alunxu/nyt-iran-coverage

NYT Iran Coverage Corpus (1970-2026)

A corpus of 46,512 New York Times articles about Iran, spanning 1970-01-06 to 2026-01-07. The dataset was curated from monthly NYT Archive API dumps, enriched with full article text recovered from archived NYT web pages through the Internet Archive Wayback Machine, and post-processed with analysis-ready fields for framing research.

Built for the DH-412 (History and the Digital, EPFL) project “A Multi-Layered and Multi-Voice Analysis of Iran's Representation in The New York Times.” The expanded version adds 2,904 pre-1979 articles, allowing cautious comparison between pre-revolution and post-revolution framing.

Files

Path	Description
iran_articles_full.parquet	Recommended file. 1970-2026 corpus with NYT metadata, recovered fulltext, voice/byline fields, component word counts, and other derived analysis columns.
iran_articles_full.csv	CSV mirror of the recommended dataset. Larger and slower to load, but convenient for manual inspection.
iran_articles.parquet	Metadata-only Iran corpus before full-text/enrichment merge.
iran_articles.csv	CSV mirror of the metadata-only corpus.
lexicons.json	Versioned project lexicons: threat, diplomacy, humanizing, source-attribution verbs, and Iran-proxy relevance terms.
images.tar.gz	Auxiliary NYT article-image archive from the image-download pipeline. Contains images/{md5(uri)[:12]}_{idx}.jpg; cross-reference via uri and multimedia_json.

Quick load:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/alunxu/nyt-iran-coverage/iran_articles_full.parquet"
)

For body-text analysis, use:

body_df = df[df["fulltext_word_count"] > 0].copy()

Coverage

33,275 / 46,512 articles (71.5%) have non-empty recovered body text. All records retain NYT API metadata such as headline, abstract, lead paragraph, section, genre, keywords, and publication date.

Decade	Articles	With full text	Coverage
1970s	5,189	857	16.5%
1980s	12,004	9,492	79.1%
1990s	9,001	8,482	94.2%
2000s	9,678	5,711	59.0%
2010s	7,182	5,508	76.7%
2020s	3,402	3,220	94.7%

Pre-1979 coverage is intentionally reported separately because it is substantively important but technically sparse: 225 / 2,904 pre-1979 articles (7.7%) have recovered body text. Pre-revolution comparisons should therefore prioritize metadata-level fields (headline, abstract, lead_paragraph, keywords_json, section_name, type_of_material, voice) and treat body-level evidence as partial.

Voice distribution

The voice field groups NYT material types into broad analytical voices. It is derived from NYT type_of_material, with news_desk used as a fallback.

Voice	Count
news	38,624
column	2,422
other	1,580
letter	1,493
editorial	1,180
review	883
blog	330

How the Dataset Was Built

1. Archive collection. Monthly NYT Archive API dumps were collected for 1970-2026. These raw monthly files contain NYT metadata for all articles in each month, not only Iran-related articles.

2. Iran filtering. A recall-oriented Iran filter was applied to NYT keyword metadata plus high-salience text fields: headline, abstract, lead paragraph, and snippet. The filter includes literal Iran terms (Iran, Iranian, Iranians) and proxy terms such as Tehran, Khomeini, Shah, Ayatollah, Persian, and major Iranian political figures or institutions. This stage constructs the corpus; it is not itself a framing classifier.

3. Full-text recovery. The NYT Archive API usually provides metadata, abstracts, and lead paragraphs, but not complete article bodies. Full text was recovered by querying archived versions of each web_url through the Wayback Machine and extracting main article text with trafilatura. The scraper uses checkpointed asynchronous requests, timestamp fallbacks, CDX snapshot discovery, and checks for short pages, duplicates, and paywall/subscription pages.

4. Post-processing and enrichment. Recovered texts were merged back by NYT uri. Boilerplate such as leading Advertisement markers was cleaned. Analysis-ready fields were then derived: voice, decade, iran_mention_count, component word counts, and byline fields recovered from the raw NYT Archive API JSON.

Column Dictionary

NYT API metadata

Column	Source / Derivation	Description
headline	NYT Archive API headline.main	Main article headline.
abstract	NYT Archive API abstract	API-provided abstract/summary. Length varies; some records contain unusually long text.
lead_paragraph	NYT Archive API lead_paragraph	Opening paragraph when available.
snippet	NYT Archive API snippet	Short display/search snippet.
pub_date	NYT Archive API pub_date	Publication timestamp in UTC.
section_name	NYT Archive API section_name	Broad NYT section, e.g. World, Opinion, Arts.
news_desk	NYT Archive API news_desk	Desk or editorial unit. Used as fallback for voice.
document_type	NYT Archive API document_type	Document category, e.g. article, multimedia, blogpost.
type_of_material	NYT Archive API type_of_material	Genre/material label, e.g. News, Op-Ed, Editorial, Review, News Analysis.
web_url	NYT Archive API web_url	Canonical NYT URL used for Wayback lookup.
uri	NYT Archive API uri	Stable NYT identifier and primary merge key.
word_count	NYT Archive API word_count	NYT-provided article word count. May differ from recovered text length.
source	NYT Archive API source	Source label as provided by NYT.
keywords_json	NYT Archive API keywords, serialized	Original NYT keyword metadata. Useful for entity/topic validation.
multimedia_json	NYT Archive API multimedia, serialized	Original NYT multimedia metadata. Used to derive image indicators.

Derived metadata and multimodal fields

Column	Derivation	Description
has_image	Derived from multimedia_json	True when NYT metadata lists at least one multimedia/image entry. This is metadata-based, not computer vision.
n_images	Derived from multimedia_json	Count of NYT multimedia/image entries.
year	Derived from pub_date	Calendar year.
month	Derived from pub_date	Calendar month.
year_month	Derived from pub_date	Monthly period for time-series aggregation.
decade	Derived from year	Decade bucket, e.g. 1970, 1980, 1990.
voice	Derived from type_of_material; news_desk fallback	Broad analytical genre: news, editorial, column, review, letter, blog, or other.
byline_original	Recovered from raw NYT Archive API byline.original	Original byline string, e.g. By NICHOLAS KRISTOF, when available.
byline_lastname	Recovered from first structured NYT byline person	Primary listed author's surname when available. Useful for columnist/staff analysis.

Full-text and component fields

Column	Derivation	Description
fulltext	Wayback Machine snapshot + trafilatura extraction	Cleaned recovered article body. Empty string when unavailable.
scrape_status	Assigned by full-text scraper	Provenance/status of recovery attempt. See below.
fulltext_word_count	Recomputed from cleaned fulltext	Number of whitespace-delimited words in recovered body text. Use > 0 for body-level analysis.
iran_mention_count	Regex count in fulltext	Counts literal Iran terms and proxy terms such as Tehran, Shah, Khomeini, IRGC, JCPOA.
headline_word_count	Whitespace split of headline	Used for component-level density calculations.
lead_word_count	Whitespace split of lead_paragraph	Used for lead/body comparisons.
abstract_word_count	Whitespace split of abstract	Used for abstract/body comparisons.

Voice Derivation Logic

voice is intentionally a coarse analytical grouping rather than a perfect reproduction of NYT editorial categories. The logic is:

• Editorial → editorial
• Op-Ed → column
• Letter → letter
• Review → review
• Web Log → blog
• News, Archives, Summary, Text, Briefing, Front Page, News Analysis, An Analysis, An Appraisal → news
• Correction, Video, Obituary, Schedule, Biography, Addendum, Interview, List, Recipe, Premium → other
• missing or ambiguous values fall back to news_desk; otherwise the safe default is news

This makes the field suitable for high-level comparisons such as news vs. editorial vs. column, while avoiding over-interpretation of fine-grained NYT production labels.

Full-Text Recovery Status

Status	Meaning	Recommended Treatment
success	Full article body was extracted from an archived NYT page.	Use for body-level text analysis.
recovered	Full text was recovered in a later retry/recovery pass.	Treat like success; kept separate for provenance.
subscription	Page was marked as subscription/paywall-like but usable body text was present.	Inspect if needed; currently contains non-empty text.
not_archived	No usable Wayback snapshot was found.	Use metadata fields only.
failed	Snapshot or request existed, but extraction did not produce usable text.	Use metadata fields only.
too_short	Extracted text was below the minimum word threshold.	Treat as no reliable body text.
extraction_failed	HTML was retrieved but article extraction failed.	Use metadata fields only.
duplicate	Recovered body duplicated another extracted text.	Excluded from body-level analysis to avoid double-counting.

Lexicons

lexicons.json provides a versioned, auditable vocabulary layer for framing analysis:

Lexicon	Terms	Purpose
threat	68	Security, conflict, coercion, nuclear, sanctions, violence.
diplomacy	41	Negotiation, agreement, reform, compromise, engagement.
humanizing	54	Culture, everyday life, family, students, art, protest, dignity.
credible_verbs	10	Source-attribution verbs such as said, confirmed, explained.
dubious_verbs	14	Attribution verbs such as claimed, denied, vowed, warned.
iran_proxy_terms	31	Iran relevance terms beyond literal Iran, such as Tehran, IRGC, Khomeini, JCPOA.

These lexicons are calibrated against war/peace journalism indicators and are meant to be transparent enough for close reading and manual audit. They should be treated as measurement instruments, not as exhaustive dictionaries of meaning.

Recommended Research Uses

The dataset supports several levels of analysis:

• Textual layer tension: headline vs. abstract vs. lead vs. full article body.
• Voice tension: news vs. editorial vs. column vs. review.
• Multimodal metadata tension: text fields vs. image/multimedia availability.
• Temporal comparison: pre-revolution, revolution, hostage crisis, Iran-Iraq War, nuclear negotiations, sanctions periods, and post-2020 coverage.
• Attribution and agency: byline, voice, and source-verb asymmetry using byline_*, voice, and lexicons.json.

Known Limitations

• Full text is incomplete. The NYT Archive API does not provide full bodies; recovered text depends on Wayback availability and extractability.
• Pre-1979 body text is sparse. Only 225 of 2,904 pre-1979 articles have recovered body text. Treat full-body pre-revolution analysis as exploratory.
• 1978 API source gap. NYT Archive API dumps contain 0 articles for 1978-09 and 1978-10, and unusually few for 1978-08. This is a source/API gap, not an Iran-filtering result.
• Image fields are metadata. has_image and n_images do not classify image content.
• NYT keywords are institutional metadata. keywords_json reflects NYT indexing practices, not an independent topic model.
• Byline extraction is partial. byline_original exists for 35,161 articles; byline_lastname exists for 28,523 articles. Missing bylines are normal for wire, archive, unsigned, or structurally incomplete records.

Ethical and Legal Note

This dataset is intended for academic research and teaching. Metadata comes from the NYT Archive API, and full-text recovery relies on publicly archived web snapshots. Users should respect the New York Times terms of service, Internet Archive usage norms, copyright restrictions, and institutional requirements governing text and data mining.

Citation

Suggested citation:

Xu, Alun. NYT Iran Coverage Corpus (1970-2026). Hugging Face Datasets, 2026.
Data derived from the New York Times Archive API and archived NYT pages via the Internet Archive Wayback Machine.