histlearn/notas-comunidade-ptbr

Community Notes BR — Enriched Portuguese Subset

A curated, enriched Portuguese-language subset of X (formerly Twitter) Community Notes, with topic modeling, named entity recognition, and reproduction of the official Matrix Factorization scoring.

Snapshot: 2026-04-07 of the official Birdwatch public dump. Coverage: 2021-10-13 → 2026-04-05. Notes: 142,448 — all in Portuguese (fastText lid.176, confidence ≥ 0.80, mean 0.98). Tweets: 90,822 unique. Entities: 536,373 across 22 named-entity types. Topics: 899 fine-grained, organized into 38 macrothemes.

Resumo em português

Subconjunto curado em PT-BR das Community Notes da X, enriquecido com modelagem de tópicos hierárquicos (BERTopic + LLM com gramática formal GBNF), reconhecimento de entidades nomeadas (GLiNER + regex + dicionários canônicos), e reprodução do algoritmo oficial de pontuação por fatoração matricial. Cobre 142k notas em PT (2021-10 a 2026-04), 90k tweets distintos, 536k entidades em 22 tipos, 899 tópicos finos agrupados em 38 macrotemas. Licença CC0; alinhada à licença da fonte original.

---

Files

File	Rows	Size	Granularity
notes_pt.parquet	142,448	35 MB	one row per note
entities.parquet	536,373	40 MB	one row per entity (long format)
topic_metadata.parquet	898	70 KB	one row per topic
macrotheme_metadata.parquet	38	6 KB	one row per macrotheme
taxonomy.json	—	—	entity-type definitions
cross_modal_validation/	—	—	quality assessment tables (see below)

Join key: noteId between notes_pt and entities. Tweet text is not redistributed (X ToS); use tweetId to hydrate via X API or consult the project's static catalog.

---

Schema

notes_pt.parquet

Column	Type	Description
noteId	string	Primary key (Community Notes note identifier).
tweetId	string	The tweet the note refers to.
note_author_participant_id	string	Anonymized note author ID (CN's own participantId).
created_at	timestamp	Note creation time (UTC, derived from createdAtMillis).
year_month	string	YYYY-MM for time-series convenience.
summary	string	Note text as published.
summary_clean	string	Whitespace-normalized, URL-stripped version (used for embeddings).
classification_algoritmo	string	MISINFORMED_OR_POTENTIALLY_MISLEADING or NOT_MISLEADING. From the original CN data.
idioma	string	Always pt (subset).
confianca_idioma	float	fastText lid.176.bin confidence, [0.80, 1.00].
status_final	string	Latest community status: CURRENTLY_RATED_HELPFUL, CURRENTLY_RATED_NOT_HELPFUL, NEEDS_MORE_RATINGS, or null.
consenso	string	4-class compaction: CRH / CRNH / NMR / Outro.
n_ratings	int	Total ratings the note received.
n_notas_no_tweet	int	Number of notes attached to the same tweet.
coreNoteIntercept	float32	Output of the official MF scoring algorithm; null when the note had insufficient ratings.
coreNoteFactor1	float32	First factor from MF; interpretable as a polarity axis.
topic_id	int	BERTopic cluster id; -1 denotes the outlier cluster (47,162 notes); null for 397 notes outside the topic-model run.
topic_name	string	LLM-generated short label (Marco-Mini-Instruct via GBNF grammar).
contexto_central	string	LLM-generated one-sentence description of the topic.
categoria_ampla	string	LLM-classified category (13 controlled values).
macrotheme_id	int	Macrotheme cluster id from hierarchical aggregation.
macrotheme_label	string	LLM-generated macrotheme label.
stratum	string	Sampling stratum for the hydrated tweet sub-sample (top_notas, bat_signals, aleatorio). Null for unsampled notes.

entities.parquet (long format)

Column	Type	Description
entity_id	string	<row_id>_<j> — unique per extraction.
noteId	string	FK to notes_pt.
tweetId	string	FK to notes_pt.
texto_entidade	string	Surface form as it appears in the note.
texto_canonico	string	Normalized form (lowercase + unidecode + dictionary lookup for programs/agencies/parties).
tipo_entidade	string	Canonical PT type (see taxonomy.json; 22 values).
label_modelo	string	Original GLiNER English label or regex pattern name.
inicio / fim	int	Character offsets in summary.
score	float	GLiNER confidence [0,1]; 0.94–1.00 for regex-extracted entities.
fonte_extracao	string	gliner | regex_url | regex_data | regex_percentual | regex_valor | regex_lei_norma | regex_processo.
contexto	string	±90-character window around the entity.
papel_no_texto	string	mencao (mentioned as object) or fonte_ou_evidencia (cited as source/evidence).
dominio	string	Canonicalized domain (URLs only).
topic_id / macrotheme_id	int	Denormalized for analysis convenience.

---

Construction methodology

Five-stage pipeline. Code is described in the accompanying paper and the project's static catalog; only the final artifacts are redistributed here.

1. ETL — bronze/silver/gold

The 5 official Birdwatch tables (notes, noteRatings, noteStatusHistory, userEnrollment, batSignals) are downloaded by snapshot date, hash-verified (MD5/SHA256), schema-drift-checked across shards, and materialized in DuckDB. Bronze keeps everything as VARCHAR with metadata; silver adds typed columns and snake_case names; gold exports compressed Parquet.

2. Portuguese filter

Each note's summary_clean (whitespace-normalized, URL-stripped via SQL) is classified by fastText lid.176.bin. Notes with idioma = pt and confianca_idioma ≥ 0.80 are retained, yielding 142,448 notes (mean confidence 0.98). A 300-note stratified sample (100 per confidence band: 0.80–0.90 / 0.90–0.95 / 0.95–1.00) was visually inspected and looked correct, although this was a quick, non-exhaustive check rather than a rigorous human evaluation.

3. Topic modeling

Notes are embedded with intfloat/multilingual-e5-large-instruct (1024-d) using a Portuguese instruction prompt that orients the embedding toward the factual matter under dispute, not the procedural surface. Clustering uses BERTopic (UMAP n_components=10, HDBSCAN min_cluster_size=30, min_samples=5) with parameters chosen by a DBCV grid search. Outliers are reduced via embedding similarity (threshold 0.975). Each topic is then labeled by Marco-Mini-Instruct (GGUF Q6_K) constrained by a GBNF grammar that forces a JSON object with three fields (contexto_central, categoria_ampla, rotulo_curto) — this guarantees parseable output and a controlled category vocabulary.

Macrothemes are obtained by hierarchical aggregation (average linkage on cosine similarity of topic embeddings, distance threshold 0.05). Each macrotheme is then labeled by a second LLM pass that receives the contexto_central of its constituent topics rather than just keywords. A two-layer procedural filter (keyword + LLM) flags topics that are about the platform itself (notes mechanics, satire without a target) rather than substantive content.

4. Named entity recognition

Each note runs through:

• GLiNER (urchade/gliner_multi-v2.1) with a custom 16-type Portuguese taxonomy (politicians, parties, agencies, courts, security forces, public figures, media outlets, digital platforms, public programs, etc.) and per-type confidence thresholds (0.40–0.55, see taxonomy.json).
• Six regex extractors (URLs, dates, monetary values, percentages, laws/decrees, judicial process numbers) for high-precision structured patterns.
• Three canonicalization dictionaries (YAML-defined, editable) mapping orthographic variants to canonical forms for public programs (Bolsa Família, SUS, Prouni, …), agency acronyms (STF, TSE, IBGE, Anvisa, …), and parties.
• Overlap resolution by IoU ≥ 0.5 with priority order favoring regex over GLiNER.

An LLM-based revision step (Marco-Mini reviewing low-confidence ambiguous entities) is implemented in the project pipeline but was not applied to this release due to GPU time constraints; the entities here are post-extraction, post-canonicalization, pre-LLM-revision.

5. Note scoring

The official Twitter/X communitynotes Matrix Factorization algorithm is run on the PT subset using the upstream scoring/src code with minimal patches (asserts → warnings to accept the smaller universe). Outputs coreNoteIntercept (overall helpfulness baseline) and coreNoteFactor1 (first polarization factor) — the same fields the algorithm produces in production, restricted to PT notes.

---

Quantitative description

Volume by status (community consensus)

consenso	n_notes	%
NMR (NEEDS_MORE_RATINGS)	116,741	82.0%
CRH (CURRENTLY_RATED_HELPFUL)	15,857	11.1%
CRNH (CURRENTLY_RATED_NOT_HELPFUL)	4,423	3.1%
Outro	5,427	3.8%

PT-BR shows a higher CRH rate (11.57%) than English (8.69%) in the same snapshot — a finding worth further study.

Volume by macrotheme (top 10 of 38)

macrotheme_id	rotulo_macro	categoria_dominante	n_topics	n_notas
0	Sátira clara	Outro	73	12,090
1	Crise Política	Política	123	9,833
2	Sátira de humorista	Política	74	7,552
3	Economia e Política	Economia	72	6,864
4	Taylor Swift e Política	Política	45	6,400
5	Cotas e Cortes no Ensino Superior	Política	47	6,370
6	Futebol e Racismo	Política	48	5,474
7	Apostas Online	Golpes e Fraudes	61	4,700
8	Musk e Censura	Tecnologia	32	4,344
9	Racismo e Liberdade	Política	35	3,525

Entities by type (top 15 of 22)

tipo_entidade	n
URL_DOMINIO	174,625
PESSOA	82,190
LOCAL	49,247
ORGANIZACAO	47,191
VEICULO_MIDIA	43,810
ATOR_POLITICO	34,117
PLATAFORMA_DIGITAL	27,763
ORGAO_PUBLICO	16,788
DATA	13,292
PARTIDO	9,045
ESTATISTICA	7,519
EVENTO_POLITICO	7,367
PROGRAMA_PUBLICO	5,773
ORGAO_SEGURANCA	4,287
VALOR_MONETARIO	4,224

131,399 notes (92%) carry at least one entity; mean of 4.08 entities per annotated note.

Missing data

Column	Missingness	Reason
coreNoteIntercept/coreNoteFactor1	27.3%	Notes with too few ratings are not scored by MF.
topic_id	0.3% (397)	Notes outside the topic-model run.
topic_name/contexto_central/categoria_ampla	33.4%	Notes in the BERTopic outlier cluster (topic_id = -1) plus the 397 above.
macrotheme_id/macrotheme_label	33.4%	Same as topic_name.
status_final	3.8%	Notes whose status is unmapped or empty in the snapshot.
stratum	97.2%	Only 4,055 notes were sampled for tweet hydration.

---

Quality assessment

No exhaustive human validation was performed. Quality is supported by four converging lines of evidence:

1. Deterministic checks. The ETL produces MD5/SHA256 hashes per shard, schema-drift detection across shards, integrity-referential checks against orphans, and snapshot-to-snapshot diffs.

2. Intrinsic metrics. Topic clustering parameters were selected by DBCV grid search (32 combinations); topic_model fit produces silhouette and stability statistics. fastText language confidence has a median of 0.99. GLiNER scores have a median of 0.91; regex sources are 0.94–1.00 by design.

3. Cross-modal consistency. The cross_modal_validation/ folder ships pre-computed tables showing that macrotheme labels match the empirical entity-type signatures of their notes — for example, Futebol e Racismo has 41% ORGANIZACAO entities (Flamengo, Palmeiras, FIFA, CBF) and ge.globo.com as the top citation source; Apostas Online combines ORGANIZACAO+PLATAFORMA_DIGITAL with archive.ph, reclameaqui and pastebin as top sources, the canonical fingerprint of fraud reporting; Crise Política is dominated by ATOR_POLITICO (Lula, Bolsonaro, STF, TSE). 14 of the top 15 macrothemes show coherent signatures across four independent dimensions (label, dominant entity type, top entities, top domains).

4. Detected failure modes (transparent self-report). Two macrothemes are flagged: Sátira clara shows generic top entities (autor, nnn, comunidade) and 100% meta-platform top domains (communitynotes.x.com, help.x.com) — likely procedural content the two-layer filter did not catch. Cotas e Cortes no Ensino Superior shows low concentration (dominant/secondary type ratio ≈ 1.04) and entities that mismatch the label, suggesting the hierarchical-clustering threshold merged unrelated clusters. Both cases are documented in cross_modal_validation/concentracao_por_macrotema.csv.

5. Spot-checked language filter. A 300-note sample stratified across confidence bands was visually inspected and looked correct; this was a quick check, not a rigorous evaluation.

---

Limitations

• No exhaustive human validation of topic labels or named entities. The LLM-based revision step for the NER is implemented but was not applied to this release.
• Tweet text not redistributed due to X's terms of service. Use tweetId for hydration. The project keeps a static catalog with note-level context for human inspection.
• Snapshot-bound. This is the 2026-04-07 snapshot; status fields evolve over time. Future snapshots may diverge.
• PT-BR coverage bias. The fastText filter at 0.80 confidence aims for high precision and may exclude some short or code-mixed notes.
• Topic outliers. 33% of notes fall in topic_id = -1 (BERTopic outliers). Outlier reduction was applied but conservative.
• Procedural leakage. As noted, Sátira clara contains residual procedural content.

---

Suggested use cases

• NER benchmarking in PT-BR with politically salient categories (ATOR_POLITICO, PARTIDO, ORGAO_JUDICIARIO, ORGAO_SEGURANCA, PROGRAMA_PUBLICO).
• Helpful-note classification at note text only, or with text + tabular signals (coreNoteFactor1, n_ratings).
• Polarization analysis along coreNoteFactor1, conditioned on macrotheme.
• Source-citation studies: which domains are cited as evidence in misinformation pushback, by topic and over time.
• Longitudinal studies of Brazilian online discourse: the 4.5-year window covers two electoral cycles, the COVID aftermath, the 2024 floods, multiple sport scandals, and major changes to Twitter/X moderation.
• Procedural-content filtering: the Sátira clara failure mode is itself a use case for training procedural-vs-substantive classifiers.

---

Citation

@inproceedings{communitynotesbr2026,
  title  = {Community Notes BR: An enriched Portuguese subset of X's crowdsourced fact-checking notes for misinformation research},
  author = {Rocha, Davi Machado da},
  booktitle = {Anais do VII Dataset Showcase Workshop (DSW), SBBD},
  year   = {2026},
  publisher = {SBC},
  doi    = {TBD}
}

License

The original Community Notes data is released by X under CC0. Derivative annotations in this release (topic labels, NER, MF reproduction) are released under CC0 1.0 Universal to match the source license. Tweet text is not included; only tweetId is published.

AI usage disclosure

This dataset incorporates outputs from generative AI models as part of its construction:

• Marco-Mini-Instruct (GGUF Q6_K, via llama-cpp-python) for structured topic and macrotheme labeling, with grammar-constrained JSON output.
• intfloat/multilingual-e5-large-instruct for embedding generation.
• urchade/gliner_multi-v2.1 for NER extraction.

Generative AI was used for parts of the prose in this datacard and the accompanying paper. AI was not used to generate the data itself or its annotations beyond the explicit pipeline steps described above.

---

Maintainer: Davi Machado da Rocha (USP). Source of upstream data: Birdwatch public dump, snapshot 2026-04-07, ©X Corp under CC0.