Canuck4Frame

Media framing of 4chan in major Canadian news media (2017–2026).

A reproducible computational-social-science pipeline that identifies, quantifies, and tracks the media frames that ten major Canadian outlets — public broadcasters and leading private titles — use when they cover the platform 4chan, via unsupervised, transformer-based topic modeling (BERTopic).

Research question. How do Canadian outlets frame 4chan, and do those frames shift over time (e.g. from "meme factory" to "far-right engine")? Is there a difference between Francophone and Anglophone coverage?

See PROTOCOL.md for the full research protocol (sampling, coding scheme, statistics, reproducibility).

---

Data source: Media Cloud (CSV export)

Canuck4Frame builds its corpus from Media Cloud, an open research database that indexes the full text of millions of news stories worldwide, with deep, reliable per-outlet coverage.

Why not GDELT? The pipeline originally targeted GDELT DOC 2.0, but GDELT's index contains essentially no "4chan" articles for the target Canadian outlets (0 for cbc.ca, radio-canada.ca, ledevoir.com, tvanouvelles.ca). Media Cloud returns a real corpus (344 metadata rows across the ten outlets).

How the corpus is obtained

Media Cloud is queried manually through its web UI (no API key required), then exported to CSV:

1. Go to https://search.mediacloud.org/ and build a search:
   • Search phrases: 4chan OR 4chan.org
   • Sources: the ten outlets below (add each by domain)
   • Dates: 2017-01-01 → present
2. Export the results to CSV and drop the file in data/mediacloud/.
3. Point config.yaml → data_source.mediacloud_csv at that file.

The export contains metadata only — id, indexed_date, language, media_name, media_url, publish_date, title, url — not the article body. Bodies are then recovered in two passes:

1. Live fetch (fetch_body) — download each URL and extract the main text with trafilatura (falls back to a <p>-tag heuristic). It uses realistic browser headers, per-domain rate limiting (so high-volume domains like thestar.com stop returning 429), retries with backoff, and incremental JSONL saves. Pages that hard-fail (paywalls, 404/410 dead links, 403 blocks, tarpits) are written to data/raw/articles_missing.csv.
2. Wayback recovery (scripts/recover_wayback.py) — for every row in articles_missing.csv, look up the closest Internet Archive (Wayback Machine) snapshot near the publish date, fetch the raw archived page, and extract the body. Recovered articles are appended to the corpus with retrieved_via: "wayback"; whatever is still unrecoverable stays in articles_missing.csv. This step is idempotent and resumable.

On the reference export this yields 280 usable articles (221 EN / 59 FR) out of 344 — the remaining 64 exist neither on the live web nor in the archive. Notably, Wayback recovery is what unblocked TVA Nouvelles (0 → 10, hard 403 on the live site) and much of Toronto Star (paywalled).

The ten outlets

Outlet	Domain	Lang	Type
CBC News	cbc.ca	en	Public
Radio-Canada	ici.radio-canada.ca	fr	Public
The Globe and Mail	theglobeandmail.com	en	Major
National Post	nationalpost.com	en	Major
CTV News	ctvnews.ca	en	Major
Global News	globalnews.ca	en	Major
Toronto Star	thestar.com	en	Major
La Presse	lapresse.ca	fr	Major
Le Devoir	ledevoir.com	fr	Major
TVA Nouvelles	tvanouvelles.ca	fr	Major

---

Project structure

canuck4frame/
├── config.yaml               # single source of truth for all parameters
├── requirements.txt
├── Dockerfile / docker-compose.yml
├── .env.example
├── src/
│   ├── config.py             # load config.yaml + .env
│   ├── collect.py            # Media Cloud CSV import + body fetching
│   ├── preprocess.py         # clean, dedupe, language-split
│   ├── modeling.py           # BERTopic (unified or split FR/EN)
│   ├── stats.py              # chi-square FR vs EN frame independence
│   └── visualization.py      # temporal + comparative charts
├── notebooks/
│   ├── 01_data_collection.ipynb
│   ├── 02_preprocessing.ipynb
│   ├── 03_bertopic_modeling.ipynb
│   └── 04_analysis_visualization.ipynb
├── scripts/
│   ├── run_pipeline.py       # end-to-end CLI
│   └── recover_wayback.py    # Wayback Machine recovery of missing bodies
├── data/
│   ├── mediacloud/           # Media Cloud CSV export(s)
│   └── raw/ processed/ results/   # generated (git-ignored)
│       # raw/ holds articles_raw.jsonl (corpus) + articles_missing.csv (unrecovered)
└── tests/

---

Installation

git clone https://github.com/marcgarnier/canuck4frame.git
cd canuck4frame

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

The first BERTopic run downloads a sentence-transformer model (a few hundred MB) and NLTK stopwords.

---

Quick start

Make sure a Media Cloud CSV export sits in data/mediacloud/ and is referenced in config.yaml (see Data source above).

Run the whole pipeline (import corpus + fetch bodies, then analyze):

python scripts/run_pipeline.py --collect

Recover bodies that failed the live fetch, from the Wayback Machine (optional but recommended — run after --collect, before analysis):

python scripts/recover_wayback.py

Re-run analysis only, on an already-collected corpus:

python scripts/run_pipeline.py --skip-collect

Or step through interactively — run the notebooks in order (01 → 04).

With Docker:

docker compose run pipeline          # full pipeline
docker compose up jupyter            # Jupyter Lab at http://localhost:8888

Artifacts land in data/results/: topic_summary.csv, frames_over_time.png, temporal_analysis.csv, comparative_fr_en.png, chi2_fr_en_result.csv, topics_over_time.html, and a saved bertopic_model/.

---

Results (reference run)

BERTopic produced 21 topics over 270 articles, leaving 39 (14 %) as outliers. Hand-coding (data/results/frames_annotation.csv) collapses them into seven frames:

Frame	Articles
far_right_extremism	51
violence_crime	43
meme_internet_culture	37
incel_misogyny	29
conspiracy_qanon	25
platform_tech	23
politics_elections	23

Framing is event-driven and shifts over the decade (frames_over_time.png): violence/crime opens the period at 61 % of 2017 articles (Las Vegas), conspiracy peaks at 47 % in 2020 (QAnon × COVID), far-right extremism takes 55 % then 62 % in 2022–2023, politics 50 % in 2024 (US election), and internet culture returns at 73 % in 2026 (a 4chan-born horror film reaching cinemas).

French and English coverage differ significantly (comparative_fr_en.png): χ² = 21.39, p = 0.0016, Cramér's V = 0.26 (moderate effect).

Frame	EN	FR
meme_internet_culture	11 %	36 %
far_right_extremism	24 %	13 %
incel_misogyny	15 %	4 %
violence_crime	20 %	13 %
platform_tech	9 %	16 %

Anglophone outlets frame 4chan as a threat (extremism, incels, violence); Francophone outlets frame it as a cultural phenomenon. ⚠️ With 270 articles across 7 frames × 2 languages, >20 % of expected cell counts are < 5 — the χ² approximation is unreliable and the p-value should be read with caution (a Fisher exact test is the natural next step).

---

Methodology (summary)

1. Collect — import the Media Cloud CSV export, download each body with trafilatura + per-domain throttling and incremental saves, then recover failed URLs from the Wayback Machine (scripts/recover_wayback.py).
2. Preprocess — strip boilerplate, drop stubs (< min_chars) and duplicates, detect language per article (langdetect), split into EN/FR sub-corpora.
3. Model — BERTopic over multilingual sentence embeddings (paraphrase-multilingual-MiniLM-L12-v2) with UMAP + HDBSCAN. Stopwords (EN+FR) removed in the c-TF-IDF step, with discriminative terms (4chan, meme, …) protected.
4. Annotate — read the 10 most representative articles per topic and assign a human frame label (frames_annotation.csv).
5. Analyze — frame share per year (stacked area), topics_over_time, and a French-vs-English comparison with a χ² test of independence.

All knobs live in config.yaml. The full procedure — sampling frame, coding scheme, and statistics — is in PROTOCOL.md.

---

Tuning for small corpora

The corpus is small (a few hundred articles). If BERTopic returns mostly outliers (topic -1):

• lower modeling.umap.n_components → 2
• lower modeling.hdbscan.min_cluster_size → 3
• lower modeling.min_topic_size → 3
• optionally set modeling.nr_topics: auto to merge similar topics

---

Ethics & scope

This project analyzes media discourse about 4chan, not content produced by 4chan users. No personal data is collected; only published news articles (public URLs and their text) are processed. Respect each outlet's terms of use and robots.txt when fetching bodies, and keep the request delay in config.yaml conservative.

---

Limitations & future work

• Media Cloud coverage shapes the corpus: indexing depth varies by outlet and the "4chan" keyword is niche, so counts per outlet are uneven.
• Body-fetch attrition. Of 344 metadata rows, 280 yielded usable text (live fetch + Wayback); 64 are lost (dead links, hard blocks not archived). CBC in particular is thin (7/22) because its old article URLs are poorly archived — a residual bias in which outlets survive to modeling.
• Frame labels are interpretive; inter-annotator agreement would strengthen them.
• Future: supervised validation of frames, more outlets, sentiment/stance layering, and effect sizes (Cramér's V) reported alongside χ².

---

Testing

pip install pytest
pytest tests/

(tests/ covers the lightweight, non-ML pieces: text cleaning, language detection, and the FR/EN chi-square statistics.)

---

License

MIT.

Suggested venues for write-up: Canadian Journal of Communication, New Media & Society, Journal of Information Technology & Politics.