Tamir39/causasent-ate-v2

CausaSent ATE v2 — Vietnamese E-commerce Aspect Term Extraction + ABSA

Vietnamese e-commerce reviews labeled with aspect terms (span), aspect category, and binary sentiment — combined from a small manually-annotated gold pool (hotel + restaurant + smartphone reviews) and a larger LLM-labeled silver pool (Tiki product reviews).

Schema

Each row in train.json / val.json / test.json is:

{
  "id": "tiki-14075703",
  "source": "tiki",
  "review": "Hàng đóng gói không cẩn thận, hộp méo mó, rách nhiều.",
  "annotations": [
    {
      "aspect_term": "Hàng",
      "aspect_term_span": [0, 4],
      "aspect_category": "appearance",
      "sentiment": "negative"
    },
    {
      "aspect_term": "đóng gói",
      "aspect_term_span": [5, 13],
      "aspect_category": "packaging",
      "sentiment": "negative"
    }
  ]
}

Hard invariants:

• review[aspect_term_span[0]:aspect_term_span[1]] == aspect_term byte-for-byte.
• aspect_category ∈ {delivery, packaging, product_quality, price, customer_service, usability, appearance} — closed 7-class taxonomy.
• sentiment ∈ {positive, negative} — neutral was dropped (only 2.7% of gold, too few to train).
• source ∈ {gold, tiki}.

Splits

Split	Reviews	Annotations
train	5,647	8,251
val	700	1,033
test	719	1,023
total	7,066	10,307

Stratified 80/10/10 by dominant (aspect_category, sentiment) per review.

Aspect coverage in train (annotation count):

aspect_category	count
appearance	2,789
product_quality	2,133
delivery	982
packaging	896
price	922
customer_service	424
usability	105

Sentiment leans negative for appearance / packaging (Tiki complaints) and positive for price / customer_service.

Data sources

• Gold (1,132 reviews, 2,205 annotations): subset of CausaSent original manual annotations (drop neutral, drop cause_span/action fields). Schema converted via Claude-assisted aspect-term span extraction; 3 NFD-encoding edge cases dropped. See Tamir39/causasent for the pre-pivot 4-tuple version.
• Tiki silver (5,934 reviews, 8,102 annotations): Vietnamese Tiki product reviews originally labeled with aspect_category + sentiment by an LLM (mid-term project); we added aspect_term_span via the same Claude pipeline. 14 records dropped on span verification.

Annotation pipeline

Both sources passed through the same Claude-based annotation pipeline (scripts/annotate_aspect_terms.py orchestrator → scripts/SUBAGENT_PROMPT.md per-batch worker → scripts/verify_annotation_batch.py byte-match check → scripts/merge_annotation_batches.py). 99.86% of gold and 99.83% of Tiki survived span verification.

A stratified sample of 300 gold records was independently re-validated; inter-annotator agreement ≈ 91% (judged term matches LLM's pick).

Intended use

Joint aspect-term extraction + aspect-category-aware sentiment classification on Vietnamese e-commerce text. Suitable for:

• Training PhoBERT/XLM-R BIO span taggers with auxiliary sentiment head.
• Comparing fine-tuned models vs. zero-shot LLM baselines on Vietnamese ATE.
• ABSA-informed action recommendation pipelines (per the paper this dataset accompanies).

License & ToS

Released under CC-BY-SA-4.0. The Tiki subset comes from publicly-displayed product reviews; we redistribute only the text and our derived annotations. If you use this dataset, please cite the original CausaSent project.