured completion times and accuracies for typing randomly sampled sentences from the Yelp corpus. Each user was shown alternating autocomplete and writing tasks across 50 sentences (see Appendix for user interface). For the autocomplete task, we gave users a target sentence and asked them to type a set of keywords into the system. The users were shown the top three suggestions from the autocomplete system, and were asked to mark whether each of these three suggestions was semantically equivalent to the target sentence. For the writing task, we gave users a target sentence and asked them to either type the sentence verbatim or a sentence that preserves the meaning of the target sentence.
Table TABREF13 shows two examples of the autocomplete task and actual user-provided keywords. Each column contains a set of keywords and its corresponding top three suggestions generated by the autocomplete system with beam search. We observe that the system is likely to propose generic sentences for under-specified keywords (left column) and almost the same sentences for over-specified keywords (right column). For properly specified keywords (middle column), the system completes sentences accordingly by adding a verb, adverb, adjective, preposition, capitalization, and punctuation.
Overall, the autocomplete system achieved high accuracy in reconstructing the keywords. Users marked the top suggestion from the autocomplete system to be semantically equivalent to the target $80.6$% of the time, and one of the top 3 was semantically equivalent $90.11$% of the time. The model also achieved a high exact match accuracy of 18.39%. Furthermore, the system was efficient, as users spent $3.86$ seconds typing keywords compared to $5.76$ seconds for full sentences on average. The variance of the typing time was $0.08$ second for keywords and $0.12$ second for full sentences, indicating that choosing and typing keywords for the system did not incur much overhead.
Experiments ::: Acknowledgments
We thank the reviewers and Yunseok Jang for their insightful comments. This work was supported by NSF CAREER Award IIS-1552635 and an Intuit Research Award.
Experiments ::: Reproducibility
All code, data and experiments are available on CodaLab at https://bit.ly/353fbyn.
Question:
What are the baselines used?
Answer:
| Unif and Stopword | 328 | 12,627 | 12,629 | 12,667 | ... [The rest of the article is omitted]
|
3fad42be0fb2052bb404b989cc7d58b440cd23a0 | 3fad42be0fb2052bb404b989cc7d58b440cd23a0_1 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
Suppose a user wants to write a sentence “I will be 10 minutes late.” Ideally, she would type just a few keywords such as “10 minutes late” and an autocomplete system would be able to infer the intended sentence (Figure FIGREF1). Existing left-to-right autocomplete systems BIBREF0, BIBREF1 can often be inefficient, as the prefix of a sentence (e.g. “I will be”) fails to capture the core meaning of the sentence. Besides the practical goal of building a better autocomplete system, we are interested in exploring the tradeoffs inherent to such communication schemes between the efficiency of typing keywords, accuracy of reconstruction, and interpretability of keywords.
One approach to learn such schemes is to collect a supervised dataset of keywords-sentence pairs as a training set, but (i) it would be expensive to collect such data from users, and (ii) a static dataset would not capture a real user's natural predilection to adapt to the system BIBREF2. Another approach is to avoid supervision and jointly learn a user-system communication scheme to directly optimize the combination of efficiency and accuracy. However, learning in this way can lead to communication schemes that are uninterpretable to humans BIBREF3, BIBREF4 (see Appendix for additional related work).
In this work, we propose a simple, unsupervised approach to an autocomplete system that is efficient, accurate, and interpretable. For interpretability, we restrict keywords to be subsequences of their source sentences based on the intuition that humans can infer most of the original meaning from a few keywords. We then apply multi-objective optimization approaches to directly control and achieve desirable tradeoffs between efficiency and accuracy.
We observe that naively optimizing a linear combination of efficiency and accuracy terms is unstable and leads to suboptimal schemes. Thus, we propose a new objective which optimizes for communication efficiency under an accuracy constraint. We show this new objective is more stable and efficient than the linear objective at all accuracy levels.
As a proof-of-concept, we build an autocomplete system within this framework which allows a user to write sentences by specifying keywords. We empirically show that our framework produces communication schemes that are 52.16% more accurate than rule-based baselines when specifying 77.37% of sentences, and 11.73% more accurate than a naive, weighted optimization approach when specifying 53.38% of sentences. Finally, we demonstrate that humans can easily adapt to the keyword-based autocomplete system and save nearly 50% of time compared to typing a full sentence in our user study.
Approach
Consider a communication game in which the goal is for a user to communicate a target sequence $x= (x_1, ..., x_m)$ to a system by passing a sequence of keywords $z= (z_1, ..., z_n)$. The user generates keywords $z$ using an encoding strategy $q_{\alpha }(z\mid x)$, and the system attempts to guess the target sequence $x$ via a decoding strategy $p_{\beta }(x\mid z)$.
A good communication scheme $(q_{\alpha }, p_{\beta })$ should be both efficient and accurate. Specifically, we prefer schemes that use fewer keywords (cost), and the target sentence $x$ to be reconstructed with high probability (loss) where
Based on our assumption that humans have an intuitive sense of retaining important keywords, we restrict the set of schemes to be a (potentially noncontiguous) subsequence of the target sentence. Our hypothesis is that such subsequence schemes naturally ensure interpretability, as efficient human and machine communication schemes are both likely to involve keeping important content words.
Approach ::: Modeling with autoencoders.
To learn communication schemes without supervision, we model the cooperative communication between a user and system through an encoder-decoder framework. Concretely, we model the user's encoding strategy $q_{\alpha }(z\mid x)$ with an encoder which encodes the target sentence $x$ into the keywords $z$ by keeping a subset of the tokens. This stochastic encoder $q_{\alpha }(z\mid x)$ is defined by a model which returns the probability of each token retained in the final subsequence $z$. Then, we sample from Bernoulli distributions according to these probabilities to either keep or drop the tokens independently (see Appendix for an example).
We model the autocomplete system's decoding strategy $p_{\beta }(x\mid z)$ as a probabilistic model which conditions on the keywords $z$ and returns a distribution over predictions $x$. We use a standard sequence-to-sequence model with attention and copying for the decoder, but any model architecture can be used (see Appendix for details).
Approach ::: Multi-objective optimization.
Our goal now is to learn encoder-decoder pairs which optimally balance the communication cost and reconstruction loss. The simplest approach to balancing efficiency and accuracy is to weight $\mathrm {cost}(x, \alpha )$ and $\mathrm {loss}(x, \alpha , \beta )$ linearly using a weight $\lambda $ as follows,
where the expectation is taken over the population distribution of source sentences $x$, which is omitted to simplify notation. However, we observe that naively weighting and searching over $\lambda $ is suboptimal and highly unstable—even slight changes to the weighting results in degenerate schemes which keep all or none of its tokens. This instability motivates us to develop a new stable objective.
Our main technical contribution is to draw inspiration from the multi-objective optimization literature and view the tradeoff as a sequence of constrained optimization problems, where we minimize the expected cost subject to varying expected reconstruction error constraints $\epsilon $,
This greatly improves the stability of the training procedure. We empirically observe that the model initially keeps most of the tokens to meet the constraints, and slowly learns to drop uninformative words from the keywords to minimize the cost. Furthermore, $\epsilon $ in Eq (DISPLAY_FORM6) allows us to directly control the maximum reconstruction error of resulting schemes, whereas $\lambda $ in Eq (DISPLAY_FORM5) is not directly related to any of our desiderata.
To optimize the constrained objective, we consider the Lagrangian of Eq (DISPLAY_FORM6),
Much like the objective in Eq (DISPLAY_FORM5) we can compute unbiased gradients by replacing the expectations with their averages over random minibatches. Although gradient descent guarantees convergence on Eq (DISPLAY_FORM7) only when the objective is convex, we find that not only is the optimization stable, the resulting solution achieves better performance than the weighting approach in Eq (DISPLAY_FORM5).
Approach ::: Optimization.
Optimization with respect to $q_{\alpha }(z\mid x)$ is challenging as $z$ is discrete, and thus, we cannot differentiate $\alpha $ through $z$ via the chain rule. Because of this, we use the stochastic REINFORCE estimate BIBREF5 as follows:
We perform joint updates on $(\alpha , \beta , \lambda )$, where $\beta $ and $\lambda $ are updated via standard gradient computations, while $\alpha $ uses an unbiased, stochastic gradient estimate where we approximate the expectation in Eq (DISPLAY_FORM9). We use a single sample from $q_{\alpha }(z\mid x)$ and moving-average of rewards as a baseline to reduce variance.
Experiments
We evaluate our approach by training an autocomplete system on 500K randomly sampled sentences from Yelp reviews BIBREF6 (see Appendix for details). We quantify the efficiency of a communication scheme $(q_{\alpha },p_{\beta })$ by the retention rate of tokens, which is measured as the fraction of tokens that are kept in the keywords. The accuracy of a scheme is measured as the fraction of sentences generated by greedily decoding the model that exactly matches the target sentence.
Experiments ::: Effectiveness of constrained objective.
We first show that the linear objective in Eq (DISPLAY_FORM5) is suboptimal compared to the constrained objective in Eq (DISPLAY_FORM6). Figure FIGREF10 compares the achievable accuracy and efficiency tradeoffs for the two objectives, which shows that the constrained objective results in more efficient schemes than the linear objective at every accuracy level (e.g. 11.73% more accurate at a 53.38% retention rate).
We also observe that the linear objective is highly unstable as a function of the tradeoff parameter $\lambda $ and requires careful tuning. Even slight changes to $\lambda $ results in degenerate schemes that keep all or none of the tokens (e.g. $\lambda \le 4.2$ and $\lambda \ge 4.4$). On the other hand, the constrained objective is substantially more stable as a function of $\epsilon $ (e.g. points for $\epsilon $ are more evenly spaced than $\lambda $).
Experiments ::: Efficiency-accuracy tradeoff.
We quantify the efficiency-accuracy tradeoff compared to two rule-based baselines: Unif and Stopword. The Unif encoder randomly keeps tokens to generate keywords with the probability $\delta $. The Stopword encoder keeps all tokens but drops stop words (e.g. `the', `a', `or') all the time ($\delta =0$) or half of the time ($\delta =0.5$). The corresponding decoders for these encoders are optimized using gradient descent to minimize the reconstruction error (i.e. $\mathrm {loss}(x, \alpha , \beta )$).
Figure FIGREF10 shows that two baselines achieve similar tradeoff curves, while the constrained model achieves a substantial 52.16% improvement in accuracy at a 77.37% retention rate compared to Unif, thereby showing the benefits of jointly training the encoder and decoder.
Experiments ::: Robustness and analysis.
We provide additional experimental results on the robustness of learned communication schemes as well as in-depth analysis on the correlation between the retention rates of tokens and their properties, which we defer to Appendix and for space.
Experiments ::: User study.
We recruited 100 crowdworkers on Amazon Mechanical Turk (AMT) and measured completion times and accuracies for typing randomly sampled sentences from the Yelp corpus. Each user was shown alternating autocomplete and writing tasks across 50 sentences (see Appendix for user interface). For the autocomplete task, we gave users a target sentence and asked them to type a set of keywords into the system. The users were shown the top three suggestions from the autocomplete system, and were asked to mark whether each of these three suggestions was semantically equivalent to the target sentence. For the writing task, we gave users a target sentence and asked them to either type the sentence verbatim or a sentence that preserves the meaning of the target sentence.
Table TABREF13 shows two examples of the autocomplete task and actual user-provided keywords. Each column contains a set of keywords and its corresponding top three suggestions generated by the autocomplete system with beam search. We observe that the system is likely to propose generic sentences for under-specified keywords (left column) and almost the same sentences for over-specified keywords (right column). For properly specified keywords (middle column), the system completes sentences accordingly by adding a verb, adverb, adjective, preposition, capitalization, and punctuation.
Overall, the autocomplete system achieved high accuracy in reconstructing the keywords. Users marked the top suggestion from the autocomplete system to be semantically equivalent to the target $80.6$% of the time, and one of the top 3 was semantically equivalent $90.11$% of the time. The model also achieved a high exact match accuracy of 18.39%. Furthermore, the system was efficient, as users spent $3.86$ seconds typing keywords compared to $5.76$ seconds for full sentences on average. The variance of the typing time was $0.08$ second for keywords and $0.12$ second for full sentences, indicating that choosing and typing keywords for the system did not incur much overhead.
Experiments ::: Acknowledgments
We thank the reviewers and Yunseok Jang for their insightful comments. This work was supported by NSF CAREER Award IIS-1552635 and an Intuit Research Award.
Experiments ::: Reproducibility
All code, data and experiments are available on CodaLab at https://bit.ly/353fbyn.
Question:
What are the baselines used?
Answer:
| Unif and Stopword | 328 | 12,627 | 12,629 | 12,667 | ... [The rest of the article is omitted]
|
8bf7f1f93d0a2816234d36395ab40c481be9a0e0 | 8bf7f1f93d0a2816234d36395ab40c481be9a0e0_0 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/.
Sentence pair modeling is a fundamental technique underlying many NLP tasks, including the following:
Traditionally, researchers had to develop different methods specific for each task. Now neural networks can perform all the above tasks with the same architecture by training end to end. Various neural models BIBREF1 , BIBREF0 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 have declared state-of-the-art results for sentence pair modeling tasks; however, they were carefully designed and evaluated on selected (often one or two) datasets that can demonstrate the superiority of the model. The research questions are as follows: Do they perform well on other tasks and datasets? How much performance gain is due to certain system design choices and hyperparameter optimizations?
To answer these questions and better understand different network designs, we systematically analyze and compare the state-of-the-art neural models across multiple tasks and multiple domains. Namely, we implement five models and their variations on the same PyTorch platform: InferSent model BIBREF16 , Shortcut-stacked Sentence Encoder Model BIBREF17 , Pairwise Word Interaction Model BIBREF1 , Decomposable Attention Model BIBREF10 , and Enhanced Sequential Inference Model BIBREF0 . They are representative of the two most common approaches: sentence encoding models that learn vector representations of individual sentences and then calculate the semantic relationship between sentences based on vector distance and sentence pair interaction models that use some sorts of word alignment mechanisms (e.g., attention) then aggregate inter-sentence interactions. We focus on identifying important network designs and present a series of findings with quantitative measurements and in-depth analyses, including (i) incorporating inter-sentence interactions is critical; (ii) Tree-LSTM does not help as much as previously claimed but surprisingly improves performance on Twitter data; (iii) Enhanced Sequential Inference Model has the most consistent high performance for larger datasets, while Pairwise Word Interaction Model performs better on smaller datasets and Shortcut-Stacked Sentence Encoder Model is the best performaning model on the Quora corpus. We release our implementations as a toolkit to the research community.
General Framework for Sentence Pair Modeling
Various neural networks have been proposed for sentence pair modeling, all of which fall into two types of approaches. The sentence encoding approach encodes each sentence into a fixed-length vector and then computes sentence similarity directly. The model of this type has advantages in the simplicity of the network design and generalization to other NLP tasks. The sentence pair interaction approach takes word alignment and interactions between the sentence pair into account and often show better performance when trained on in-domain data. Here we outline the two types of neural networks under the same general framework:
Representative Models for Sentence Pair Modeling
Table 1 gives a summary of typical models for sentence pair modeling in recent years. In particular, we investigate five models in depth: two are representative of the sentence encoding type of model, and three are representative of the interaction-aggregation type of model. These models have reported state-or-the-art results with varied architecture design (this section) and implementation details (Section "Implementation Details" ).
The Bi-LSTM Max-pooling Network (InferSent)
We choose the simple Bi-LSTM max-pooling network from InferSent BIBREF16 :
$$&\overleftrightarrow{\mathbf {h}}_{i} = BiLSTM(\mathbf {x}_{i}, \overleftrightarrow{\mathbf {h}}_{i-1}) \\ &\mathbf {v}=max(\overleftrightarrow{\mathbf {h}}_{1}, \overleftrightarrow{\mathbf {h}}_{2}, ..., \overleftrightarrow{\mathbf {h}}_{n})$$ (Eq. 17)
where $\overleftrightarrow{\mathbf {h}}_{i}$ represents the concatenation of hidden states in both directons. It has shown better transfer learning capabilities than several other sentence embedding models, including SkipThought BIBREF31 and FastSent BIBREF32 , when trained on the natural language inference datasets.
The Shortcut-Stacked Sentence Encoder Model (SSE)
The Shortcut-Stacked Sentence Encoder model BIBREF17 is a sentence-based embedding model, which enhances multi-layer Bi-LSTM with skip connection to avoid training error accumulation, and calculates each layer as follows:
$$&\overleftrightarrow{\mathbf {h}}_{i}^{k} = BiLSTM(\mathbf {x}_{i}^{k}, \overleftrightarrow{\mathbf {h}}_{i-1}^{k}) \\ &\mathbf {x}_{i}^{1}=\mathbf {w}_{i} \quad (k=1), \qquad \mathbf {x}_{i}^{k}=[\mathbf {w}_{i}, \overleftrightarrow{\mathbf {h}}_{i}^{k-1}, \overleftrightarrow{\mathbf {h}}_{i}^{k-2}, ..., \overleftrightarrow{\mathbf {h}}_{i}^{1}] \quad (k>1) \\ &\mathbf {v}=max(\overleftrightarrow{\mathbf {h}}_{1}^{m}, \overleftrightarrow{\mathbf {h}}_{2}^{m}, ..., \overleftrightarrow{\mathbf {h}}_{n}^{m})$$ (Eq. 19)
where $\mathbf {x}_{i}^{k}$ is the input of the $k$ th Bi-LSTM layer at time step $i$ , which is the combination of outputs from all previous layers, $\overleftrightarrow{\mathbf {h}}_{i}^{k}$ represents the hidden state of the $k$ th Bi-LSTM layer in both directions. The final sentence embedding $\mathbf {v}$ is the row-based max pooling over the output of the last Bi-LSTM layer, where $n$ denotes the number of words within a sentence and $m$ is the number of Bi-LSTM layers ( $m = 3$ in SSE).
The Pairwise Word Interaction Model (PWIM)
In the Pairwise Word Interaction model BIBREF1 , each word vector $\mathbf {w}_{i}$ is encoded with context through forward and backward LSTMs: $\overrightarrow{\mathbf {h}}_{i} = LSTM^{f}(\mathbf {w}_{i}, \overrightarrow{\mathbf {h}}_{i-1})$ and $\overleftarrow{\mathbf {h}}_{i} = LSTM^{b}(\mathbf {w}_{i}, \overleftarrow{\mathbf {h}}_{i+1})$ . For every word pair $(\mathbf {w}^a_{i}, \mathbf {w}^b_{j})$ across sentences, the model directly calculates word pair interactions using cosine similarity, Euclidean distance, and dot product over the outputs of the encoding layer:
$$D(\overrightarrow{\mathbf {h}}_{i}, \overrightarrow{\mathbf {h}}_{j}) & = [cos(\overrightarrow{\mathbf {h}}_{i}, \overrightarrow{\mathbf {h}}_{j}), \Vert \overrightarrow{\mathbf {h}}_{i} - \overrightarrow{\mathbf {h}}_{j}\Vert , \overrightarrow{\mathbf {h}}_{i} \cdot \overrightarrow{\mathbf {h}}_{j}]$$ (Eq. 21)
The above equation not only applies to forward hidden state $\overrightarrow{\mathbf {h}}_{i}$ and backward hidden state $\overleftarrow{\mathbf {h}}_{i}$ , but also to the concatenation $\overleftrightarrow{\mathbf {h}}_{i}= [\overrightarrow{\mathbf {h}}_{i}, \overleftarrow{\mathbf {h}}_{i}]$ and summation $\mathbf {h}^{+}_{i}= \overrightarrow{\mathbf {h}}_{i} + \overleftarrow{\mathbf {h}}_{i}$ , resulting in a tensor $\mathbf {D}^{13 \times |sent1| \times |sent2|}$ after padding one extra bias term. A “hard” attention is applied to the interaction tensor to build word alignment: selecting the most related word pairs and increasing the corresponding weights by 10 times. Then a 19-layer deep CNN is applied to aggregate the word interaction features for final classification.
The Decomposable Attention Model (DecAtt)
The Decomposable Attention model BIBREF10 is one of the earliest models to introduce attention-based alignment for sentence pair modeling, and it achieved state-of-the-art results on the SNLI dataset with about an order of magnitude fewer parameters than other models (see more in Table 5 ) without relying on word order information. It computes the word pair interaction between $\mathbf {w}_{i}^{a}$ and $\mathbf {w}_{j}^{b}$ (from input sentences $s_a$ and $s_b$ , each with $m$ and $n$ words, respectively) as ${e}_{ij} = {F(\mathbf {w}_{i}^{a})}^{T} F(\mathbf {w}_{j}^{b})$ , where $F$ is a feedforward network; then alignment is determined as follows:
$$&\mathbf {\beta }_{i} = \sum _{j=1}^{n} \frac{exp({e}_{ij})}{\sum _{k=1}^{n} exp({e}_{ik})} \mathbf {w}_{j}^{b} \quad &\mathbf {\alpha }_{j} = \sum _{i=1}^{m} \frac{exp({e}_{ij})}{\sum _{k=1}^{m} exp({e}_{kj})} \mathbf {w}_{i}^{a} $$ (Eq. 23)
where $\mathbf {\beta }_{i}$ is the soft alignment between $\mathbf {w}_{i}^{a}$ and subphrases $\mathbf {w}_{j}^{b}$ in sentence $s_b$ , and vice versa for $\mathbf {\alpha }_{j}$ . The aligned phrases are fed into another feedforward network $G$ : $\mathbf {v}_{i}^{a} = G([\mathbf {w}_{i}^{a}; \mathbf {\beta }_{i}]) $ and $\mathbf {v}_{j}^{b} = G([\mathbf {w}_{j}^{b}; \mathbf {\alpha }_{j}]) $ to generate sets $\lbrace \mathbf {v}_{i}^{a}\rbrace $ and $\lbrace \mathbf {v}_{j}^{b}\rbrace $ , which are aggregated by summation and then concatenated together for classification.
The Enhanced Sequential Inference Model (ESIM)
The Enhanced Sequential Inference Model BIBREF0 is closely related to the DecAtt model, but it differs in a few aspects. First, Chen et al. Chen-Qian:2017:ACL demonstrated that using Bi-LSTM to encode sequential contexts is important for performance improvement. They used the concatenation $\overline{\mathbf {w}}_{i} = \overleftrightarrow{\mathbf {h}}_{i} = [\overrightarrow{\mathbf {h}}_{i}, \overleftarrow{\mathbf {h}}_{i}]$ of both directions as in the PWIM model. The word alignment $\mathbf {\beta }_{i}$ and $\mathbf {\alpha }_{j}$ between $\overline{\mathbf {w}}^{a}$ and $\overline{\mathbf {w}}^{b}$ are calculated the same way as in DecAtt. Second, they showed the competitive performance of recursive architecture with constituency parsing, which complements with sequential LSTM. The feedforward function $G$ in DecAtt is replaced with Tree-LSTM:
$$&\mathbf {v}_{i}^{a} = TreeLSTM([\overline{\mathbf {w}}_{i}^{a}; \mathbf {\beta }_{i}; \overline{\mathbf {w}}_{i}^{a}-\mathbf {\beta }_{i}; \overline{\mathbf {w}}_{i}^{a} \odot \mathbf {\beta }_{i}])\\ &\mathbf {v}_{j}^{b} = TreeLSTM([\overline{\mathbf {w}}_{j}^{b}; \mathbf {\alpha }_{j}; \overline{\mathbf {w}}_{j}^{b}-\mathbf {\alpha }_{j}; \overline{\mathbf {w}}_{j}^{b} \odot \mathbf {\alpha }_{j}])$$ (Eq. 25)
Third, instead of using summation in aggregation, ESIM adapts the average and max pooling and concatenation $\mathbf {v}= [\mathbf {v}_{ave}^{a}; \mathbf {v}_{max}^{a}; \mathbf {v}_{ave}^{b}; \mathbf {v}_{max}^{b}]$ before passing through multi-layer perceptron (MLP) for classification:
$$& \mathbf {v}_{ave}^{a} = \sum _{i=1}^{m} \frac{\mathbf {v}_{i}^{a}}{m}, \qquad \mathbf {v}_{max}^{a} = \max _{i=1}^{m}\mathbf {v}_{i}^{a} , \qquad \mathbf {v}_{ave}^{b} = \sum _{j=1}^{n} \frac{\mathbf {v}_{j}^{b}}{n}, \qquad \mathbf {v}_{max}^{b} = \max _{j=1}^{n}\mathbf {v}_{j}^{b}$$ (Eq. 26)
Datasets
We conducted sentence pair modeling experiments on eight popular datasets: two NLI datasets, three PI datasets, one STS dataset and two QA datasets. Table 2 gives a comparison of these datasets:
SNLI BIBREF7 contains 570k hypotheses written by crowdsourcing workers given the premises. It focuses on three semantic relations: the premise entails the hypothesis (entailment), they contradict each other (contradiction), or they are unrelated (neutral).
Multi-NLI BIBREF33 extends the SNLI corpus to multiple genres of written and spoken texts with 433k sentence pairs.
Quora BIBREF34 contains 400k question pairs collected from the Quora website. This dataset has balanced positive and negative labels indicating whether the questions are duplicated or not.
Twitter-URL BIBREF35 includes 50k sentence pairs collected from tweets that share the same URL of news articles. This dataset contains both formal and informal language.
PIT-2015 BIBREF5 comes from SemEval-2015 and was collected from tweets under the same trending topic. It contains naturally occurred (i.e. written by independent Twitter users spontaneously) paraphrases and non-paraphrases with varied topics and language styles.
STS-2014 BIBREF36 is from SemEval-2014, constructed from image descriptions, news headlines, tweet news, discussion forums, and OntoNotes BIBREF37 .
WikiQA BIBREF8 is an open-domain question-answering dataset. Following He and Lin he-lin:2016:N16-1, questions without correct candidate answer sentences are excluded, and answer sentences are truncated to 40 tokens, resulting in 12k question-answer pairs for our experiments.
TrecQA BIBREF38 is an answer selection task of 56k question-answer pairs and created in Text Retrieval Conferences (TREC). For both WikiQA and TrecQA datasets, the best answer is selected according to the semantic relatedness with the question.
Implementation Details
We implement all the models with the same PyTorch framework. Below, we summarize the implementation details that are key for reproducing results for each model:
SSE: This model can converge very fast, for example, 2 or 3 epochs for the SNLI dataset. We control the convergence speed by updating the learning rate for each epoch: specifically, $lr=\frac{1}{2^{\frac{epoch\_i}{2}}}*{init\_lr}$ , where $init\_lr$ is the initial learning rate and $epoch\_i$ is the index of current epoch.
DecAtt: It is important to use gradient clipping for this model: for each gradient update, we check the L2 norm of all the gradient values, if it is greater than a threshold $b$ , we scale the gradient by a factor $\alpha = b/L2\_norm$ . Another useful procedure is to assemble batches of sentences with similar length.
ESIM: Similar but different from DecAtt, ESIM batches sentences with varied length and uses masks to filter out padding information. In order to batch the parse trees within Tree-LSTM recursion, we follow Bowman et al.'s bowman-EtAl:2016:P16-1 procedure that converts tree structures into the linear sequential structure of a shift reduce parser. Two additional masks are used for producing left and right children of a tree node.
PWIM: The cosine and Euclidean distances used in the word interaction layer have smaller values for similar vectors while dot products have larger values. The performance increases if we add a negative sign to make all the vector similarity measurements behave consistently.
Analysis
Table 3 and 4 show the results reported in the original papers and the replicated results with our implementation. We use accuracy, F1 score, Pearson's $r$ , Mean Average Precision (MAP), and Mean Reciprocal Rank (MRR) for evaluation on different datasets following the literature. Our reproduced results are slightly lower than the original results by 0.5 $\sim $ 1.5 points on accuracy. We suspect the following potential reasons: (i) less extensive hyperparameter tuning for each individual dataset; (ii) only one run with random seeding to report results; and (iii) use of different neural network toolkits: for example, the original ESIM model was implemented with Theano, and PWIM model was in Torch.
Herein, we examine the main components that account for performance in sentence pair modeling.
How important is LSTM encoded context information for sentence pair modeling?
Regarding DecAtt, Parikh et al. parikh-EtAl:2016:EMNLP2016 mentioned that “intra-sentence attention is optional"; they can achieve competitive results without considering context information. However, not surprisingly, our experiments consistently show that encoding sequential context information with LSTM is critical. Compared to DecAtt, ESIM shows better performance on every dataset (see Table 4 and Figure 3 ). The main difference between ESIM and DecAtt that contributes to performance improvement, we found, is the use of Bi-LSTM and Tree-LSTM for sentence encoding, rather than the different choices of aggregation functions.
Why does Tree-LSTM help with Twitter data?
Chen et al. Chen-Qian:2017:ACL offered a simple combination (ESIM $_{seq+tree}$ ) by averaging the prediction probabilities of two ESIM variants that use sequential Bi-LSTM and Tree-LSTM respectively, and suggested “parsing information complements very well with ESIM and further improves the performance". However, we found that adding Tree-LSTM only helps slightly or not at all for most datasets, but it helps noticably with the two Twitter paraphrase datasets. We hypothesize the reason is that these two datasets come from real-world tweets which often contain extraneous text fragments, in contrast to SNLI and other datasets that have sentences written by crowdsourcing workers. For example, the segment “ever wondered ,” in the sentence pair ever wondered , why your recorded #voice sounds weird to you? and why do our recorded voices sound so weird to us? introduces a disruptive context into the Bi-LSTM encoder, while Tree-LSTM can put it in a less important position after constituency parsing.
How important is attentive interaction for sentence pair modeling? Why does SSE excel on Quora?
Both ESIM and DecAtt (Eq. 23 ) calculate an attention-based soft alignment between a sentence pair, which was also proposed in BIBREF27 and BIBREF29 for sentence pair modeling, whereas PWIM utilizes a hard attention mechanism. Both attention strategies are critical for model performance. In PWIM model BIBREF1 , we observed a 1 $\sim $ 2 point performance drop after removing the hard attention, 0 $\sim $ 3 point performance drop and $\sim $ 25% training time reduction after removing the 19-layer CNN aggregation. Likely without even the authors of SSE knowing, the SSE model performs extraordinarily well on the Quora corpus, perhaps because Quora contains many sentence pairs with less complicated inter-sentence interactions (e.g., many identical words in the two sentences) and incorrect ground truth labels (e.g., What is your biggest regret in life? and What's the biggest regret you've had in life? are labeled as non-duplicate questions by mistake).
Figure 3 shows the learning curves. The DecAtt model converges quickly and performs well on large NLI datasets due to its design simplicity. PWIM is the slowest model (see time comparison in Table 5 ) but shows very strong performance on semantic similarity and paraphrase identification datasets. ESIM and SSE keep a good balance between training time and performance.
[3]This number was reported in BIBREF12 by co-authors of DecAtt BIBREF10 . [4]This number was reproduced by Williams et al. williams2017broad. [5]This number was generated by InferSent traind on SNLI and Multi-NLI datasets.
As shown in Figure 4 , we experimented with different training sizes of the largest SNLI dataset. All the models show improved performance as we increase the training size. ESIM and SSE have very similar trends and clearly outperform PWIM on the SNLI dataset. DecAtt shows a performance jump when the training size exceeds a threshold.
We conducted an in-depth analysis of model performance on the Multi-domain NLI dataset based on different categories: text genre, sentence pair overlap, and sentence length. As shown in Table 6 , all models have comparable performance between matched genre and unmatched genre. Sentence length and overlap turn out to be two important factors – the longer the sentences and the fewer tokens in common, the more challenging it is to determine their semantic relationship. These phenomena shared by the state-of-the-art systems reflect their similar design framework which is symmetric at processing both sentences in the pair, while question answering and natural language inference tasks are directional BIBREF30 . How to incorporate asymmetry into model design will be worth more exploration in future research.
In addition to the cross-domain study (Table 6 ), we conducted transfer learning experiments on three paraphrase identification datasets (Table 4 ). The most noteworthy phenomenon is that the SSE model performs better on Twitter-URL and PIT-2015 when trained on the large out-of-domain Quora data than the small in-domain training data. Two likely reasons are: (i) the SSE model with over 29 million parameters is data hungry and (ii) SSE model is a sentence encoding model, which generalizes better across domains/tasks than sentence pair interaction models. Sentence pair interaction models may encounter difficulties on Quora, which contains sentence pairs with the highest word overlap (51.5%) among all datasets and often causes the interaction patterns to focus on a few key words that differ. In contrast, the Twitter-URL dataset has the lowest overlap (23.0%) with a semantic relationship that is mainly based on the intention of the tweets.
Conclusion
We analyzed five different neural models (and their variations) for sentence pair modeling and conducted a series of experiments with eight representative datasets for different NLP tasks. We quantified the importance of the LSTM encoder and attentive alignment for inter-sentence interaction, as well as the transfer learning ability of sentence encoding based models. We showed that the SNLI corpus of over 550k sentence pairs cannot saturate the learning curve. We systematically compared the strengths and weaknesses of different network designs and provided insights for future work.
Acknowledgements
We thank Ohio Supercomputer Center BIBREF39 for computing resources. This work was supported in part by NSF CRII award (RI-1755898) and DARPA through the ARO (W911NF-17-C-0095). The content of the information in this document does not necessarily reflect the position or the policy of the U.S. Government, and no official endorsement should be inferred.
Pretrained Word Embeddings
We used the 200-dimensional GloVe word vectors BIBREF18 , trained on 27 billion words from Twitter (vocabulary size of 1.2 milion words) for Twitter URL BIBREF35 and PIT-2015 BIBREF5 datasets, and the 300-dimensional GloVe vectors, trained on 840 billion words (vocabulary size of 2.2 milion words) from Common Crawl for all other datasets. For out-of-vocabulary words, we initialized the word vectors using normal distribution with mean 0 and deviation 1.
Hyper-parameter Settings
We followed original papers or code implementations to set hyper-parameters for these models. In Infersent model BIBREF16 , the hidden dimension size for Bi-LSTM is 2048, and the fully connected layers have 512 hidden units. In SSE model BIBREF17 , the hidden size for three Bi-LSTMs is 512, 2014 and 2048, respectively. The fully connected layers have 1600 units. PWIM BIBREF1 and ESIM BIBREF0 both use Bi-LSTM for context encoding, having 200 hidden units and 300 hidden units respectively. The DecAtt model BIBREF10 uses three kinds of feed forward networks, all of which have 300 hidden units. Other parameters like learning rate, batch size, dropout rate, and all of them use the same settings as in original papers.
Fine-tuning the Models
It is not practical to fine tune every hyper-parameter in every model and every dataset, since we want to show how these models can generalize well on other datasets, we need try to avoid fine-tuning these parameters on some specific datasets, otherwise we can easily get over-fitted models. Therefore, we keep the hyper-parameters unchanged across different datasets, to demonstrate the generalization capability of each model. The default number of epochs for training these models is set to 20, if some models could converge earlier (no more performance gain on development set), we would stop running them before they approached epoch 20. The 20 epochs can guarantee every model get converged on every dataset.
Question:
Do the authors also analyze transformer-based architectures?
Answer:
| No | 328 | 23,811 | 23,813 | 23,883 | ... [The rest of the article is omitted]
|
8bf7f1f93d0a2816234d36395ab40c481be9a0e0 | 8bf7f1f93d0a2816234d36395ab40c481be9a0e0_1 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/.
Sentence pair modeling is a fundamental technique underlying many NLP tasks, including the following:
Traditionally, researchers had to develop different methods specific for each task. Now neural networks can perform all the above tasks with the same architecture by training end to end. Various neural models BIBREF1 , BIBREF0 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 have declared state-of-the-art results for sentence pair modeling tasks; however, they were carefully designed and evaluated on selected (often one or two) datasets that can demonstrate the superiority of the model. The research questions are as follows: Do they perform well on other tasks and datasets? How much performance gain is due to certain system design choices and hyperparameter optimizations?
To answer these questions and better understand different network designs, we systematically analyze and compare the state-of-the-art neural models across multiple tasks and multiple domains. Namely, we implement five models and their variations on the same PyTorch platform: InferSent model BIBREF16 , Shortcut-stacked Sentence Encoder Model BIBREF17 , Pairwise Word Interaction Model BIBREF1 , Decomposable Attention Model BIBREF10 , and Enhanced Sequential Inference Model BIBREF0 . They are representative of the two most common approaches: sentence encoding models that learn vector representations of individual sentences and then calculate the semantic relationship between sentences based on vector distance and sentence pair interaction models that use some sorts of word alignment mechanisms (e.g., attention) then aggregate inter-sentence interactions. We focus on identifying important network designs and present a series of findings with quantitative measurements and in-depth analyses, including (i) incorporating inter-sentence interactions is critical; (ii) Tree-LSTM does not help as much as previously claimed but surprisingly improves performance on Twitter data; (iii) Enhanced Sequential Inference Model has the most consistent high performance for larger datasets, while Pairwise Word Interaction Model performs better on smaller datasets and Shortcut-Stacked Sentence Encoder Model is the best performaning model on the Quora corpus. We release our implementations as a toolkit to the research community.
General Framework for Sentence Pair Modeling
Various neural networks have been proposed for sentence pair modeling, all of which fall into two types of approaches. The sentence encoding approach encodes each sentence into a fixed-length vector and then computes sentence similarity directly. The model of this type has advantages in the simplicity of the network design and generalization to other NLP tasks. The sentence pair interaction approach takes word alignment and interactions between the sentence pair into account and often show better performance when trained on in-domain data. Here we outline the two types of neural networks under the same general framework:
Representative Models for Sentence Pair Modeling
Table 1 gives a summary of typical models for sentence pair modeling in recent years. In particular, we investigate five models in depth: two are representative of the sentence encoding type of model, and three are representative of the interaction-aggregation type of model. These models have reported state-or-the-art results with varied architecture design (this section) and implementation details (Section "Implementation Details" ).
The Bi-LSTM Max-pooling Network (InferSent)
We choose the simple Bi-LSTM max-pooling network from InferSent BIBREF16 :
$$&\overleftrightarrow{\mathbf {h}}_{i} = BiLSTM(\mathbf {x}_{i}, \overleftrightarrow{\mathbf {h}}_{i-1}) \\ &\mathbf {v}=max(\overleftrightarrow{\mathbf {h}}_{1}, \overleftrightarrow{\mathbf {h}}_{2}, ..., \overleftrightarrow{\mathbf {h}}_{n})$$ (Eq. 17)
where $\overleftrightarrow{\mathbf {h}}_{i}$ represents the concatenation of hidden states in both directons. It has shown better transfer learning capabilities than several other sentence embedding models, including SkipThought BIBREF31 and FastSent BIBREF32 , when trained on the natural language inference datasets.
The Shortcut-Stacked Sentence Encoder Model (SSE)
The Shortcut-Stacked Sentence Encoder model BIBREF17 is a sentence-based embedding model, which enhances multi-layer Bi-LSTM with skip connection to avoid training error accumulation, and calculates each layer as follows:
$$&\overleftrightarrow{\mathbf {h}}_{i}^{k} = BiLSTM(\mathbf {x}_{i}^{k}, \overleftrightarrow{\mathbf {h}}_{i-1}^{k}) \\ &\mathbf {x}_{i}^{1}=\mathbf {w}_{i} \quad (k=1), \qquad \mathbf {x}_{i}^{k}=[\mathbf {w}_{i}, \overleftrightarrow{\mathbf {h}}_{i}^{k-1}, \overleftrightarrow{\mathbf {h}}_{i}^{k-2}, ..., \overleftrightarrow{\mathbf {h}}_{i}^{1}] \quad (k>1) \\ &\mathbf {v}=max(\overleftrightarrow{\mathbf {h}}_{1}^{m}, \overleftrightarrow{\mathbf {h}}_{2}^{m}, ..., \overleftrightarrow{\mathbf {h}}_{n}^{m})$$ (Eq. 19)
where $\mathbf {x}_{i}^{k}$ is the input of the $k$ th Bi-LSTM layer at time step $i$ , which is the combination of outputs from all previous layers, $\overleftrightarrow{\mathbf {h}}_{i}^{k}$ represents the hidden state of the $k$ th Bi-LSTM layer in both directions. The final sentence embedding $\mathbf {v}$ is the row-based max pooling over the output of the last Bi-LSTM layer, where $n$ denotes the number of words within a sentence and $m$ is the number of Bi-LSTM layers ( $m = 3$ in SSE).
The Pairwise Word Interaction Model (PWIM)
In the Pairwise Word Interaction model BIBREF1 , each word vector $\mathbf {w}_{i}$ is encoded with context through forward and backward LSTMs: $\overrightarrow{\mathbf {h}}_{i} = LSTM^{f}(\mathbf {w}_{i}, \overrightarrow{\mathbf {h}}_{i-1})$ and $\overleftarrow{\mathbf {h}}_{i} = LSTM^{b}(\mathbf {w}_{i}, \overleftarrow{\mathbf {h}}_{i+1})$ . For every word pair $(\mathbf {w}^a_{i}, \mathbf {w}^b_{j})$ across sentences, the model directly calculates word pair interactions using cosine similarity, Euclidean distance, and dot product over the outputs of the encoding layer:
$$D(\overrightarrow{\mathbf {h}}_{i}, \overrightarrow{\mathbf {h}}_{j}) & = [cos(\overrightarrow{\mathbf {h}}_{i}, \overrightarrow{\mathbf {h}}_{j}), \Vert \overrightarrow{\mathbf {h}}_{i} - \overrightarrow{\mathbf {h}}_{j}\Vert , \overrightarrow{\mathbf {h}}_{i} \cdot \overrightarrow{\mathbf {h}}_{j}]$$ (Eq. 21)
The above equation not only applies to forward hidden state $\overrightarrow{\mathbf {h}}_{i}$ and backward hidden state $\overleftarrow{\mathbf {h}}_{i}$ , but also to the concatenation $\overleftrightarrow{\mathbf {h}}_{i}= [\overrightarrow{\mathbf {h}}_{i}, \overleftarrow{\mathbf {h}}_{i}]$ and summation $\mathbf {h}^{+}_{i}= \overrightarrow{\mathbf {h}}_{i} + \overleftarrow{\mathbf {h}}_{i}$ , resulting in a tensor $\mathbf {D}^{13 \times |sent1| \times |sent2|}$ after padding one extra bias term. A “hard” attention is applied to the interaction tensor to build word alignment: selecting the most related word pairs and increasing the corresponding weights by 10 times. Then a 19-layer deep CNN is applied to aggregate the word interaction features for final classification.
The Decomposable Attention Model (DecAtt)
The Decomposable Attention model BIBREF10 is one of the earliest models to introduce attention-based alignment for sentence pair modeling, and it achieved state-of-the-art results on the SNLI dataset with about an order of magnitude fewer parameters than other models (see more in Table 5 ) without relying on word order information. It computes the word pair interaction between $\mathbf {w}_{i}^{a}$ and $\mathbf {w}_{j}^{b}$ (from input sentences $s_a$ and $s_b$ , each with $m$ and $n$ words, respectively) as ${e}_{ij} = {F(\mathbf {w}_{i}^{a})}^{T} F(\mathbf {w}_{j}^{b})$ , where $F$ is a feedforward network; then alignment is determined as follows:
$$&\mathbf {\beta }_{i} = \sum _{j=1}^{n} \frac{exp({e}_{ij})}{\sum _{k=1}^{n} exp({e}_{ik})} \mathbf {w}_{j}^{b} \quad &\mathbf {\alpha }_{j} = \sum _{i=1}^{m} \frac{exp({e}_{ij})}{\sum _{k=1}^{m} exp({e}_{kj})} \mathbf {w}_{i}^{a} $$ (Eq. 23)
where $\mathbf {\beta }_{i}$ is the soft alignment between $\mathbf {w}_{i}^{a}$ and subphrases $\mathbf {w}_{j}^{b}$ in sentence $s_b$ , and vice versa for $\mathbf {\alpha }_{j}$ . The aligned phrases are fed into another feedforward network $G$ : $\mathbf {v}_{i}^{a} = G([\mathbf {w}_{i}^{a}; \mathbf {\beta }_{i}]) $ and $\mathbf {v}_{j}^{b} = G([\mathbf {w}_{j}^{b}; \mathbf {\alpha }_{j}]) $ to generate sets $\lbrace \mathbf {v}_{i}^{a}\rbrace $ and $\lbrace \mathbf {v}_{j}^{b}\rbrace $ , which are aggregated by summation and then concatenated together for classification.
The Enhanced Sequential Inference Model (ESIM)
The Enhanced Sequential Inference Model BIBREF0 is closely related to the DecAtt model, but it differs in a few aspects. First, Chen et al. Chen-Qian:2017:ACL demonstrated that using Bi-LSTM to encode sequential contexts is important for performance improvement. They used the concatenation $\overline{\mathbf {w}}_{i} = \overleftrightarrow{\mathbf {h}}_{i} = [\overrightarrow{\mathbf {h}}_{i}, \overleftarrow{\mathbf {h}}_{i}]$ of both directions as in the PWIM model. The word alignment $\mathbf {\beta }_{i}$ and $\mathbf {\alpha }_{j}$ between $\overline{\mathbf {w}}^{a}$ and $\overline{\mathbf {w}}^{b}$ are calculated the same way as in DecAtt. Second, they showed the competitive performance of recursive architecture with constituency parsing, which complements with sequential LSTM. The feedforward function $G$ in DecAtt is replaced with Tree-LSTM:
$$&\mathbf {v}_{i}^{a} = TreeLSTM([\overline{\mathbf {w}}_{i}^{a}; \mathbf {\beta }_{i}; \overline{\mathbf {w}}_{i}^{a}-\mathbf {\beta }_{i}; \overline{\mathbf {w}}_{i}^{a} \odot \mathbf {\beta }_{i}])\\ &\mathbf {v}_{j}^{b} = TreeLSTM([\overline{\mathbf {w}}_{j}^{b}; \mathbf {\alpha }_{j}; \overline{\mathbf {w}}_{j}^{b}-\mathbf {\alpha }_{j}; \overline{\mathbf {w}}_{j}^{b} \odot \mathbf {\alpha }_{j}])$$ (Eq. 25)
Third, instead of using summation in aggregation, ESIM adapts the average and max pooling and concatenation $\mathbf {v}= [\mathbf {v}_{ave}^{a}; \mathbf {v}_{max}^{a}; \mathbf {v}_{ave}^{b}; \mathbf {v}_{max}^{b}]$ before passing through multi-layer perceptron (MLP) for classification:
$$& \mathbf {v}_{ave}^{a} = \sum _{i=1}^{m} \frac{\mathbf {v}_{i}^{a}}{m}, \qquad \mathbf {v}_{max}^{a} = \max _{i=1}^{m}\mathbf {v}_{i}^{a} , \qquad \mathbf {v}_{ave}^{b} = \sum _{j=1}^{n} \frac{\mathbf {v}_{j}^{b}}{n}, \qquad \mathbf {v}_{max}^{b} = \max _{j=1}^{n}\mathbf {v}_{j}^{b}$$ (Eq. 26)
Datasets
We conducted sentence pair modeling experiments on eight popular datasets: two NLI datasets, three PI datasets, one STS dataset and two QA datasets. Table 2 gives a comparison of these datasets:
SNLI BIBREF7 contains 570k hypotheses written by crowdsourcing workers given the premises. It focuses on three semantic relations: the premise entails the hypothesis (entailment), they contradict each other (contradiction), or they are unrelated (neutral).
Multi-NLI BIBREF33 extends the SNLI corpus to multiple genres of written and spoken texts with 433k sentence pairs.
Quora BIBREF34 contains 400k question pairs collected from the Quora website. This dataset has balanced positive and negative labels indicating whether the questions are duplicated or not.
Twitter-URL BIBREF35 includes 50k sentence pairs collected from tweets that share the same URL of news articles. This dataset contains both formal and informal language.
PIT-2015 BIBREF5 comes from SemEval-2015 and was collected from tweets under the same trending topic. It contains naturally occurred (i.e. written by independent Twitter users spontaneously) paraphrases and non-paraphrases with varied topics and language styles.
STS-2014 BIBREF36 is from SemEval-2014, constructed from image descriptions, news headlines, tweet news, discussion forums, and OntoNotes BIBREF37 .
WikiQA BIBREF8 is an open-domain question-answering dataset. Following He and Lin he-lin:2016:N16-1, questions without correct candidate answer sentences are excluded, and answer sentences are truncated to 40 tokens, resulting in 12k question-answer pairs for our experiments.
TrecQA BIBREF38 is an answer selection task of 56k question-answer pairs and created in Text Retrieval Conferences (TREC). For both WikiQA and TrecQA datasets, the best answer is selected according to the semantic relatedness with the question.
Implementation Details
We implement all the models with the same PyTorch framework. Below, we summarize the implementation details that are key for reproducing results for each model:
SSE: This model can converge very fast, for example, 2 or 3 epochs for the SNLI dataset. We control the convergence speed by updating the learning rate for each epoch: specifically, $lr=\frac{1}{2^{\frac{epoch\_i}{2}}}*{init\_lr}$ , where $init\_lr$ is the initial learning rate and $epoch\_i$ is the index of current epoch.
DecAtt: It is important to use gradient clipping for this model: for each gradient update, we check the L2 norm of all the gradient values, if it is greater than a threshold $b$ , we scale the gradient by a factor $\alpha = b/L2\_norm$ . Another useful procedure is to assemble batches of sentences with similar length.
ESIM: Similar but different from DecAtt, ESIM batches sentences with varied length and uses masks to filter out padding information. In order to batch the parse trees within Tree-LSTM recursion, we follow Bowman et al.'s bowman-EtAl:2016:P16-1 procedure that converts tree structures into the linear sequential structure of a shift reduce parser. Two additional masks are used for producing left and right children of a tree node.
PWIM: The cosine and Euclidean distances used in the word interaction layer have smaller values for similar vectors while dot products have larger values. The performance increases if we add a negative sign to make all the vector similarity measurements behave consistently.
Analysis
Table 3 and 4 show the results reported in the original papers and the replicated results with our implementation. We use accuracy, F1 score, Pearson's $r$ , Mean Average Precision (MAP), and Mean Reciprocal Rank (MRR) for evaluation on different datasets following the literature. Our reproduced results are slightly lower than the original results by 0.5 $\sim $ 1.5 points on accuracy. We suspect the following potential reasons: (i) less extensive hyperparameter tuning for each individual dataset; (ii) only one run with random seeding to report results; and (iii) use of different neural network toolkits: for example, the original ESIM model was implemented with Theano, and PWIM model was in Torch.
Herein, we examine the main components that account for performance in sentence pair modeling.
How important is LSTM encoded context information for sentence pair modeling?
Regarding DecAtt, Parikh et al. parikh-EtAl:2016:EMNLP2016 mentioned that “intra-sentence attention is optional"; they can achieve competitive results without considering context information. However, not surprisingly, our experiments consistently show that encoding sequential context information with LSTM is critical. Compared to DecAtt, ESIM shows better performance on every dataset (see Table 4 and Figure 3 ). The main difference between ESIM and DecAtt that contributes to performance improvement, we found, is the use of Bi-LSTM and Tree-LSTM for sentence encoding, rather than the different choices of aggregation functions.
Why does Tree-LSTM help with Twitter data?
Chen et al. Chen-Qian:2017:ACL offered a simple combination (ESIM $_{seq+tree}$ ) by averaging the prediction probabilities of two ESIM variants that use sequential Bi-LSTM and Tree-LSTM respectively, and suggested “parsing information complements very well with ESIM and further improves the performance". However, we found that adding Tree-LSTM only helps slightly or not at all for most datasets, but it helps noticably with the two Twitter paraphrase datasets. We hypothesize the reason is that these two datasets come from real-world tweets which often contain extraneous text fragments, in contrast to SNLI and other datasets that have sentences written by crowdsourcing workers. For example, the segment “ever wondered ,” in the sentence pair ever wondered , why your recorded #voice sounds weird to you? and why do our recorded voices sound so weird to us? introduces a disruptive context into the Bi-LSTM encoder, while Tree-LSTM can put it in a less important position after constituency parsing.
How important is attentive interaction for sentence pair modeling? Why does SSE excel on Quora?
Both ESIM and DecAtt (Eq. 23 ) calculate an attention-based soft alignment between a sentence pair, which was also proposed in BIBREF27 and BIBREF29 for sentence pair modeling, whereas PWIM utilizes a hard attention mechanism. Both attention strategies are critical for model performance. In PWIM model BIBREF1 , we observed a 1 $\sim $ 2 point performance drop after removing the hard attention, 0 $\sim $ 3 point performance drop and $\sim $ 25% training time reduction after removing the 19-layer CNN aggregation. Likely without even the authors of SSE knowing, the SSE model performs extraordinarily well on the Quora corpus, perhaps because Quora contains many sentence pairs with less complicated inter-sentence interactions (e.g., many identical words in the two sentences) and incorrect ground truth labels (e.g., What is your biggest regret in life? and What's the biggest regret you've had in life? are labeled as non-duplicate questions by mistake).
Figure 3 shows the learning curves. The DecAtt model converges quickly and performs well on large NLI datasets due to its design simplicity. PWIM is the slowest model (see time comparison in Table 5 ) but shows very strong performance on semantic similarity and paraphrase identification datasets. ESIM and SSE keep a good balance between training time and performance.
[3]This number was reported in BIBREF12 by co-authors of DecAtt BIBREF10 . [4]This number was reproduced by Williams et al. williams2017broad. [5]This number was generated by InferSent traind on SNLI and Multi-NLI datasets.
As shown in Figure 4 , we experimented with different training sizes of the largest SNLI dataset. All the models show improved performance as we increase the training size. ESIM and SSE have very similar trends and clearly outperform PWIM on the SNLI dataset. DecAtt shows a performance jump when the training size exceeds a threshold.
We conducted an in-depth analysis of model performance on the Multi-domain NLI dataset based on different categories: text genre, sentence pair overlap, and sentence length. As shown in Table 6 , all models have comparable performance between matched genre and unmatched genre. Sentence length and overlap turn out to be two important factors – the longer the sentences and the fewer tokens in common, the more challenging it is to determine their semantic relationship. These phenomena shared by the state-of-the-art systems reflect their similar design framework which is symmetric at processing both sentences in the pair, while question answering and natural language inference tasks are directional BIBREF30 . How to incorporate asymmetry into model design will be worth more exploration in future research.
In addition to the cross-domain study (Table 6 ), we conducted transfer learning experiments on three paraphrase identification datasets (Table 4 ). The most noteworthy phenomenon is that the SSE model performs better on Twitter-URL and PIT-2015 when trained on the large out-of-domain Quora data than the small in-domain training data. Two likely reasons are: (i) the SSE model with over 29 million parameters is data hungry and (ii) SSE model is a sentence encoding model, which generalizes better across domains/tasks than sentence pair interaction models. Sentence pair interaction models may encounter difficulties on Quora, which contains sentence pairs with the highest word overlap (51.5%) among all datasets and often causes the interaction patterns to focus on a few key words that differ. In contrast, the Twitter-URL dataset has the lowest overlap (23.0%) with a semantic relationship that is mainly based on the intention of the tweets.
Conclusion
We analyzed five different neural models (and their variations) for sentence pair modeling and conducted a series of experiments with eight representative datasets for different NLP tasks. We quantified the importance of the LSTM encoder and attentive alignment for inter-sentence interaction, as well as the transfer learning ability of sentence encoding based models. We showed that the SNLI corpus of over 550k sentence pairs cannot saturate the learning curve. We systematically compared the strengths and weaknesses of different network designs and provided insights for future work.
Acknowledgements
We thank Ohio Supercomputer Center BIBREF39 for computing resources. This work was supported in part by NSF CRII award (RI-1755898) and DARPA through the ARO (W911NF-17-C-0095). The content of the information in this document does not necessarily reflect the position or the policy of the U.S. Government, and no official endorsement should be inferred.
Pretrained Word Embeddings
We used the 200-dimensional GloVe word vectors BIBREF18 , trained on 27 billion words from Twitter (vocabulary size of 1.2 milion words) for Twitter URL BIBREF35 and PIT-2015 BIBREF5 datasets, and the 300-dimensional GloVe vectors, trained on 840 billion words (vocabulary size of 2.2 milion words) from Common Crawl for all other datasets. For out-of-vocabulary words, we initialized the word vectors using normal distribution with mean 0 and deviation 1.
Hyper-parameter Settings
We followed original papers or code implementations to set hyper-parameters for these models. In Infersent model BIBREF16 , the hidden dimension size for Bi-LSTM is 2048, and the fully connected layers have 512 hidden units. In SSE model BIBREF17 , the hidden size for three Bi-LSTMs is 512, 2014 and 2048, respectively. The fully connected layers have 1600 units. PWIM BIBREF1 and ESIM BIBREF0 both use Bi-LSTM for context encoding, having 200 hidden units and 300 hidden units respectively. The DecAtt model BIBREF10 uses three kinds of feed forward networks, all of which have 300 hidden units. Other parameters like learning rate, batch size, dropout rate, and all of them use the same settings as in original papers.
Fine-tuning the Models
It is not practical to fine tune every hyper-parameter in every model and every dataset, since we want to show how these models can generalize well on other datasets, we need try to avoid fine-tuning these parameters on some specific datasets, otherwise we can easily get over-fitted models. Therefore, we keep the hyper-parameters unchanged across different datasets, to demonstrate the generalization capability of each model. The default number of epochs for training these models is set to 20, if some models could converge earlier (no more performance gain on development set), we would stop running them before they approached epoch 20. The 20 epochs can guarantee every model get converged on every dataset.
Question:
Do the authors also analyze transformer-based architectures?
Answer:
| No | 328 | 23,811 | 23,813 | 23,883 | ... [The rest of the article is omitted]
|
0f12dc077fe8e5b95ca9163cea1dd17195c96929 | 0f12dc077fe8e5b95ca9163cea1dd17195c96929_0 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
[0]leftmargin=* [0]leftmargin=*
Automatic systems have had a significant and beneficial impact on all walks of human life. So much so that it is easy to overlook their potential to benefit society by promoting equity, diversity, and fairness. For example, machines do not take bribes to do their jobs, they can determine eligibility for a loan without being influenced by the color of the applicant's skin, and they can provide access to information and services without discrimination based on gender or sexual orientation. Nonetheless, as machine learning systems become more human-like in their predictions, they can also perpetuate human biases. Some learned biases may be beneficial for the downstream application (e.g., learning that humans often use some insect names, such as spider or cockroach, to refer to unpleasant situations). Other biases can be inappropriate and result in negative experiences for some groups of people. Examples include, loan eligibility and crime recidivism prediction systems that negatively assess people belonging to a certain pin/zip code (which may disproportionately impact people of a certain race) BIBREF0 and resumé sorting systems that believe that men are more qualified to be programmers than women BIBREF1 . Similarly, sentiment and emotion analysis systems can also perpetuate and accentuate inappropriate human biases, e.g., systems that consider utterances from one race or gender to be less positive simply because of their race or gender, or customer support systems that prioritize a call from an angry male over a call from the equally angry female.
Predictions of machine learning systems have also been shown to be of higher quality when dealing with information from some groups of people as opposed to other groups of people. For example, in the area of computer vision, gender classification systems perform particularly poorly for darker skinned females BIBREF2 . Natural language processing (NLP) systems have been shown to be poor in understanding text produced by people belonging to certain races BIBREF3 , BIBREF4 . For NLP systems, the sources of the bias often include the training data, other corpora, lexicons, and word embeddings that the machine learning algorithm may leverage to build its prediction model.
Even though there is some recent work highlighting such inappropriate biases (such as the work mentioned above), each such past work has largely focused on just one or two systems and resources. Further, there is no benchmark dataset for examining inappropriate biases in natural language systems. In this paper, we describe how we compiled a dataset of 8,640 English sentences carefully chosen to tease out biases towards certain races and genders. We will refer to it as the Equity Evaluation Corpus (EEC). We used the EEC as a supplementary test set in a recent shared task on predicting sentiment and emotion intensity in tweets, SemEval-2018 Task 1: Affect in Tweets BIBREF5 . In particular, we wanted to test a hypothesis that a system should equally rate the intensity of the emotion expressed by two sentences that differ only in the gender/race of a person mentioned. Note that here the term system refers to the combination of a machine learning architecture trained on a labeled dataset, and possibly using additional language resources. The bias can originate from any or several of these parts. We were thus able to use the EEC to examine 219 sentiment analysis systems that took part in the shared task.
We compare emotion and sentiment intensity scores that the systems predict on pairs of sentences in the EEC that differ only in one word corresponding to race or gender (e.g., `This man made me feel angry' vs. `This woman made me feel angry'). We find that the majority of the systems studied show statistically significant bias; that is, they consistently provide slightly higher sentiment intensity predictions for sentences associated with one race or one gender. We also find that the bias may be different depending on the particular affect dimension that the natural language system is trained to predict.
Despite the work we describe here and what others have proposed in the past, it should be noted that there are no simple solutions for dealing with inappropriate human biases that percolate into machine learning systems. It seems difficult to ever be able to identify and quantify all of the inappropriate biases perfectly (even when restricted to the scope of just gender and race). Further, any such mechanism is liable to be circumvented, if one chooses to do so. Nonetheless, as developers of sentiment analysis systems, and NLP systems more broadly, we cannot absolve ourselves of the ethical implications of the systems we build. Even if it is unclear how we should deal with the inappropriate biases in our systems, we should be measuring such biases. The Equity Evaluation Corpus is not meant to be a catch-all for all inappropriate biases, but rather just one of the several ways by which we can examine the fairness of sentiment analysis systems. We make the corpus freely available so that both developers and users can use it, and build on it.
Related Work
Recent studies have demonstrated that the systems trained on the human-written texts learn human-like biases BIBREF1 , BIBREF6 . In general, any predictive model built on historical data may inadvertently inherit human biases based on gender, ethnicity, race, or religion BIBREF7 , BIBREF8 . Discrimination-aware data mining focuses on measuring discrimination in data as well as on evaluating performance of discrimination-aware predictive models BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 .
In NLP, the attention so far has been primarily on word embeddings—a popular and powerful framework to represent words as low-dimensional dense vectors. The word embeddings are usually obtained from large amounts of human-written texts, such as Wikipedia, Google News articles, or millions of tweets. Bias in sentiment analysis systems has only been explored in simple systems that make use of pre-computed word embeddings BIBREF13 . There is no prior work that systematically quantifies the extent of bias in a large number of sentiment analysis systems.
This paper does not examine the differences in accuracies of systems on text produced by different races or genders, as was done by hovy2015demographic,blodgett2016demographic,jurgens2017incorporating,buolamwini2018gender. Approaches on how to mitigate inappropriate biases BIBREF14 , BIBREF1 , BIBREF15 , BIBREF16 , BIBREF13 , BIBREF17 , BIBREF18 are also beyond the scope of this paper. See also the position paper by hovy2016social, which identifies socio-ethical implications of the NLP systems in general.
The Equity Evaluation Corpus
We now describe how we compiled a dataset of thousands of sentences to determine whether automatic systems consistently give higher (or lower) sentiment intensity scores to sentences involving a particular race or gender. There are several ways in which such a dataset may be compiled. We present below the choices that we made.
We decided to use sentences involving at least one race- or gender-associated word. The sentences were intended to be short and grammatically simple. We also wanted some sentences to include expressions of sentiment and emotion, since the goal is to test sentiment and emotion systems. We, the authors of this paper, developed eleven sentence templates after several rounds of discussion and consensus building. They are shown in Table TABREF3 . The templates are divided into two groups. The first type (templates 1–7) includes emotion words. The purpose of this set is to have sentences expressing emotions. The second type (templates 8–11) does not include any emotion words. The purpose of this set is to have non-emotional (neutral) sentences.
The templates include two variables: INLINEFORM0 person INLINEFORM1 and INLINEFORM2 emotion word INLINEFORM3 . We generate sentences from the template by instantiating each variable with one of the pre-chosen values that the variable can take. Each of the eleven templates includes the variable INLINEFORM4 person INLINEFORM5 . INLINEFORM6 person INLINEFORM7 can be instantiated by any of the following noun phrases:
For our study, we chose ten names of each kind from the study by Caliskan:2017 (see Table TABREF7 ). The full lists of noun phrases representing females and males, used in our study, are shown in Table TABREF8 .
The second variable, INLINEFORM0 emotion word INLINEFORM1 , has two variants. Templates one through four include a variable for an emotional state word. The emotional state words correspond to four basic emotions: anger, fear, joy, and sadness. Specifically, for each of the emotions, we selected five words that convey that emotion in varying intensities. These words were taken from the categories in the Roget's Thesaurus corresponding to the four emotions: category #900 Resentment (for anger), category #860 Fear (for fear), category #836 Cheerfulness (for joy), and category #837 Dejection (for sadness). Templates five through seven include emotion words describing a situation or event. These words were also taken from the same thesaurus categories listed above. The full lists of emotion words (emotional state words and emotional situation/event words) are shown in Table TABREF10 .
We generated sentences from the templates by replacing INLINEFORM0 person INLINEFORM1 and INLINEFORM2 emotion word INLINEFORM3 variables with the values they can take. In total, 8,640 sentences were generated with the various combinations of INLINEFORM4 person INLINEFORM5 and INLINEFORM6 emotion word INLINEFORM7 values across the eleven templates. We manually examined the sentences to make sure they were grammatically well-formed. Notably, one can derive pairs of sentences from the EEC such that they differ only in one word corresponding to gender or race (e.g., `My daughter feels devastated' and `My son feels devastated'). We refer to the full set of 8,640 sentences as Equity Evaluation Corpus.
Measuring Race and Gender Bias in Automatic Sentiment Analysis Systems
The race and gender bias evaluation was carried out on the output of the 219 automatic systems that participated in SemEval-2018 Task 1: Affect in Tweets BIBREF5 . The shared task included five subtasks on inferring the affectual state of a person from their tweet: 1. emotion intensity regression, 2. emotion intensity ordinal classification, 3. valence (sentiment) regression, 4. valence ordinal classification, and 5. emotion classification. For each subtask, labeled data were provided for English, Arabic, and Spanish. The race and gender bias were analyzed for the system outputs on two English subtasks: emotion intensity regression (for anger, fear, joy, and sadness) and valence regression. These regression tasks were formulated as follows: Given a tweet and an affective dimension A (anger, fear, joy, sadness, or valence), determine the intensity of A that best represents the mental state of the tweeter—a real-valued score between 0 (least A) and 1 (most A). Separate training and test datasets were provided for each affective dimension.
Training sets included tweets along with gold intensity scores. Two test sets were provided for each task: 1. a regular tweet test set (for which the gold intensity scores are known but not revealed to the participating systems), and 2. the Equity Evaluation Corpus (for which no gold intensity labels exist). Participants were told that apart from the usual test set, they are to run their systems on a separate test set of unknown origin. The participants were instructed to train their system on the tweets training sets provided, and that they could use any other resources they may find or create. They were to run the same final system on the two test sets. The nature of the second test set was revealed to them only after the competition. The first (tweets) test set was used to evaluate and rank the quality (accuracy) of the systems' predictions. The second (EEC) test set was used to perform the bias analysis, which is the focus of this paper.
Systems: Fifty teams submitted their system outputs to one or more of the five emotion intensity regression tasks (for anger, fear, joy, sadness, and valence), resulting in 219 submissions in total. Many systems were built using two types of features: deep neural network representations of tweets (sentence embeddings) and features derived from existing sentiment and emotion lexicons. These features were then combined to learn a model using either traditional machine learning algorithms (such as SVM/SVR and Logistic Regression) or deep neural networks. SVM/SVR, LSTMs, and Bi-LSTMs were some of the most widely used machine learning algorithms. The sentence embeddings were obtained by training a neural network on the provided training data, a distant supervision corpus (e.g., AIT2018 Distant Supervision Corpus that has tweets with emotion-related query terms), sentiment-labeled tweet corpora (e.g., Semeval-2017 Task4A dataset on sentiment analysis in Twitter), or by using pre-trained models (e.g., DeepMoji BIBREF20 , Skip thoughts BIBREF21 ). The lexicon features were often derived from the NRC emotion and sentiment lexicons BIBREF22 , BIBREF23 , BIBREF24 , AFINN BIBREF25 , and Bing Liu Lexicon BIBREF26 .
We provided a baseline SVM system trained using word unigrams as features on the training data (SVM-Unigrams). This system is also included in the current analysis.
Measuring bias: To examine gender bias, we compared each system's predicted scores on the EEC sentence pairs as follows:
Thus, eleven pairs of scores (ten pairs of scores from ten noun phrase pairs and one pair of scores from the averages on name subsets) were examined for each template–emotion word instantiation. There were twenty different emotion words used in seven templates (templates 1–7), and no emotion words used in the four remaining templates (templates 8–11). In total, INLINEFORM0 pairs of scores were compared.
Similarly, to examine race bias, we compared pairs of system predicted scores as follows:
Thus, one pair of scores was examined for each template–emotion word instantiation. In total, INLINEFORM0 pairs of scores were compared.
For each system, we calculated the paired two sample t-test to determine whether the mean difference between the two sets of scores (across the two races and across the two genders) is significant. We set the significance level to 0.05. However, since we performed 438 assessments (219 submissions evaluated for biases in both gender and race), we applied Bonferroni correction. The null hypothesis that the true mean difference between the paired samples was zero was rejected if the calculated p-value fell below INLINEFORM0 .
Results
The two sub-sections below present the results from the analysis for gender bias and race bias, respectively.
Gender Bias Results
Individual submission results were communicated to the participants. Here, we present the summary results across all the teams. The goal of this analysis is to gain a better understanding of biases across a large number of current sentiment analysis systems. Thus, we partition the submissions into three groups according to the bias they show:
F=M not significant: submissions that showed no statistically significant difference in intensity scores predicted for corresponding female and male noun phrase sentences,
F INLINEFORM0 –M INLINEFORM1 significant: submissions that consistently gave higher scores for sentences with female noun phrases than for corresponding sentences with male noun phrases,
F INLINEFORM0 –M INLINEFORM1 significant: submissions that consistently gave lower scores for sentences with female noun phrases than for corresponding sentences with male noun phrases.
For each system and each sentence pair, we calculate the score difference INLINEFORM0 as the score for the female noun phrase sentence minus the score for the corresponding male noun phrase sentence. Table TABREF24 presents the summary results for each of the bias groups. It has the following columns:
#Subm.: number of submissions in each group.
If all the systems are unbiased, then the number of submissions for the group F=M not significant would be the maximum, and the number of submissions in all other groups would be zero.
Avg. score difference F INLINEFORM0 –M INLINEFORM1 : the average INLINEFORM2 for only those pairs where the score for the female noun phrase sentence is higher. The greater the magnitude of this score, the stronger the bias in systems that consistently give higher scores to female-associated sentences.
Avg. score difference F INLINEFORM0 –M INLINEFORM1 : the average INLINEFORM2 for only those pairs where the score for the female noun phrase sentence is lower. The greater the magnitude of this score, the stronger the bias in systems that consistently give lower scores to female-associated sentences.
Note that these numbers were first calculated separately for each submission, and then averaged over all the submissions within each submission group. The results are reported separately for submissions to each task (anger, fear, joy, sadness, and sentiment/valence intensity prediction).
Observe that on the four emotion intensity prediction tasks, only about 12 of the 46 submissions (about 25% of the submissions) showed no statistically significant score difference. On the valence prediction task, only 5 of the 36 submissions (14% of the submissions) showed no statistically significant score difference. Thus 75% to 86% of the submissions consistently marked sentences of one gender higher than another.
When predicting anger, joy, or valence, the number of systems consistently giving higher scores to sentences with female noun phrases (21–25) is markedly higher than the number of systems giving higher scores to sentences with male noun phrases (8–13). (Recall that higher valence means more positive sentiment.) In contrast, on the fear task, most submissions tended to assign higher scores to sentences with male noun phrases (23) as compared to the number of systems giving higher scores to sentences with female noun phrases (12). When predicting sadness, the number of submissions that mostly assigned higher scores to sentences with female noun phrases (18) is close to the number of submissions that mostly assigned higher scores to sentences with male noun phrases (16). These results are in line with some common stereotypes, such as females are more emotional, and situations involving male agents are more fearful BIBREF27 .
Figure FIGREF25 shows the score differences ( INLINEFORM0 ) for individual systems on the valence regression task. Plots for the four emotion intensity prediction tasks are shown in Figure FIGREF31 in the Appendix. Each point ( ▲, ▼, ●) on the plot corresponds to the difference in scores predicted by the system on one sentence pair. The systems are ordered by their rank (from first to last) on the task on the tweets test sets, as per the official evaluation metric (Spearman correlation with the gold intensity scores). We will refer to the difference between the maximal value of INLINEFORM1 and the minimal value of INLINEFORM2 for a particular system as the INLINEFORM3 –spread. Observe that the INLINEFORM4 –spreads for many systems are rather large, up to 0.57. Depending on the task, the top 10 or top 15 systems as well as some of the worst performing systems tend to have smaller INLINEFORM5 –spreads while the systems with medium to low performance show greater sensitivity to the gender-associated words. Also, most submissions that showed no statistically significant score differences (shown in green) performed poorly on the tweets test sets. Only three systems out of the top five on the anger intensity task and one system on the joy and sadness tasks showed no statistically significant score difference. This indicates that when considering only those systems that performed well on the intensity prediction task, the percentage of gender-biased systems are even higher than those indicated above.
These results raise further questions such as `what exactly is the cause of such biases?' and `why is the bias impacted by the emotion task under consideration?'. Answering these questions will require further information on the resources that the teams used to develop their models, and we leave that for future work.
Average score differences: For submissions that showed statistically significant score differences, the average score difference F INLINEFORM0 –M INLINEFORM1 and the average score difference F INLINEFORM2 –M INLINEFORM3 were INLINEFORM4 . Since the intensity scores range from 0 to 1, 0.03 is 3% of the full range. The maximal score difference ( INLINEFORM5 ) across all the submissions was as high as 0.34. Note, however, that these INLINEFORM6 s are the result of changing just one word in a sentence. In more complex sentences, several gender-associated words can appear, which may have a bigger impact. Also, whether consistent score differences of this magnitude will have significant repercussions in downstream applications, depends on the particular application.
Analyses on only the neutral sentences in EEC and only the emotional sentences in EEC: We also performed a separate analysis using only those sentences from the EEC that included no emotion words. Recall that there are four templates that contain no emotion words. Tables TABREF26 shows these results. We observe similar trends as in the analysis on the full set. One noticeable difference is that the number of submissions that showed statistically significant score difference is much smaller for this data subset. However, the total number of comparisons on the subset (44) is much smaller than the total number of comparisons on the full set (1,584), which makes the statistical test less powerful. Note also that the average score differences on the subset (columns 3 and 4 in Table TABREF26 ) tend to be higher than the differences on the full set (columns 3 and 4 in Table TABREF24 ). This indicates that gender-associated words can have a bigger impact on system predictions for neutral sentences.
We also performed an analysis by restricting the dataset to contain only the sentences with the emotion words corresponding to the emotion task (i.e., submissions to the anger intensity prediction task were evaluated only on sentences with anger words). The results (not shown here) were similar to the results on the full set.
Race Bias Results
We did a similar analysis for race as we did for gender. For each submission on each task, we calculated the difference between the average predicted score on the set of sentences with African American (AA) names and the average predicted score on the set of sentences with European American (EA) names. Then, we aggregated the results over all such sentence pairs in the EEC.
Table TABREF29 shows the results. The table has the same form and structure as the gender result tables. Observe that the number of submissions with no statistically significant score difference for sentences pertaining to the two races is about 5–11 (about 11% to 24%) for the four emotions and 3 (about 8%) for valence. These numbers are even lower than what was found for gender.
The majority of the systems assigned higher scores to sentences with African American names on the tasks of anger, fear, and sadness intensity prediction. On the joy and valence tasks, most submissions tended to assign higher scores to sentences with European American names. These tendencies reflect some common stereotypes that associate African Americans with more negative emotions BIBREF28 .
Figure FIGREF28 shows the score differences for individual systems on race sentence pairs on the valence regression task. Plots for the four emotion intensity prediction tasks are shown in Figure FIGREF32 in the Appendix. Here, the INLINEFORM0 –spreads are smaller than on the gender sentence pairs—from 0 to 0.15. As in the gender analysis, on the valence task the top 13 systems as well as some of the worst performing systems have smaller INLINEFORM1 –spread while the systems with medium to low performance show greater sensitivity to the race-associated names. However, we do not observe the same pattern in the emotion intensity tasks. Also, similar to the gender analysis, most submissions that showed no statistically significant score differences obtained lower scores on the tweets test sets. Only one system out of the top five showed no statistically significant score difference on the anger and fear intensity tasks, and none on the other tasks. Once again, just as in the case of gender, this raises questions of the exact causes of such biases. We hope to explore this in future work.
Discussion
As mentioned in the introduction, bias can originate from any or several parts of a system: the labeled and unlabeled datasets used to learn different parts of the model, the language resources used (e.g., pre-trained word embeddings, lexicons), the learning method used (algorithm, features, parameters), etc. In our analysis, we found systems trained using a variety of algorithms (traditional as well as deep neural networks) and a variety of language resources showing gender and race biases. Further experiments may tease out the extent of bias in each of these parts.
We also analyzed the output of our baseline SVM system trained using word unigrams (SVM-Unigrams). The system does not use any language resources other than the training data. We observe that this baseline system also shows small bias in gender and race. The INLINEFORM0 -spreads for this system were quite small: 0.09 to 0.2 on the gender sentence pairs and less than 0.002 on the race sentence pairs. The predicted intensity scores tended to be higher on the sentences with male noun phrases than on the sentences with female noun phrases for the tasks of anger, fear, and sadness intensity prediction. This tendency was reversed on the task of valence prediction. On the race sentence pairs, the system predicted higher intensity scores on the sentences with European American names for all four emotion intensity prediction tasks, and on the sentences with African American names for the task of valence prediction. This indicates that the training data contains some biases (in the form of some unigrams associated with a particular gender or race tending to appear in tweets labeled with certain emotions). The labeled datasets for the shared task were created using a fairly standard approach: polling Twitter with task-related query terms (in this case, emotion words) and then manually annotating the tweets with task-specific labels. The SVM-Unigram bias results show that data collected by distant supervision can be a source of bias. However, it should be noted that different learning methods in combination with different language resources can accentuate, reverse, or mask the bias present in the training data to different degrees.
Conclusions and Future Work
We created the Equity Evaluation Corpus (EEC), which consists of 8,640 sentences specifically chosen to tease out gender and race biases in natural language processing systems. We used the EEC to analyze 219 NLP systems that participated in a recent international shared task on predicting sentiment and emotion intensity. We found that more than 75% of the systems tend to mark sentences involving one gender/race with higher intensity scores than the sentences involving the other gender/race. We found such biases to be more widely prevalent for race than for gender. We also found that the bias can be different depending on the particular affect dimension involved.
We found the score differences across genders and across races to be somewhat small on average ( INLINEFORM0 , which is INLINEFORM1 of the 0 to 1 score range). However, for some systems the score differences reached as high as 0.34 (34%). What impact a consistent bias, even with an average magnitude INLINEFORM2 , might have in downstream applications merits further investigation.
We plan to extend the EEC with sentences associated with country names, professions (e.g., doctors, police officers, janitors, teachers, etc.), fields of study (e.g., arts vs. sciences), as well as races (e.g., Asian, mixed, etc.) and genders (e.g., agender, androgyne, trans, queer, etc.) not included in the current study. We can then use the corpus to examine biases across each of those variables as well. We are also interested in exploring which systems (or what techniques) accentuate inappropriate biases in the data and which systems mitigate such biases. Finally, we are interested in exploring how the quality of sentiment analysis predictions varies when applied to text produced by different demographic groups, such as people of different races, genders, and ethnicities.
The Equity Evaluation Corpus and the proposed methodology to examine bias are not meant to be comprehensive. However, using several approaches and datasets such as the one proposed here can bring about a more thorough examination of inappropriate biases in modern machine learning systems.
Appendix
Figures FIGREF31 and FIGREF32 show box plots of the score differences for each system on the four emotion intensity regression tasks on the gender and race sentence pairs, respectively. Each point on a plot corresponds to the difference in scores predicted by the system on one sentence pair. The systems are ordered by their performance rank (from first to last) on the task as per the official evaluation metric on the tweets test sets.
Question:
What criteria are used to select the 8,640 English sentences?
Answer:
| Sentences involving at least one race- or gender-associated word, sentence have to be short and grammatically simple, sentence have to include expressions of sentiment and emotion. | 328 | 30,036 | 30,038 | 30,109 | ... [The rest of the article is omitted]
|
0f12dc077fe8e5b95ca9163cea1dd17195c96929 | 0f12dc077fe8e5b95ca9163cea1dd17195c96929_1 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
[0]leftmargin=* [0]leftmargin=*
Automatic systems have had a significant and beneficial impact on all walks of human life. So much so that it is easy to overlook their potential to benefit society by promoting equity, diversity, and fairness. For example, machines do not take bribes to do their jobs, they can determine eligibility for a loan without being influenced by the color of the applicant's skin, and they can provide access to information and services without discrimination based on gender or sexual orientation. Nonetheless, as machine learning systems become more human-like in their predictions, they can also perpetuate human biases. Some learned biases may be beneficial for the downstream application (e.g., learning that humans often use some insect names, such as spider or cockroach, to refer to unpleasant situations). Other biases can be inappropriate and result in negative experiences for some groups of people. Examples include, loan eligibility and crime recidivism prediction systems that negatively assess people belonging to a certain pin/zip code (which may disproportionately impact people of a certain race) BIBREF0 and resumé sorting systems that believe that men are more qualified to be programmers than women BIBREF1 . Similarly, sentiment and emotion analysis systems can also perpetuate and accentuate inappropriate human biases, e.g., systems that consider utterances from one race or gender to be less positive simply because of their race or gender, or customer support systems that prioritize a call from an angry male over a call from the equally angry female.
Predictions of machine learning systems have also been shown to be of higher quality when dealing with information from some groups of people as opposed to other groups of people. For example, in the area of computer vision, gender classification systems perform particularly poorly for darker skinned females BIBREF2 . Natural language processing (NLP) systems have been shown to be poor in understanding text produced by people belonging to certain races BIBREF3 , BIBREF4 . For NLP systems, the sources of the bias often include the training data, other corpora, lexicons, and word embeddings that the machine learning algorithm may leverage to build its prediction model.
Even though there is some recent work highlighting such inappropriate biases (such as the work mentioned above), each such past work has largely focused on just one or two systems and resources. Further, there is no benchmark dataset for examining inappropriate biases in natural language systems. In this paper, we describe how we compiled a dataset of 8,640 English sentences carefully chosen to tease out biases towards certain races and genders. We will refer to it as the Equity Evaluation Corpus (EEC). We used the EEC as a supplementary test set in a recent shared task on predicting sentiment and emotion intensity in tweets, SemEval-2018 Task 1: Affect in Tweets BIBREF5 . In particular, we wanted to test a hypothesis that a system should equally rate the intensity of the emotion expressed by two sentences that differ only in the gender/race of a person mentioned. Note that here the term system refers to the combination of a machine learning architecture trained on a labeled dataset, and possibly using additional language resources. The bias can originate from any or several of these parts. We were thus able to use the EEC to examine 219 sentiment analysis systems that took part in the shared task.
We compare emotion and sentiment intensity scores that the systems predict on pairs of sentences in the EEC that differ only in one word corresponding to race or gender (e.g., `This man made me feel angry' vs. `This woman made me feel angry'). We find that the majority of the systems studied show statistically significant bias; that is, they consistently provide slightly higher sentiment intensity predictions for sentences associated with one race or one gender. We also find that the bias may be different depending on the particular affect dimension that the natural language system is trained to predict.
Despite the work we describe here and what others have proposed in the past, it should be noted that there are no simple solutions for dealing with inappropriate human biases that percolate into machine learning systems. It seems difficult to ever be able to identify and quantify all of the inappropriate biases perfectly (even when restricted to the scope of just gender and race). Further, any such mechanism is liable to be circumvented, if one chooses to do so. Nonetheless, as developers of sentiment analysis systems, and NLP systems more broadly, we cannot absolve ourselves of the ethical implications of the systems we build. Even if it is unclear how we should deal with the inappropriate biases in our systems, we should be measuring such biases. The Equity Evaluation Corpus is not meant to be a catch-all for all inappropriate biases, but rather just one of the several ways by which we can examine the fairness of sentiment analysis systems. We make the corpus freely available so that both developers and users can use it, and build on it.
Related Work
Recent studies have demonstrated that the systems trained on the human-written texts learn human-like biases BIBREF1 , BIBREF6 . In general, any predictive model built on historical data may inadvertently inherit human biases based on gender, ethnicity, race, or religion BIBREF7 , BIBREF8 . Discrimination-aware data mining focuses on measuring discrimination in data as well as on evaluating performance of discrimination-aware predictive models BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 .
In NLP, the attention so far has been primarily on word embeddings—a popular and powerful framework to represent words as low-dimensional dense vectors. The word embeddings are usually obtained from large amounts of human-written texts, such as Wikipedia, Google News articles, or millions of tweets. Bias in sentiment analysis systems has only been explored in simple systems that make use of pre-computed word embeddings BIBREF13 . There is no prior work that systematically quantifies the extent of bias in a large number of sentiment analysis systems.
This paper does not examine the differences in accuracies of systems on text produced by different races or genders, as was done by hovy2015demographic,blodgett2016demographic,jurgens2017incorporating,buolamwini2018gender. Approaches on how to mitigate inappropriate biases BIBREF14 , BIBREF1 , BIBREF15 , BIBREF16 , BIBREF13 , BIBREF17 , BIBREF18 are also beyond the scope of this paper. See also the position paper by hovy2016social, which identifies socio-ethical implications of the NLP systems in general.
The Equity Evaluation Corpus
We now describe how we compiled a dataset of thousands of sentences to determine whether automatic systems consistently give higher (or lower) sentiment intensity scores to sentences involving a particular race or gender. There are several ways in which such a dataset may be compiled. We present below the choices that we made.
We decided to use sentences involving at least one race- or gender-associated word. The sentences were intended to be short and grammatically simple. We also wanted some sentences to include expressions of sentiment and emotion, since the goal is to test sentiment and emotion systems. We, the authors of this paper, developed eleven sentence templates after several rounds of discussion and consensus building. They are shown in Table TABREF3 . The templates are divided into two groups. The first type (templates 1–7) includes emotion words. The purpose of this set is to have sentences expressing emotions. The second type (templates 8–11) does not include any emotion words. The purpose of this set is to have non-emotional (neutral) sentences.
The templates include two variables: INLINEFORM0 person INLINEFORM1 and INLINEFORM2 emotion word INLINEFORM3 . We generate sentences from the template by instantiating each variable with one of the pre-chosen values that the variable can take. Each of the eleven templates includes the variable INLINEFORM4 person INLINEFORM5 . INLINEFORM6 person INLINEFORM7 can be instantiated by any of the following noun phrases:
For our study, we chose ten names of each kind from the study by Caliskan:2017 (see Table TABREF7 ). The full lists of noun phrases representing females and males, used in our study, are shown in Table TABREF8 .
The second variable, INLINEFORM0 emotion word INLINEFORM1 , has two variants. Templates one through four include a variable for an emotional state word. The emotional state words correspond to four basic emotions: anger, fear, joy, and sadness. Specifically, for each of the emotions, we selected five words that convey that emotion in varying intensities. These words were taken from the categories in the Roget's Thesaurus corresponding to the four emotions: category #900 Resentment (for anger), category #860 Fear (for fear), category #836 Cheerfulness (for joy), and category #837 Dejection (for sadness). Templates five through seven include emotion words describing a situation or event. These words were also taken from the same thesaurus categories listed above. The full lists of emotion words (emotional state words and emotional situation/event words) are shown in Table TABREF10 .
We generated sentences from the templates by replacing INLINEFORM0 person INLINEFORM1 and INLINEFORM2 emotion word INLINEFORM3 variables with the values they can take. In total, 8,640 sentences were generated with the various combinations of INLINEFORM4 person INLINEFORM5 and INLINEFORM6 emotion word INLINEFORM7 values across the eleven templates. We manually examined the sentences to make sure they were grammatically well-formed. Notably, one can derive pairs of sentences from the EEC such that they differ only in one word corresponding to gender or race (e.g., `My daughter feels devastated' and `My son feels devastated'). We refer to the full set of 8,640 sentences as Equity Evaluation Corpus.
Measuring Race and Gender Bias in Automatic Sentiment Analysis Systems
The race and gender bias evaluation was carried out on the output of the 219 automatic systems that participated in SemEval-2018 Task 1: Affect in Tweets BIBREF5 . The shared task included five subtasks on inferring the affectual state of a person from their tweet: 1. emotion intensity regression, 2. emotion intensity ordinal classification, 3. valence (sentiment) regression, 4. valence ordinal classification, and 5. emotion classification. For each subtask, labeled data were provided for English, Arabic, and Spanish. The race and gender bias were analyzed for the system outputs on two English subtasks: emotion intensity regression (for anger, fear, joy, and sadness) and valence regression. These regression tasks were formulated as follows: Given a tweet and an affective dimension A (anger, fear, joy, sadness, or valence), determine the intensity of A that best represents the mental state of the tweeter—a real-valued score between 0 (least A) and 1 (most A). Separate training and test datasets were provided for each affective dimension.
Training sets included tweets along with gold intensity scores. Two test sets were provided for each task: 1. a regular tweet test set (for which the gold intensity scores are known but not revealed to the participating systems), and 2. the Equity Evaluation Corpus (for which no gold intensity labels exist). Participants were told that apart from the usual test set, they are to run their systems on a separate test set of unknown origin. The participants were instructed to train their system on the tweets training sets provided, and that they could use any other resources they may find or create. They were to run the same final system on the two test sets. The nature of the second test set was revealed to them only after the competition. The first (tweets) test set was used to evaluate and rank the quality (accuracy) of the systems' predictions. The second (EEC) test set was used to perform the bias analysis, which is the focus of this paper.
Systems: Fifty teams submitted their system outputs to one or more of the five emotion intensity regression tasks (for anger, fear, joy, sadness, and valence), resulting in 219 submissions in total. Many systems were built using two types of features: deep neural network representations of tweets (sentence embeddings) and features derived from existing sentiment and emotion lexicons. These features were then combined to learn a model using either traditional machine learning algorithms (such as SVM/SVR and Logistic Regression) or deep neural networks. SVM/SVR, LSTMs, and Bi-LSTMs were some of the most widely used machine learning algorithms. The sentence embeddings were obtained by training a neural network on the provided training data, a distant supervision corpus (e.g., AIT2018 Distant Supervision Corpus that has tweets with emotion-related query terms), sentiment-labeled tweet corpora (e.g., Semeval-2017 Task4A dataset on sentiment analysis in Twitter), or by using pre-trained models (e.g., DeepMoji BIBREF20 , Skip thoughts BIBREF21 ). The lexicon features were often derived from the NRC emotion and sentiment lexicons BIBREF22 , BIBREF23 , BIBREF24 , AFINN BIBREF25 , and Bing Liu Lexicon BIBREF26 .
We provided a baseline SVM system trained using word unigrams as features on the training data (SVM-Unigrams). This system is also included in the current analysis.
Measuring bias: To examine gender bias, we compared each system's predicted scores on the EEC sentence pairs as follows:
Thus, eleven pairs of scores (ten pairs of scores from ten noun phrase pairs and one pair of scores from the averages on name subsets) were examined for each template–emotion word instantiation. There were twenty different emotion words used in seven templates (templates 1–7), and no emotion words used in the four remaining templates (templates 8–11). In total, INLINEFORM0 pairs of scores were compared.
Similarly, to examine race bias, we compared pairs of system predicted scores as follows:
Thus, one pair of scores was examined for each template–emotion word instantiation. In total, INLINEFORM0 pairs of scores were compared.
For each system, we calculated the paired two sample t-test to determine whether the mean difference between the two sets of scores (across the two races and across the two genders) is significant. We set the significance level to 0.05. However, since we performed 438 assessments (219 submissions evaluated for biases in both gender and race), we applied Bonferroni correction. The null hypothesis that the true mean difference between the paired samples was zero was rejected if the calculated p-value fell below INLINEFORM0 .
Results
The two sub-sections below present the results from the analysis for gender bias and race bias, respectively.
Gender Bias Results
Individual submission results were communicated to the participants. Here, we present the summary results across all the teams. The goal of this analysis is to gain a better understanding of biases across a large number of current sentiment analysis systems. Thus, we partition the submissions into three groups according to the bias they show:
F=M not significant: submissions that showed no statistically significant difference in intensity scores predicted for corresponding female and male noun phrase sentences,
F INLINEFORM0 –M INLINEFORM1 significant: submissions that consistently gave higher scores for sentences with female noun phrases than for corresponding sentences with male noun phrases,
F INLINEFORM0 –M INLINEFORM1 significant: submissions that consistently gave lower scores for sentences with female noun phrases than for corresponding sentences with male noun phrases.
For each system and each sentence pair, we calculate the score difference INLINEFORM0 as the score for the female noun phrase sentence minus the score for the corresponding male noun phrase sentence. Table TABREF24 presents the summary results for each of the bias groups. It has the following columns:
#Subm.: number of submissions in each group.
If all the systems are unbiased, then the number of submissions for the group F=M not significant would be the maximum, and the number of submissions in all other groups would be zero.
Avg. score difference F INLINEFORM0 –M INLINEFORM1 : the average INLINEFORM2 for only those pairs where the score for the female noun phrase sentence is higher. The greater the magnitude of this score, the stronger the bias in systems that consistently give higher scores to female-associated sentences.
Avg. score difference F INLINEFORM0 –M INLINEFORM1 : the average INLINEFORM2 for only those pairs where the score for the female noun phrase sentence is lower. The greater the magnitude of this score, the stronger the bias in systems that consistently give lower scores to female-associated sentences.
Note that these numbers were first calculated separately for each submission, and then averaged over all the submissions within each submission group. The results are reported separately for submissions to each task (anger, fear, joy, sadness, and sentiment/valence intensity prediction).
Observe that on the four emotion intensity prediction tasks, only about 12 of the 46 submissions (about 25% of the submissions) showed no statistically significant score difference. On the valence prediction task, only 5 of the 36 submissions (14% of the submissions) showed no statistically significant score difference. Thus 75% to 86% of the submissions consistently marked sentences of one gender higher than another.
When predicting anger, joy, or valence, the number of systems consistently giving higher scores to sentences with female noun phrases (21–25) is markedly higher than the number of systems giving higher scores to sentences with male noun phrases (8–13). (Recall that higher valence means more positive sentiment.) In contrast, on the fear task, most submissions tended to assign higher scores to sentences with male noun phrases (23) as compared to the number of systems giving higher scores to sentences with female noun phrases (12). When predicting sadness, the number of submissions that mostly assigned higher scores to sentences with female noun phrases (18) is close to the number of submissions that mostly assigned higher scores to sentences with male noun phrases (16). These results are in line with some common stereotypes, such as females are more emotional, and situations involving male agents are more fearful BIBREF27 .
Figure FIGREF25 shows the score differences ( INLINEFORM0 ) for individual systems on the valence regression task. Plots for the four emotion intensity prediction tasks are shown in Figure FIGREF31 in the Appendix. Each point ( ▲, ▼, ●) on the plot corresponds to the difference in scores predicted by the system on one sentence pair. The systems are ordered by their rank (from first to last) on the task on the tweets test sets, as per the official evaluation metric (Spearman correlation with the gold intensity scores). We will refer to the difference between the maximal value of INLINEFORM1 and the minimal value of INLINEFORM2 for a particular system as the INLINEFORM3 –spread. Observe that the INLINEFORM4 –spreads for many systems are rather large, up to 0.57. Depending on the task, the top 10 or top 15 systems as well as some of the worst performing systems tend to have smaller INLINEFORM5 –spreads while the systems with medium to low performance show greater sensitivity to the gender-associated words. Also, most submissions that showed no statistically significant score differences (shown in green) performed poorly on the tweets test sets. Only three systems out of the top five on the anger intensity task and one system on the joy and sadness tasks showed no statistically significant score difference. This indicates that when considering only those systems that performed well on the intensity prediction task, the percentage of gender-biased systems are even higher than those indicated above.
These results raise further questions such as `what exactly is the cause of such biases?' and `why is the bias impacted by the emotion task under consideration?'. Answering these questions will require further information on the resources that the teams used to develop their models, and we leave that for future work.
Average score differences: For submissions that showed statistically significant score differences, the average score difference F INLINEFORM0 –M INLINEFORM1 and the average score difference F INLINEFORM2 –M INLINEFORM3 were INLINEFORM4 . Since the intensity scores range from 0 to 1, 0.03 is 3% of the full range. The maximal score difference ( INLINEFORM5 ) across all the submissions was as high as 0.34. Note, however, that these INLINEFORM6 s are the result of changing just one word in a sentence. In more complex sentences, several gender-associated words can appear, which may have a bigger impact. Also, whether consistent score differences of this magnitude will have significant repercussions in downstream applications, depends on the particular application.
Analyses on only the neutral sentences in EEC and only the emotional sentences in EEC: We also performed a separate analysis using only those sentences from the EEC that included no emotion words. Recall that there are four templates that contain no emotion words. Tables TABREF26 shows these results. We observe similar trends as in the analysis on the full set. One noticeable difference is that the number of submissions that showed statistically significant score difference is much smaller for this data subset. However, the total number of comparisons on the subset (44) is much smaller than the total number of comparisons on the full set (1,584), which makes the statistical test less powerful. Note also that the average score differences on the subset (columns 3 and 4 in Table TABREF26 ) tend to be higher than the differences on the full set (columns 3 and 4 in Table TABREF24 ). This indicates that gender-associated words can have a bigger impact on system predictions for neutral sentences.
We also performed an analysis by restricting the dataset to contain only the sentences with the emotion words corresponding to the emotion task (i.e., submissions to the anger intensity prediction task were evaluated only on sentences with anger words). The results (not shown here) were similar to the results on the full set.
Race Bias Results
We did a similar analysis for race as we did for gender. For each submission on each task, we calculated the difference between the average predicted score on the set of sentences with African American (AA) names and the average predicted score on the set of sentences with European American (EA) names. Then, we aggregated the results over all such sentence pairs in the EEC.
Table TABREF29 shows the results. The table has the same form and structure as the gender result tables. Observe that the number of submissions with no statistically significant score difference for sentences pertaining to the two races is about 5–11 (about 11% to 24%) for the four emotions and 3 (about 8%) for valence. These numbers are even lower than what was found for gender.
The majority of the systems assigned higher scores to sentences with African American names on the tasks of anger, fear, and sadness intensity prediction. On the joy and valence tasks, most submissions tended to assign higher scores to sentences with European American names. These tendencies reflect some common stereotypes that associate African Americans with more negative emotions BIBREF28 .
Figure FIGREF28 shows the score differences for individual systems on race sentence pairs on the valence regression task. Plots for the four emotion intensity prediction tasks are shown in Figure FIGREF32 in the Appendix. Here, the INLINEFORM0 –spreads are smaller than on the gender sentence pairs—from 0 to 0.15. As in the gender analysis, on the valence task the top 13 systems as well as some of the worst performing systems have smaller INLINEFORM1 –spread while the systems with medium to low performance show greater sensitivity to the race-associated names. However, we do not observe the same pattern in the emotion intensity tasks. Also, similar to the gender analysis, most submissions that showed no statistically significant score differences obtained lower scores on the tweets test sets. Only one system out of the top five showed no statistically significant score difference on the anger and fear intensity tasks, and none on the other tasks. Once again, just as in the case of gender, this raises questions of the exact causes of such biases. We hope to explore this in future work.
Discussion
As mentioned in the introduction, bias can originate from any or several parts of a system: the labeled and unlabeled datasets used to learn different parts of the model, the language resources used (e.g., pre-trained word embeddings, lexicons), the learning method used (algorithm, features, parameters), etc. In our analysis, we found systems trained using a variety of algorithms (traditional as well as deep neural networks) and a variety of language resources showing gender and race biases. Further experiments may tease out the extent of bias in each of these parts.
We also analyzed the output of our baseline SVM system trained using word unigrams (SVM-Unigrams). The system does not use any language resources other than the training data. We observe that this baseline system also shows small bias in gender and race. The INLINEFORM0 -spreads for this system were quite small: 0.09 to 0.2 on the gender sentence pairs and less than 0.002 on the race sentence pairs. The predicted intensity scores tended to be higher on the sentences with male noun phrases than on the sentences with female noun phrases for the tasks of anger, fear, and sadness intensity prediction. This tendency was reversed on the task of valence prediction. On the race sentence pairs, the system predicted higher intensity scores on the sentences with European American names for all four emotion intensity prediction tasks, and on the sentences with African American names for the task of valence prediction. This indicates that the training data contains some biases (in the form of some unigrams associated with a particular gender or race tending to appear in tweets labeled with certain emotions). The labeled datasets for the shared task were created using a fairly standard approach: polling Twitter with task-related query terms (in this case, emotion words) and then manually annotating the tweets with task-specific labels. The SVM-Unigram bias results show that data collected by distant supervision can be a source of bias. However, it should be noted that different learning methods in combination with different language resources can accentuate, reverse, or mask the bias present in the training data to different degrees.
Conclusions and Future Work
We created the Equity Evaluation Corpus (EEC), which consists of 8,640 sentences specifically chosen to tease out gender and race biases in natural language processing systems. We used the EEC to analyze 219 NLP systems that participated in a recent international shared task on predicting sentiment and emotion intensity. We found that more than 75% of the systems tend to mark sentences involving one gender/race with higher intensity scores than the sentences involving the other gender/race. We found such biases to be more widely prevalent for race than for gender. We also found that the bias can be different depending on the particular affect dimension involved.
We found the score differences across genders and across races to be somewhat small on average ( INLINEFORM0 , which is INLINEFORM1 of the 0 to 1 score range). However, for some systems the score differences reached as high as 0.34 (34%). What impact a consistent bias, even with an average magnitude INLINEFORM2 , might have in downstream applications merits further investigation.
We plan to extend the EEC with sentences associated with country names, professions (e.g., doctors, police officers, janitors, teachers, etc.), fields of study (e.g., arts vs. sciences), as well as races (e.g., Asian, mixed, etc.) and genders (e.g., agender, androgyne, trans, queer, etc.) not included in the current study. We can then use the corpus to examine biases across each of those variables as well. We are also interested in exploring which systems (or what techniques) accentuate inappropriate biases in the data and which systems mitigate such biases. Finally, we are interested in exploring how the quality of sentiment analysis predictions varies when applied to text produced by different demographic groups, such as people of different races, genders, and ethnicities.
The Equity Evaluation Corpus and the proposed methodology to examine bias are not meant to be comprehensive. However, using several approaches and datasets such as the one proposed here can bring about a more thorough examination of inappropriate biases in modern machine learning systems.
Appendix
Figures FIGREF31 and FIGREF32 show box plots of the score differences for each system on the four emotion intensity regression tasks on the gender and race sentence pairs, respectively. Each point on a plot corresponds to the difference in scores predicted by the system on one sentence pair. The systems are ordered by their performance rank (from first to last) on the task as per the official evaluation metric on the tweets test sets.
Question:
What criteria are used to select the 8,640 English sentences?
Answer:
| generated with the various combinations of INLINEFORM4 person INLINEFORM5 and INLINEFORM6 emotion word INLINEFORM7 values across the eleven templates, differ only in one word corresponding to gender or race | 328 | 30,036 | 30,038 | 30,109 | ... [The rest of the article is omitted]
|
518dae6f936882152c162058895db4eca815e649 | 518dae6f936882152c162058895db4eca815e649_0 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/ Deep neural networks have been widely used in text classification and have achieved promising results BIBREF0 , BIBREF1 , BIBREF2 . Most focus on content information and use models such as convolutional neural networks (CNN) BIBREF3 or recursive neural networks BIBREF4 . However, for user-generated posts on social media like Facebook or Twitter, there is more information that should not be ignored. On social media platforms, a user can act either as the author of a post or as a reader who expresses his or her comments about the post.
In this paper, we classify posts taking into account post authorship, likes, topics, and comments. In particular, users and their “likes” hold strong potential for text mining. For example, given a set of posts that are related to a specific topic, a user's likes and dislikes provide clues for stance labeling. From a user point of view, users with positive attitudes toward the issue leave positive comments on the posts with praise or even just the post's content; from a post point of view, positive posts attract users who hold positive stances. We also investigate the influence of topics: different topics are associated with different stance labeling tendencies and word usage. For example we discuss women's rights and unwanted babies on the topic of abortion, but we criticize medicine usage or crime when on the topic of marijuana BIBREF5 . Even for posts on a specific topic like nuclear power, a variety of arguments are raised: green energy, radiation, air pollution, and so on. As for comments, we treat them as additional text information. The arguments in the comments and the commenters (the users who leave the comments) provide hints on the post's content and further facilitate stance classification.
In this paper, we propose the user-topic-comment neural network (UTCNN), a deep learning model that utilizes user, topic, and comment information. We attempt to learn user and topic representations which encode user interactions and topic influences to further enhance text classification, and we also incorporate comment information. We evaluate this model on a post stance classification task on forum-style social media platforms. The contributions of this paper are as follows: 1. We propose UTCNN, a neural network for text in modern social media channels as well as legacy social media, forums, and message boards — anywhere that reveals users, their tastes, as well as their replies to posts. 2. When classifying social media post stances, we leverage users, including authors and likers. User embeddings can be generated even for users who have never posted anything. 3. We incorporate a topic model to automatically assign topics to each post in a single topic dataset. 4. We show that overall, the proposed method achieves the highest performance in all instances, and that all of the information extracted, whether users, topics, or comments, still has its contributions.
Extra-Linguistic Features for Stance Classification
In this paper we aim to use text as well as other features to see how they complement each other in a deep learning model. In the stance classification domain, previous work has showed that text features are limited, suggesting that adding extra-linguistic constraints could improve performance BIBREF6 , BIBREF7 , BIBREF8 . For example, Hasan and Ng as well as Thomas et al. require that posts written by the same author have the same stance BIBREF9 , BIBREF10 . The addition of this constraint yields accuracy improvements of 1–7% for some models and datasets. Hasan and Ng later added user-interaction constraints and ideology constraints BIBREF7 : the former models the relationship among posts in a sequence of replies and the latter models inter-topic relationships, e.g., users who oppose abortion could be conservative and thus are likely to oppose gay rights.
For work focusing on online forum text, since posts are linked through user replies, sequential labeling methods have been used to model relationships between posts. For example, Hasan and Ng use hidden Markov models (HMMs) to model dependent relationships to the preceding post BIBREF9 ; Burfoot et al. use iterative classification to repeatedly generate new estimates based on the current state of knowledge BIBREF11 ; Sridhar et al. use probabilistic soft logic (PSL) to model reply links via collaborative filtering BIBREF12 . In the Facebook dataset we study, we use comments instead of reply links. However, as the ultimate goal in this paper is predicting not comment stance but post stance, we treat comments as extra information for use in predicting post stance.
Deep Learning on Extra-Linguistic Features
In recent years neural network models have been applied to document sentiment classification BIBREF13 , BIBREF4 , BIBREF14 , BIBREF15 , BIBREF2 . Text features can be used in deep networks to capture text semantics or sentiment. For example, Dong et al. use an adaptive layer in a recursive neural network for target-dependent Twitter sentiment analysis, where targets are topics such as windows 7 or taylor swift BIBREF16 , BIBREF17 ; recursive neural tensor networks (RNTNs) utilize sentence parse trees to capture sentence-level sentiment for movie reviews BIBREF4 ; Le and Mikolov predict sentiment by using paragraph vectors to model each paragraph as a continuous representation BIBREF18 . They show that performance can thus be improved by more delicate text models.
Others have suggested using extra-linguistic features to improve the deep learning model. The user-word composition vector model (UWCVM) BIBREF19 is inspired by the possibility that the strength of sentiment words is user-specific; to capture this they add user embeddings in their model. In UPNN, a later extension, they further add a product-word composition as product embeddings, arguing that products can also show different tendencies of being rated or reviewed BIBREF20 . Their addition of user information yielded 2–10% improvements in accuracy as compared to the above-mentioned RNTN and paragraph vector methods. We also seek to inject user information into the neural network model. In comparison to the research of Tang et al. on sentiment classification for product reviews, the difference is two-fold. First, we take into account multiple users (one author and potentially many likers) for one post, whereas only one user (the reviewer) is involved in a review. Second, we add comment information to provide more features for post stance classification. None of these two factors have been considered previously in a deep learning model for text stance classification. Therefore, we propose UTCNN, which generates and utilizes user embeddings for all users — even for those who have not authored any posts — and incorporates comments to further improve performance.
Method
In this section, we first describe CNN-based document composition, which captures user- and topic-dependent document-level semantic representation from word representations. Then we show how to add comment information to construct the user-topic-comment neural network (UTCNN).
User- and Topic-dependent Document Composition
As shown in Figure FIGREF4 , we use a general CNN BIBREF3 and two semantic transformations for document composition . We are given a document with an engaged user INLINEFORM0 , a topic INLINEFORM1 , and its composite INLINEFORM2 words, each word INLINEFORM3 of which is associated with a word embedding INLINEFORM4 where INLINEFORM5 is the vector dimension. For each word embedding INLINEFORM6 , we apply two dot operations as shown in Equation EQREF6 : DISPLAYFORM0
where INLINEFORM0 models the user reading preference for certain semantics, and INLINEFORM1 models the topic semantics; INLINEFORM2 and INLINEFORM3 are the dimensions of transformed user and topic embeddings respectively. We use INLINEFORM4 to model semantically what each user prefers to read and/or write, and use INLINEFORM5 to model the semantics of each topic. The dot operation of INLINEFORM6 and INLINEFORM7 transforms the global representation INLINEFORM8 to a user-dependent representation. Likewise, the dot operation of INLINEFORM9 and INLINEFORM10 transforms INLINEFORM11 to a topic-dependent representation.
After the two dot operations on INLINEFORM0 , we have user-dependent and topic-dependent word vectors INLINEFORM1 and INLINEFORM2 , which are concatenated to form a user- and topic-dependent word vector INLINEFORM3 . Then the transformed word embeddings INLINEFORM4 are used as the CNN input. Here we apply three convolutional layers on the concatenated transformed word embeddings INLINEFORM5 : DISPLAYFORM0
where INLINEFORM0 is the index of words; INLINEFORM1 is a non-linear activation function (we use INLINEFORM2 ); INLINEFORM5 is the convolutional filter with input length INLINEFORM6 and output length INLINEFORM7 , where INLINEFORM8 is the window size of the convolutional operation; and INLINEFORM9 and INLINEFORM10 are the output and bias of the convolution layer INLINEFORM11 , respectively. In our experiments, the three window sizes INLINEFORM12 in the three convolution layers are one, two, and three, encoding unigram, bigram, and trigram semantics accordingly.
After the convolutional layer, we add a maximum pooling layer among convolutional outputs to obtain the unigram, bigram, and trigram n-gram representations. This is succeeded by an average pooling layer for an element-wise average of the three maximized convolution outputs.
UTCNN Model Description
Figure FIGREF10 illustrates the UTCNN model. As more than one user may interact with a given post, we first add a maximum pooling layer after the user matrix embedding layer and user vector embedding layer to form a moderator matrix embedding INLINEFORM0 and a moderator vector embedding INLINEFORM1 for moderator INLINEFORM2 respectively, where INLINEFORM3 is used for the semantic transformation in the document composition process, as mentioned in the previous section. The term moderator here is to denote the pseudo user who provides the overall semantic/sentiment of all the engaged users for one document. The embedding INLINEFORM4 models the moderator stance preference, that is, the pattern of the revealed user stance: whether a user is willing to show his preference, whether a user likes to show impartiality with neutral statements and reasonable arguments, or just wants to show strong support for one stance. Ideally, the latent user stance is modeled by INLINEFORM5 for each user. Likewise, for topic information, a maximum pooling layer is added after the topic matrix embedding layer and topic vector embedding layer to form a joint topic matrix embedding INLINEFORM6 and a joint topic vector embedding INLINEFORM7 for topic INLINEFORM8 respectively, where INLINEFORM9 models the semantic transformation of topic INLINEFORM10 as in users and INLINEFORM11 models the topic stance tendency. The latent topic stance is also modeled by INLINEFORM12 for each topic.
As for comments, we view them as short documents with authors only but without likers nor their own comments. Therefore we apply document composition on comments although here users are commenters (users who comment). It is noticed that the word embeddings INLINEFORM0 for the same word in the posts and comments are the same, but after being transformed to INLINEFORM1 in the document composition process shown in Figure FIGREF4 , they might become different because of their different engaged users. The output comment representation together with the commenter vector embedding INLINEFORM2 and topic vector embedding INLINEFORM3 are concatenated and a maximum pooling layer is added to select the most important feature for comments. Instead of requiring that the comment stance agree with the post, UTCNN simply extracts the most important features of the comment contents; they could be helpful, whether they show obvious agreement or disagreement. Therefore when combining comment information here, the maximum pooling layer is more appropriate than other pooling or merging layers. Indeed, we believe this is one reason for UTCNN's performance gains.
Finally, the pooled comment representation, together with user vector embedding INLINEFORM0 , topic vector embedding INLINEFORM1 , and document representation are fed to a fully connected network, and softmax is applied to yield the final stance label prediction for the post.
Experiment
We start with the experimental dataset and then describe the training process as well as the implementation of the baselines. We also implement several variations to reveal the effects of features: authors, likers, comment, and commenters. In the results section we compare our model with related work.
Dataset
We tested the proposed UTCNN on two different datasets: FBFans and CreateDebate. FBFans is a privately-owned, single-topic, Chinese, unbalanced, social media dataset, and CreateDebate is a public, multiple-topic, English, balanced, forum dataset. Results using these two datasets show the applicability and superiority for different topics, languages, data distributions, and platforms.
The FBFans dataset contains data from anti-nuclear-power Chinese Facebook fan groups from September 2013 to August 2014, including posts and their author and liker IDs. There are a total of 2,496 authors, 505,137 likers, 33,686 commenters, and 505,412 unique users. Two annotators were asked to take into account only the post content to label the stance of the posts in the whole dataset as supportive, neutral, or unsupportive (hereafter denoted as Sup, Neu, and Uns). Sup/Uns posts were those in support of or against anti-reconstruction; Neu posts were those evincing a neutral standpoint on the topic, or were irrelevant. Raw agreement between annotators is 0.91, indicating high agreement. Specifically, Cohen’s Kappa for Neu and not Neu labeling is 0.58 (moderate), and for Sup or Uns labeling is 0.84 (almost perfect). Posts with inconsistent labels were filtered out, and the development and testing sets were randomly selected from what was left. Posts in the development and testing sets involved at least one user who appeared in the training set. The number of posts for each stance is shown on the left-hand side of Table TABREF12 . About twenty percent of the posts were labeled with a stance, and the number of supportive (Sup) posts was much larger than that of the unsupportive (Uns) ones: this is thus highly skewed data, which complicates stance classification. On average, 161.1 users were involved in one post. The maximum was 23,297 and the minimum was one (the author). For comments, on average there were 3 comments per post. The maximum was 1,092 and the minimum was zero.
To test whether the assumption of this paper – posts attract users who hold the same stance to like them – is reliable, we examine the likes from authors of different stances. Posts in FBFans dataset are used for this analysis. We calculate the like statistics of each distinct author from these 32,595 posts. As the numbers of authors in the Sup, Neu and Uns stances are largely imbalanced, these numbers are normalized by the number of users of each stance. Table TABREF13 shows the results. Posts with stances (i.e., not neutral) attract users of the same stance. Neutral posts also attract both supportive and neutral users, like what we observe in supportive posts, but just the neutral posts can attract even more neutral likers. These results do suggest that users prefer posts of the same stance, or at least posts of no obvious stance which might cause annoyance when reading, and hence support the user modeling in our approach.
The CreateDebate dataset was collected from an English online debate forum discussing four topics: abortion (ABO), gay rights (GAY), Obama (OBA), and marijuana (MAR). The posts are annotated as for (F) and against (A). Replies to posts in this dataset are also labeled with stance and hence use the same data format as posts. The labeling results are shown in the right-hand side of Table TABREF12 . We observe that the dataset is more balanced than the FBFans dataset. In addition, there are 977 unique users in the dataset. To compare with Hasan and Ng's work, we conducted five-fold cross-validation and present the annotation results as the average number of all folds BIBREF9 , BIBREF5 .
The FBFans dataset has more integrated functions than the CreateDebate dataset; thus our model can utilize all linguistic and extra-linguistic features. For the CreateDebate dataset, on the other hand, the like and comment features are not available (as there is a stance label for each reply, replies are evaluated as posts as other previous work) but we still implemented our model using the content, author, and topic information.
Settings
In the UTCNN training process, cross-entropy was used as the loss function and AdaGrad as the optimizer. For FBFans dataset, we learned the 50-dimensional word embeddings on the whole dataset using GloVe BIBREF21 to capture the word semantics; for CreateDebate dataset we used the publicly available English 50-dimensional word embeddings, pre-trained also using GloVe. These word embeddings were fixed in the training process. The learning rate was set to 0.03. All user and topic embeddings were randomly initialized in the range of [-0.1 0.1]. Matrix embeddings for users and topics were sized at 250 ( INLINEFORM0 ); vector embeddings for users and topics were set to length 10.
We applied the LDA topic model BIBREF22 on the FBFans dataset to determine the latent topics with which to build topic embeddings, as there is only one general known topic: nuclear power plants. We learned 100 latent topics and assigned the top three topics for each post. For the CreateDebate dataset, which itself constitutes four topics, the topic labels for posts were used directly without additionally applying LDA.
For the FBFans data we report class-based f-scores as well as the macro-average f-score ( INLINEFORM0 ) shown in equation EQREF19 . DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are the average precision and recall of the three class. We adopted the macro-average f-score as the evaluation metric for the overall performance because (1) the experimental dataset is severely imbalanced, which is common for contentious issues; and (2) for stance classification, content in minor-class posts is usually more important for further applications. For the CreateDebate dataset, accuracy was adopted as the evaluation metric to compare the results with related work BIBREF7 , BIBREF9 , BIBREF12 .
Baselines
We pit our model against the following baselines: 1) SVM with unigram, bigram, and trigram features, which is a standard yet rather strong classifier for text features; 2) SVM with average word embedding, where a document is represented as a continuous representation by averaging the embeddings of the composite words; 3) SVM with average transformed word embeddings (the INLINEFORM0 in equation EQREF6 ), where a document is represented as a continuous representation by averaging the transformed embeddings of the composite words; 4) two mature deep learning models on text classification, CNN BIBREF3 and Recurrent Convolutional Neural Networks (RCNN) BIBREF0 , where the hyperparameters are based on their work; 5) the above SVM and deep learning models with comment information; 6) UTCNN without user information, representing a pure-text CNN model where we use the same user matrix and user embeddings INLINEFORM1 and INLINEFORM2 for each user; 7) UTCNN without the LDA model, representing how UTCNN works with a single-topic dataset; 8) UTCNN without comments, in which the model predicts the stance label given only user and topic information. All these models were trained on the training set, and parameters as well as the SVM kernel selections (linear or RBF) were fine-tuned on the development set. Also, we adopt oversampling on SVMs, CNN and RCNN because the FBFans dataset is highly imbalanced.
Results on FBFans Dataset
In Table TABREF22 we show the results of UTCNN and the baselines on the FBFans dataset. Here Majority yields good performance on Neu since FBFans is highly biased to the neutral class. The SVM models perform well on Sup and Neu but perform poorly for Uns, showing that content information in itself is insufficient to predict stance labels, especially for the minor class. With the transformed word embedding feature, SVM can achieve comparable performance as SVM with n-gram feature. However, the much fewer feature dimension of the transformed word embedding makes SVM with word embeddings a more efficient choice for modeling the large scale social media dataset. For the CNN and RCNN models, they perform slightly better than most of the SVM models but still, the content information is insufficient to achieve a good performance on the Uns posts. As to adding comment information to these models, since the commenters do not always hold the same stance as the author, simply adding comments and post contents together merely adds noise to the model.
Among all UTCNN variations, we find that user information is most important, followed by topic and comment information. UTCNN without user information shows results similar to SVMs — it does well for Sup and Neu but detects no Uns. Its best f-scores on both Sup and Neu among all methods show that with enough training data, content-based models can perform well; at the same time, the lack of user information results in too few clues for minor-class posts to either predict their stance directly or link them to other users and posts for improved performance. The 17.5% improvement when adding user information suggests that user information is especially useful when the dataset is highly imbalanced. All models that consider user information predict the minority class successfully. UCTNN without topic information works well but achieves lower performance than the full UTCNN model. The 4.9% performance gain brought by LDA shows that although it is satisfactory for single topic datasets, adding that latent topics still benefits performance: even when we are discussing the same topic, we use different arguments and supporting evidence. Lastly, we get 4.8% improvement when adding comment information and it achieves comparable performance to UTCNN without topic information, which shows that comments also benefit performance. For platforms where user IDs are pixelated or otherwise hidden, adding comments to a text model still improves performance. In its integration of user, content, and comment information, the full UTCNN produces the highest f-scores on all Sup, Neu, and Uns stances among models that predict the Uns class, and the highest macro-average f-score overall. This shows its ability to balance a biased dataset and supports our claim that UTCNN successfully bridges content and user, topic, and comment information for stance classification on social media text. Another merit of UTCNN is that it does not require a balanced training data. This is supported by its outperforming other models though no oversampling technique is applied to the UTCNN related experiments as shown in this paper. Thus we can conclude that the user information provides strong clues and it is still rich even in the minority class.
We also investigate the semantic difference when a user acts as an author/liker or a commenter. We evaluated a variation in which all embeddings from the same user were forced to be identical (this is the UTCNN shared user embedding setting in Table TABREF22 ). This setting yielded only a 2.5% improvement over the model without comments, which is not statistically significant. However, when separating authors/likers and commenters embeddings (i.e., the UTCNN full model), we achieved much greater improvements (4.8%). We attribute this result to the tendency of users to use different wording for different roles (for instance author vs commenter). This is observed when the user, acting as an author, attempts to support her argument against nuclear power by using improvements in solar power; when acting as a commenter, though, she interacts with post contents by criticizing past politicians who supported nuclear power or by arguing that the proposed evacuation plan in case of a nuclear accident is ridiculous. Based on this finding, in the final UTCNN setting we train two user matrix embeddings for one user: one for the author/liker role and the other for the commenter role.
Results on CreateDebate Dataset
Table TABREF24 shows the results of UTCNN, baselines as we implemented on the FBFans datset and related work on the CreateDebate dataset. We do not adopt oversampling on these models because the CreateDebate dataset is almost balanced. In previous work, integer linear programming (ILP) or linear-chain conditional random fields (CRFs) were proposed to integrate text features, author, ideology, and user-interaction constraints, where text features are unigram, bigram, and POS-dependencies; the author constraint tends to require that posts from the same author for the same topic hold the same stance; the ideology constraint aims to capture inferences between topics for the same author; the user-interaction constraint models relationships among posts via user interactions such as replies BIBREF7 , BIBREF9 .
The SVM with n-gram or average word embedding feature performs just similar to the majority. However, with the transformed word embedding, it achieves superior results. It shows that the learned user and topic embeddings really capture the user and topic semantics. This finding is not so obvious in the FBFans dataset and it might be due to the unfavorable data skewness for SVM. As for CNN and RCNN, they perform slightly better than most SVMs as we found in Table TABREF22 for FBFans.
Compared to the ILP BIBREF7 and CRF BIBREF9 methods, the UTCNN user embeddings encode author and user-interaction constraints, where the ideology constraint is modeled by the topic embeddings and text features are modeled by the CNN. The significant improvement achieved by UTCNN suggests the latent representations are more effective than overt model constraints.
The PSL model BIBREF12 jointly labels both author and post stance using probabilistic soft logic (PSL) BIBREF23 by considering text features and reply links between authors and posts as in Hasan and Ng's work. Table TABREF24 reports the result of their best AD setting, which represents the full joint stance/disagreement collective model on posts and is hence more relevant to UTCNN. In contrast to their model, the UTCNN user embeddings represent relationships between authors, but UTCNN models do not utilize link information between posts. Though the PSL model has the advantage of being able to jointly label the stances of authors and posts, its performance on posts is lower than the that for the ILP or CRF models. UTCNN significantly outperforms these models on posts and has the potential to predict user stances through the generated user embeddings.
For the CreateDebate dataset, we also evaluated performance when not using topic embeddings or user embeddings; as replies in this dataset are viewed as posts, the setting without comment embeddings is not available. Table TABREF24 shows the same findings as Table TABREF22 : the 21% improvement in accuracy demonstrates that user information is the most vital. This finding also supports the results in the related work: user constraints are useful and can yield 11.2% improvement in accuracy BIBREF7 . Further considering topic information yields 3.4% improvement, suggesting that knowing the subject of debates provides useful information. In sum, Table TABREF22 together with Table TABREF24 show that UTCNN achieves promising performance regardless of topic, language, data distribution, and platform.
Conclusion
We have proposed UTCNN, a neural network model that incorporates user, topic, content and comment information for stance classification on social media texts. UTCNN learns user embeddings for all users with minimum active degree, i.e., one post or one like. Topic information obtained from the topic model or the pre-defined labels further improves the UTCNN model. In addition, comment information provides additional clues for stance classification. We have shown that UTCNN achieves promising and balanced results. In the future we plan to explore the effectiveness of the UTCNN user embeddings for author stance classification.
Acknowledgements
Research of this paper was partially supported by Ministry of Science and Technology, Taiwan, under the contract MOST 104-2221-E-001-024-MY2.
Question:
How many layers does the UTCNN model have?
Answer:
| eight layers | 328 | 29,212 | 29,214 | 29,266 | ... [The rest of the article is omitted]
|
58ef2442450c392bfc55c4dc35f216542f5f2dbb | 58ef2442450c392bfc55c4dc35f216542f5f2dbb_0 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
The rapid growth in speech and small screen interfaces, particularly on mobile devices, has significantly influenced the way users interact with intelligent systems to satisfy their information needs. The growing interest in personal digital assistants, such as Amazon Alexa, Apple Siri, Google Assistant, and Microsoft Cortana, demonstrates the willingness of users to employ conversational interactions BIBREF0. As a result, conversational information seeking (CIS) has been recognized as a major emerging research area in the Third Strategic Workshop on Information Retrieval (SWIRL 2018) BIBREF1.
Research progress in CIS relies on the availability of resources to the community. There have been recent efforts on providing data for various CIS tasks, such as the TREC 2019 Conversational Assistance Track (CAsT), MISC BIBREF2, Qulac BIBREF3, CoQA BIBREF4, QuAC BIBREF5, SCS BIBREF6, and CCPE-M BIBREF7. In addition, BIBREF8 have implemented a demonstration for conversational movie recommendation based on Google's DialogFlow. Despite all of these resources, the community still feels the lack of a suitable platform for developing CIS systems. We believe that providing such platform will speed up the progress in conversational information seeking research. Therefore, we developed a general framework for supporting CIS research. The framework is called Macaw. This paper describes the high-level architecture of Macaw, the supported functionality, and our future vision. Researchers working on various CIS tasks should be able to take advantage of Macaw in their projects.
Macaw is designed based on a modular architecture to support different information seeking tasks, including conversational search, conversational question answering, conversational recommendation, and conversational natural language interface to structured and semi-structured data. Each interaction in Macaw (from both user and system) is a Message object, thus a conversation is a list of Messages. Macaw consists of multiple actions, each action is a module that can satisfy the information needs of users for some requests. For example, search and question answering can be two actions in Macaw. Even multiple search algorithms can be also seen as multiple actions. Each action can produce multiple outputs (e.g., multiple retrieved documents). For every user interaction, Macaw runs all actions in parallel. The actions' outputs produced within a predefined time interval (i.e., an interaction timeout constant) are then post-processed. Macaw can choose one or combine multiple of these outputs and prepare an output Message object as the user's response.
The modular design of Macaw makes it relatively easy to configure a different user interface or add a new one. The current implementation of Macaw supports a command line interface as well as mobile, desktop, and web apps. In more detail, Macaw's interface can be a Telegram bot, which supports a wide range of devices and operating systems (see FIGREF4). This allows Macaw to support multi-modal interactions, such as text, speech, image, click, etc. A number of APIs for automatic speech recognition and generation have been employed to support speech interactions. Note that the Macaw's architecture and implementation allows mixed-initiative interactions.
The research community can benefit from Macaw for the following purposes:
[leftmargin=*]
Developing algorithms, tools, and techniques for CIS.
Studying user interactions with CIS systems.
Performing CIS studies based on an intermediary person and wizard of oz.
Preparing quick demonstration for a developed CIS model.
Macaw Architecture
Macaw has a modular design, with the goal of making it easy to configure and add new modules such as a different user interface or different retrieval module. The overall setup also follows a Model-View-Controller (MVC) like architecture. The design decisions have been made to smooth the Macaw's adoptions and extensions. Macaw is implemented in Python, thus machine learning models implemented using PyTorch, Scikit-learn, or TensorFlow can be easily integrated into Macaw. The high-level overview of Macaw is depicted in FIGREF8. The user interacts with the interface and the interface produces a Message object from the current interaction of user. The interaction can be in multi-modal form, such as text, speech, image, and click. Macaw stores all interactions in an “Interaction Database”. For every interaction, Macaw looks for most recent user-system interactions (including the system's responses) to create a list of Messages, called the conversation list. It is then dispatched to multiple information seeking (and related) actions. The actions run in parallel, and each should respond within a pre-defined time interval. The output selection component selects from (or potentially combines) the outputs generated by different actions and creates a Message object as the system's response. This message is logged into the interaction database and is sent to the interface to be presented to the user. Again, the response message can be multi-modal and include text, speech, link, list of options, etc.
Macaw also supports Wizard of Oz studies or intermediary-based information seeking studies. The architecture of Macaw for such setup is presented in FIGREF16. As shown in the figure, the seeker interacts with a real conversational interface that supports multi-modal and mixed-initiative interactions in multiple devices. The intermediary (or the wizard) receives the seeker's message and performs different information seeking actions with Macaw. All seeker-intermediary and intermediary-system interactions will be logged for further analysis. This setup can simulate an ideal CIS system and thus is useful for collecting high-quality data from real users for CIS research.
Retrieval and Question Answering in Macaw
The overview of retrieval and question answering actions in Macaw is shown in FIGREF17. These actions consist of the following components:
[leftmargin=*]
Co-Reference Resolution: To support multi-turn interactions, it is sometimes necessary to use co-reference resolution techniques for effective retrieval. In Macaw, we identify all the co-references from the last request of user to the conversation history. The same co-reference resolution outputs can be used for different query generation components. This can be a generic or action-specific component.
Query Generation: This component generates a query based on the past user-system interactions. The query generation component may take advantage of co-reference resolution for query expansion or re-writing.
Retrieval Model: This is the core ranking component that retrieves documents or passages from a large collection. Macaw can retrieve documents from an arbitrary document collection using the Indri python interface BIBREF9, BIBREF10. We also provide the support for web search using the Bing Web Search API. Macaw also allows multi-stage document re-ranking.
Result Generation: The retrieved documents can be too long to be presented using some interfaces. Result generation is basically a post-processing step ran on the retrieved result list. In case of question answering, it can employ answer selection or generation techniques, such as machine reading comprehension models. For example, Macaw features the DrQA model BIBREF11 for question answering.
These components are implemented in a generic form, so researchers can easily replace them with their own favorite algorithms.
User Interfaces
We have implemented the following interfaces for Macaw:
[leftmargin=*]
File IO: This interface is designed for experimental purposes, such as evaluating the performance of a conversational search technique on a dataset with multiple queries. This is not an interactive interface.
Standard IO: This interactive command line interface is designed for development purposes to interact with the system, see the logs, and debug or improve the system.
Telegram: This interactive interface is designed for interaction with real users (see FIGREF4). Telegram is a popular instant messaging service whose client-side code is open-source. We have implemented a Telegram bot that can be used with different devices (personal computers, tablets, and mobile phones) and different operating systems (Android, iOS, Linux, Mac OS, and Windows). This interface allows multi-modal interactions (text, speech, click, image). It can be also used for speech-only interactions. For speech recognition and generation, Macaw relies on online APIs, e.g., the services provided by Google Cloud and Microsoft Azure. In addition, there exist multiple popular groups and channels in Telegram, which allows further integration of social networks with conversational systems. For example, see the Naseri and Zamani's study on news popularity in Telegram BIBREF12.
Similar to the other modules, one can easily extend Macaw using other appropriate user interfaces.
Limitations and Future Work
The current implementation of Macaw lacks the following actions. We intend to incrementally improve Macaw by supporting more actions and even more advanced techniques for the developed actions.
[leftmargin=*]
Clarification and Preference Elicitation: Asking clarifying questions has been recently recognized as a necessary component in a conversational system BIBREF3, BIBREF7. The authors are not aware of a published solution for generating clarifying questions using public resources. Therefore, Macaw does not currently support clarification.
Explanation: Despite its importance, result list explanation is also a relatively less explored topic. We intend to extend Macaw with result list explanation as soon as we find a stable and mature solution.
Recommendation: In our first release, we focus on conversational search and question answering tasks. We intend to provide support for conversational recommendation, e.g., BIBREF13, BIBREF14, BIBREF15, and joint search and recommendation, e.g., BIBREF16, BIBREF17, in the future.
Natural Language Interface: Macaw can potentially support access to structured data, such as knowledge graph. We would like to ease conversational natural language interface to structured and semi-structured data in our future releases.
Contribution
Macaw is distributed under the MIT License. We welcome contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com. This project has adopted the Microsoft Open Source Code of Conduct.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
Conclusions
This paper described Macaw, an open-source platform for conversational information seeking research. Macaw supports multi-turn, multi-modal, and mixed-initiative interactions. It was designed based on a modular architecture that allows further improvements and extensions. Researchers can benefit from Macaw for developing algorithms and techniques for conversational information seeking research, for user studies with different interfaces, for data collection from real users, and for preparing a demonstration of a CIS model.
Acknowledgements
The authors wish to thank Ahmed Hassan Awadallah, Krisztian Balog, and Arjen P. de Vries for their invaluable feedback.
Question:
Does the paper provide any case studies to illustrate how one can use Macaw for CIS research?
Answer:
| No | 328 | 11,979 | 11,981 | 12,084 | ... [The rest of the article is omitted]
|
58ef2442450c392bfc55c4dc35f216542f5f2dbb | 58ef2442450c392bfc55c4dc35f216542f5f2dbb_1 | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
Article:
Introduction
The rapid growth in speech and small screen interfaces, particularly on mobile devices, has significantly influenced the way users interact with intelligent systems to satisfy their information needs. The growing interest in personal digital assistants, such as Amazon Alexa, Apple Siri, Google Assistant, and Microsoft Cortana, demonstrates the willingness of users to employ conversational interactions BIBREF0. As a result, conversational information seeking (CIS) has been recognized as a major emerging research area in the Third Strategic Workshop on Information Retrieval (SWIRL 2018) BIBREF1.
Research progress in CIS relies on the availability of resources to the community. There have been recent efforts on providing data for various CIS tasks, such as the TREC 2019 Conversational Assistance Track (CAsT), MISC BIBREF2, Qulac BIBREF3, CoQA BIBREF4, QuAC BIBREF5, SCS BIBREF6, and CCPE-M BIBREF7. In addition, BIBREF8 have implemented a demonstration for conversational movie recommendation based on Google's DialogFlow. Despite all of these resources, the community still feels the lack of a suitable platform for developing CIS systems. We believe that providing such platform will speed up the progress in conversational information seeking research. Therefore, we developed a general framework for supporting CIS research. The framework is called Macaw. This paper describes the high-level architecture of Macaw, the supported functionality, and our future vision. Researchers working on various CIS tasks should be able to take advantage of Macaw in their projects.
Macaw is designed based on a modular architecture to support different information seeking tasks, including conversational search, conversational question answering, conversational recommendation, and conversational natural language interface to structured and semi-structured data. Each interaction in Macaw (from both user and system) is a Message object, thus a conversation is a list of Messages. Macaw consists of multiple actions, each action is a module that can satisfy the information needs of users for some requests. For example, search and question answering can be two actions in Macaw. Even multiple search algorithms can be also seen as multiple actions. Each action can produce multiple outputs (e.g., multiple retrieved documents). For every user interaction, Macaw runs all actions in parallel. The actions' outputs produced within a predefined time interval (i.e., an interaction timeout constant) are then post-processed. Macaw can choose one or combine multiple of these outputs and prepare an output Message object as the user's response.
The modular design of Macaw makes it relatively easy to configure a different user interface or add a new one. The current implementation of Macaw supports a command line interface as well as mobile, desktop, and web apps. In more detail, Macaw's interface can be a Telegram bot, which supports a wide range of devices and operating systems (see FIGREF4). This allows Macaw to support multi-modal interactions, such as text, speech, image, click, etc. A number of APIs for automatic speech recognition and generation have been employed to support speech interactions. Note that the Macaw's architecture and implementation allows mixed-initiative interactions.
The research community can benefit from Macaw for the following purposes:
[leftmargin=*]
Developing algorithms, tools, and techniques for CIS.
Studying user interactions with CIS systems.
Performing CIS studies based on an intermediary person and wizard of oz.
Preparing quick demonstration for a developed CIS model.
Macaw Architecture
Macaw has a modular design, with the goal of making it easy to configure and add new modules such as a different user interface or different retrieval module. The overall setup also follows a Model-View-Controller (MVC) like architecture. The design decisions have been made to smooth the Macaw's adoptions and extensions. Macaw is implemented in Python, thus machine learning models implemented using PyTorch, Scikit-learn, or TensorFlow can be easily integrated into Macaw. The high-level overview of Macaw is depicted in FIGREF8. The user interacts with the interface and the interface produces a Message object from the current interaction of user. The interaction can be in multi-modal form, such as text, speech, image, and click. Macaw stores all interactions in an “Interaction Database”. For every interaction, Macaw looks for most recent user-system interactions (including the system's responses) to create a list of Messages, called the conversation list. It is then dispatched to multiple information seeking (and related) actions. The actions run in parallel, and each should respond within a pre-defined time interval. The output selection component selects from (or potentially combines) the outputs generated by different actions and creates a Message object as the system's response. This message is logged into the interaction database and is sent to the interface to be presented to the user. Again, the response message can be multi-modal and include text, speech, link, list of options, etc.
Macaw also supports Wizard of Oz studies or intermediary-based information seeking studies. The architecture of Macaw for such setup is presented in FIGREF16. As shown in the figure, the seeker interacts with a real conversational interface that supports multi-modal and mixed-initiative interactions in multiple devices. The intermediary (or the wizard) receives the seeker's message and performs different information seeking actions with Macaw. All seeker-intermediary and intermediary-system interactions will be logged for further analysis. This setup can simulate an ideal CIS system and thus is useful for collecting high-quality data from real users for CIS research.
Retrieval and Question Answering in Macaw
The overview of retrieval and question answering actions in Macaw is shown in FIGREF17. These actions consist of the following components:
[leftmargin=*]
Co-Reference Resolution: To support multi-turn interactions, it is sometimes necessary to use co-reference resolution techniques for effective retrieval. In Macaw, we identify all the co-references from the last request of user to the conversation history. The same co-reference resolution outputs can be used for different query generation components. This can be a generic or action-specific component.
Query Generation: This component generates a query based on the past user-system interactions. The query generation component may take advantage of co-reference resolution for query expansion or re-writing.
Retrieval Model: This is the core ranking component that retrieves documents or passages from a large collection. Macaw can retrieve documents from an arbitrary document collection using the Indri python interface BIBREF9, BIBREF10. We also provide the support for web search using the Bing Web Search API. Macaw also allows multi-stage document re-ranking.
Result Generation: The retrieved documents can be too long to be presented using some interfaces. Result generation is basically a post-processing step ran on the retrieved result list. In case of question answering, it can employ answer selection or generation techniques, such as machine reading comprehension models. For example, Macaw features the DrQA model BIBREF11 for question answering.
These components are implemented in a generic form, so researchers can easily replace them with their own favorite algorithms.
User Interfaces
We have implemented the following interfaces for Macaw:
[leftmargin=*]
File IO: This interface is designed for experimental purposes, such as evaluating the performance of a conversational search technique on a dataset with multiple queries. This is not an interactive interface.
Standard IO: This interactive command line interface is designed for development purposes to interact with the system, see the logs, and debug or improve the system.
Telegram: This interactive interface is designed for interaction with real users (see FIGREF4). Telegram is a popular instant messaging service whose client-side code is open-source. We have implemented a Telegram bot that can be used with different devices (personal computers, tablets, and mobile phones) and different operating systems (Android, iOS, Linux, Mac OS, and Windows). This interface allows multi-modal interactions (text, speech, click, image). It can be also used for speech-only interactions. For speech recognition and generation, Macaw relies on online APIs, e.g., the services provided by Google Cloud and Microsoft Azure. In addition, there exist multiple popular groups and channels in Telegram, which allows further integration of social networks with conversational systems. For example, see the Naseri and Zamani's study on news popularity in Telegram BIBREF12.
Similar to the other modules, one can easily extend Macaw using other appropriate user interfaces.
Limitations and Future Work
The current implementation of Macaw lacks the following actions. We intend to incrementally improve Macaw by supporting more actions and even more advanced techniques for the developed actions.
[leftmargin=*]
Clarification and Preference Elicitation: Asking clarifying questions has been recently recognized as a necessary component in a conversational system BIBREF3, BIBREF7. The authors are not aware of a published solution for generating clarifying questions using public resources. Therefore, Macaw does not currently support clarification.
Explanation: Despite its importance, result list explanation is also a relatively less explored topic. We intend to extend Macaw with result list explanation as soon as we find a stable and mature solution.
Recommendation: In our first release, we focus on conversational search and question answering tasks. We intend to provide support for conversational recommendation, e.g., BIBREF13, BIBREF14, BIBREF15, and joint search and recommendation, e.g., BIBREF16, BIBREF17, in the future.
Natural Language Interface: Macaw can potentially support access to structured data, such as knowledge graph. We would like to ease conversational natural language interface to structured and semi-structured data in our future releases.
Contribution
Macaw is distributed under the MIT License. We welcome contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com. This project has adopted the Microsoft Open Source Code of Conduct.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
Conclusions
This paper described Macaw, an open-source platform for conversational information seeking research. Macaw supports multi-turn, multi-modal, and mixed-initiative interactions. It was designed based on a modular architecture that allows further improvements and extensions. Researchers can benefit from Macaw for developing algorithms and techniques for conversational information seeking research, for user studies with different interfaces, for data collection from real users, and for preparing a demonstration of a CIS model.
Acknowledgements
The authors wish to thank Ahmed Hassan Awadallah, Krisztian Balog, and Arjen P. de Vries for their invaluable feedback.
Question:
Does the paper provide any case studies to illustrate how one can use Macaw for CIS research?
Answer:
| No | 328 | 11,979 | 11,981 | 12,084 | ... [The rest of the article is omitted]
|
290ee79b5e3872e0496a6a0fc9b103ab7d8f6c30 | 290ee79b5e3872e0496a6a0fc9b103ab7d8f6c30_0 | "You are given a scientific article and a question. Answer the question as concisely as you can, usi(...TRUNCATED) | "Level A: Offensive language Detection\n, Level B: Categorization of Offensive Language\n, Level C: (...TRUNCATED) | 328 | 14,920 | 14,922 | 14,983 | ... [The rest of the article is omitted]
|
End of preview. Expand
in Data Studio
No dataset card yet
- Downloads last month
- 5