alvations/poliinformatics-2014

PoliInformatics 2014 — NLP Unshared Task

Modernised release of the PoliInformatics 2014 Unshared Task dataset — a corpus of US Congressional and Federal Reserve documents about the 2008 financial crisis, collected by John Wilkerson (UW) and released by Noah Smith (CMU).

Original distribution: https://sites.google.com/site/unsharedtask2014/data.

This release streams the upstream tarball (unshared-task-poliinformatics-2014-v1.0.tgz, 222 MB; mirrored at alvations/stash) directly into parquet — no on-disk extraction — and exposes both the cleaned text (one config per subcorpus) and the raw binary blobs of every original file as a separate config.

Configs

Config	Granularity	Source	Rows	Parquet
congressional-bills	one section per row	tab-delimited .txt from each bill	7,184	9.3 MB
congressional-hearings	one utterance per row	per-hearing .csv (Speaker, Speech)	27,406	14.4 MB
fomc	one utterance per row	per-meeting .csv (Discourse #, Speaker, Speech)	6,287	3.8 MB
congressional-reports	one report per row (HTML pages stripped + concatenated)	raw-pages/*.html	2	1.2 MB
fcic	one FCIC report per row (HTML pages stripped + concatenated)	raw-pages/*.html	2	1.4 MB
raw	one source file per row, binary blob	every PDF / HTML / CSV / TXT in the tarball	13,565	368 MB

Usage

from datasets import load_dataset

# utterance-level FOMC transcripts (Bernanke, Kohn, et al. on the financial crisis)
fomc = load_dataset("alvations/poliinformatics-2014", "fomc", split="train")
print(fomc[0]["speaker"], fomc[0]["text"][:120])
# CHAIRMAN BERNANKE.   Good early morning.  [Laughter]  I would like to start by thanking the Committee for the very useful

# TARP hearings — utterance-level
hearings = load_dataset("alvations/poliinformatics-2014", "congressional-hearings", split="train")

# Bills — section-level (Dodd-Frank, ARRA, Gramm-Leach-Bliley, 110thALL, 111thALL etc.)
bills = load_dataset("alvations/poliinformatics-2014", "congressional-bills", split="train")
print([d for d in set(bills["doc_id"])])

# Raw originals (PDF / HTML / CSV bytes) for users who want to redo the extraction
raw = load_dataset("alvations/poliinformatics-2014", "raw", split="train")
print(raw[0]["file_path"], raw[0]["file_type"], raw[0]["size_bytes"])

Schemas

congressional-bills

Column	Type	Description
subcorpus	string	always "congressional-bills"
doc_id	string	bill name (e.g. DoddFrank, ARRA, GrammLeachBliley, 111thALL)
congress	int32	Congress number (e.g. 111)
chamber	string	hr / s / hres / sres / hjres / sjres / hconres / sconres
bill_no	string	bill number within the chamber
version	string	bill version code (e.g. rfs, eas, enr)
full_id	string	concatenated bill identifier (e.g. hr4173-111-eas)
page_no	int32	section count for this bill version
date	string	issued-on date (M/D/YY)
section_idx	int32	section index within the bill
char_count	int32	character length of the section text
text	string	the section text

congressional-hearings / fomc (utterance-level)

Column	Type	Description
subcorpus	string	"congressional-hearings" or "fomc"
doc_id	string	hearing / meeting identifier (e.g. CHRG-110shrg50420, FOMC20071211meeting)
topic	string	only set for hearings (e.g. TARP); empty for FOMC
speaker	string	speaker label as it appears in the transcript (e.g. CHAIRMAN BERNANKE., MR. KOHN.)
utterance_idx	int32	row order within the source CSV
discourse_num	int32	discourse number when present (FOMC); −1 otherwise
text	string	the utterance text

congressional-reports / fcic (document-level)

Built by collecting all raw-pages/*.html (or <basename>-raw/*.html) files per report, BeautifulSoup-stripping the markup, and concatenating in page order.

Column	Type	Description
subcorpus	string	"congressional-reports" or "fcic"
doc_id	string	report folder basename
title	string	parent topic folder (e.g. AnatomyofCollapse, StockMarketPlunge, FCICFinalReport)
n_pages	int32	number of HTML pages concatenated
text	string	cleaned plain-text body

raw

Column	Type	Description
subcorpus	string	the upstream top-level folder name
doc_id	string	the second-level folder (or filename for bills)
file_path	string	full path within the tarball
file_type	string	extension (pdf, html, htm, csv, txt, py, …)
size_bytes	int64	file size
bytes	binary	raw file contents

Build pipeline

build.py streams the source tarball via tarfile.ZipFile.open(mode="r:gz") and:

1. For each *.csv in CongressionalHearings / FOMC — parse with csv.DictReader (after stripping NULs and lifting the field-size cap to handle very long speeches) and emit one row per utterance.
2. For each bill .txt — regex-scan for record anchors (\d+\t\d+\t\d+\t\d+\t[a-z]+\t...) since the source records are concatenated without line breaks and contain embedded quotes/newlines. Emit one row per section.
3. For each report / FCIC folder containing *raw*/*.html — collect the page bytes, BS-strip in page order, and emit one row per report.
4. For every regular file in the tarball — emit one row in the raw config with the bytes column.

No PDFs are parsed; the HTML page extracts are the cleaner upstream form.

Provenance

• Upstream collected primarily by John Wilkerson (jwilker@uw.edu) and released by Noah Smith (nasmith@cs.cmu.edu).
• Upstream README: https://sites.google.com/site/unsharedtask2014/data
• Contact: unsharedtask@gmail.com
• Source mirror: alvations/stash · poliinformatics-2014/

Citation

The official task description and discussion of the data live at the Unshared Task 2014 site. The accompanying overview paper:

@inproceedings{smith-etal-2014-poliinformatics,
  title     = {Overview of the 2014 {NLP} Unshared Task in {P}oli{I}nformatics},
  author    = {Smith, Noah A. and Cardie, Claire and Washington, Anne L. and Wilkerson, John D.},
  booktitle = {Proceedings of the {ACL} 2014 Workshop on Language Technologies and Computational Social Science},
  pages     = {5--7},
  year      = {2014},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/W14-2502/},
}

License

The upstream release does not attach an explicit redistribution license, but the source materials are all US Government documents (Congressional bills, hearings, reports; Federal Reserve / FOMC meeting transcripts; FCIC reports) which are in the public domain in the United States. Reuse should still cite the PoliInformatics organisers and acknowledge the original collectors.