histlearn/notas-comunidade-ptbr

Community Notes BR: Enriched Portuguese Dataset for NLP and Text Mining

Community Notes BR is a curated Portuguese-language subset of X Community Notes enriched for Natural Language Processing (PLN), text mining, and research on collaborative misinformation moderation. It combines note-level metadata, topic and macrotheme labels, named entities, source-domain extraction, Matrix Factorization scoring fields, an optional Universal Dependencies syntax layer, and automatic quality-audit tables.

The dataset does not redistribute tweet text. It contains Community Notes text and identifiers released by the official Community Notes public data program; use tweetId for hydration when your own legal and API access conditions allow it.

Snapshot: official Community Notes public data from 2026-04-07. Upstream source: https://communitynotes.x.com/guide/en/under-the-hood/download-data Coverage: 2021-10-13 to 2026-04-05. Notes: 142,448 Portuguese notes, selected with fastText lid.176 confidence >= 0.80. Tweets: 90,822 unique tweet IDs. Authors: 25,483 anonymized Community Notes participant IDs. Entities: 510,352 post-processed entity rows, with 21 observed canonical entity types. Topics: 898 valid fine-grained topics, plus BERTopic outliers (topic_id = -1), organized into 38 macrothemes. Syntax layer: 292,275 automatic UD sentences and 5,599,679 token rows. Current documentation state: updated after the entity post-processing and quality-audit rounds of June 2026.

Resumo em português

Este é um subconjunto curado em PT-BR das Community Notes da X, enriquecido para pesquisa em PLN, mineração de texto e moderação colaborativa. Ele inclui 142.448 notas, 90.822 tweets distintos, 510.352 entidades nomeadas pós-processadas, 898 tópicos válidos, 38 macrotemas, campos de pontuação por fatoração matricial e uma camada opcional de sintaxe Universal Dependencies. A versão atual também publica tabelas de auditoria em quality_audit/, criadas para tornar explícitos os problemas e limites de offsets de entidades, alinhamento com UD e integridade entre notas, tópicos e entidades.

As anotações são automáticas e não passaram por revisão humana linha a linha. As tabelas de auditoria devem ser usadas como parte do dataset, especialmente quando a análise depende de spans, offsets ou alinhamento entre entidades e sentenças.

---

Files

File or folder	Rows	Size	Granularity / role
notes_pt.parquet	142,448	42.4 MB	one row per Portuguese Community Note
entities.parquet	510,352	68.8 MB	one row per extracted entity
topic_metadata.parquet	898	78.3 KB	one row per valid BERTopic cluster
macrotheme_metadata.parquet	38	9.4 KB	one row per macrotheme
DATA_DICTIONARY.md	-	-	explicit data dictionary with column-level descriptions
taxonomy.json	-	5.0 KB	entity extraction taxonomy and thresholds
syntax_ud_bosque/sentences/*.parquet	292,275	22.8 MB	automatic UD sentence spans
syntax_ud_bosque/tokens/*.parquet	5,599,679	89.7 MB	automatic UD token-level annotations
syntax_ud_bosque/notes_udpipe_bosque.conllu.zst	-	42.3 MB	compressed CoNLL-U export
quality_audit/	-	90.7 MB	automatic validation tables
cross_modal_validation/	-	64.6 KB	macrotheme/entity/domain consistency previews
source_code/	-	2.3 MB	sanitized construction and analysis notebooks

Primary join key: noteId between notes_pt.parquet, entities.parquet, the UD files, and the audit tables. tweetId is kept for tweet hydration and cross-reference, but tweet text is not included.

The Dataset Viewer is intentionally limited to the main analysis tables and the optional UD syntax layer. The automatic quality-audit tables remain available under quality_audit/ in the repository files.

The core dataset is released under CC0. The optional syntax_ud_bosque/ layer was generated with UDPipe using the portuguese-bosque-ud-2.17-251125 model, whose license is reported by the API as CC BY-NC-SA; see syntax_ud_bosque/udpipe_bosque_metadata.json.

---

Schema

For a complete column-level data dictionary, including keys, Parquet types, controlled values and interpretation notes, see DATA_DICTIONARY.md. The summary below highlights the main analysis tables.

notes_pt.parquet

Column	Description
noteId	Primary key, the official Community Notes note identifier.
tweetId	The tweet the note refers to.
note_author_participant_id	Anonymized Community Notes participant ID.
created_at, year_month	Note creation time in UTC and monthly convenience key.
summary	Note text as published in the Community Notes data.
summary_clean	Whitespace-normalized and URL-stripped text used by embeddings and UD parsing.
classification_algoritmo	Original Community Notes note classification, such as MISINFORMED_OR_POTENTIALLY_MISLEADING or NOT_MISLEADING.
idioma, confianca_idioma	fastText language label and confidence. All retained rows have idioma = pt and confidence >= 0.80.
status_final, consenso	Latest Community Notes status and compact status class (CRH, CRNH, NMR, Outro).
n_ratings, n_notas_no_tweet	Rating count for the note and number of notes attached to the same tweet.
coreNoteIntercept, coreNoteFactor1	Outputs from the official Matrix Factorization scoring code, restricted to the PT subset. Null when a note has insufficient ratings for scoring.
topic_id, topic_name, contexto_central, categoria_ampla	Fine-grained topic ID and LLM-generated topic metadata. topic_id = -1 denotes BERTopic outliers.
macrotheme_id, macrotheme_label	Hierarchical macrotheme ID and LLM-generated label.
stratum	Hydration sample stratum (top_notas, bat_signals, aleatorio); null for unsampled notes.

entities.parquet

Column	Description
entity_id	Unique ID of the extraction, formatted as <row_id>_<j>. No duplicates are present in the current release.
noteId, tweetId	Foreign keys to the note and tweet.
texto_entidade	Surface form extracted from the note/extraction context.
texto_canonico	Normalized form after lowercasing, unidecode and dictionary-based canonicalization where applicable.
tipo_entidade	Canonical Portuguese entity type. The current table has 21 observed types; taxonomy.json documents extraction rules and regex aliases.
label_modelo	Original GLiNER label or regex extractor name.
inicio, fim	Generated character offsets from extraction. Treat them as automatic spans, not as gold offsets against summary_clean; use quality_audit/entity_offset_audit.parquet before span-level analysis.
score	GLiNER confidence or high-confidence regex score.
fonte_extracao	Extraction source: GLiNER or one of the regex extractors for URLs, dates, percentages, monetary values, legal references or court-process numbers.
contexto	Local context window around the entity.
papel_no_texto	mencao or fonte_ou_evidencia.
dominio	Canonicalized domain for URL/domain entities.
topic_id, macrotheme_id	Denormalized topic and macrotheme IDs copied from the source note for analysis convenience.

syntax_ud_bosque/

The optional syntax layer is automatic UDPipe output over notes_pt.summary_clean.

Folder	Description
sentences/part-*.parquet	Sentence-level spans, text and sentence IDs.
tokens/part-*.parquet	Token-level CoNLL-U fields plus noteId and sentence metadata.
notes_udpipe_bosque.conllu.zst	Compressed CoNLL-U export with note-level comments.

quality_audit/

File	Rows	Description
entity_offset_audit.parquet	510,352	Checks whether texto_entidade and offsets validate against notes_pt.summary_clean.
ud_sentence_tokenization_audit.parquet	292,275	Checks whether UD sentence text can be located in summary_clean, and whether token counts and token IDs are coherent.
entity_ud_alignment.parquet	510,352	Aligns entities with UD sentence spans when offsets validate.
notes_topic_integrity_audit.parquet	142,448	Validates note-level topic and macrotheme references.
entities_topic_integrity_audit.parquet	510,352	Checks whether entity topic/macrotheme IDs match their source note.
quality_audit_summary.parquet / .csv	27	Compact count table for all audit statuses.

---

Construction Methodology

1. Community Notes ETL

The five official Community Notes tables (notes, noteRatings, noteStatusHistory, userEnrollment, batSignals) were downloaded by snapshot date, hash-verified, checked for schema drift across shards and materialized in DuckDB. Bronze tables preserve the source fields as strings, silver tables add typed columns and snake_case names, and gold tables export compressed Parquet files.

2. Portuguese filter

Each note's summary_clean was classified with fastText lid.176.bin. Notes with idioma = pt and confianca_idioma >= 0.80 were retained, yielding 142,448 notes. A 300-note stratified sample across confidence bands was visually inspected and looked correct, but this was a quick sanity check rather than a formal human evaluation.

3. Topic and macrotheme modeling

Notes were embedded with intfloat/multilingual-e5-large-instruct using a Portuguese instruction prompt oriented toward the factual matter under dispute. BERTopic was fitted with UMAP and HDBSCAN parameters selected by a DBCV grid search. Outliers were reduced conservatively by embedding similarity.

Each valid topic was labeled by Marco-Mini-Instruct with grammar-constrained JSON output. Macrothemes were obtained by hierarchical aggregation over topic embeddings and labeled in a second LLM pass using topic descriptions rather than only keywords. A procedural-content filter was used to separate substantive topics from notes about Community Notes mechanics, satire without a target, and similar meta-platform content.

4. Named entity recognition

Entity extraction combines:

• GLiNER (urchade/gliner_multi-v2.1) with a Portuguese taxonomy for people, political actors, parties, public bodies, courts, security forces, media outlets, platforms, public programs, organizations, locations and political events.
• Regex extractors for structured patterns such as domains/URLs, dates, percentages, monetary values, legal references and judicial process numbers.
• Canonical dictionaries for public programs, public agencies, parties, platforms and recurring media/domain variants.
• Overlap resolution favoring higher-precision regex outputs and canonical types.

The current release includes automatic post-processing after the first extraction. It removed generic role nouns and extraction markers, corrected known media/platform/party type conflicts, applied a POS/UD-assisted filter to single-token false positives, and refined the separation among PESSOA, PESSOA_PUBLICA and ATOR_POLITICO. The post-processing was conservative and automatic; it was not a human line-by-line adjudication.

Post-processing history for entities.parquet:

Stage	Rows	Notes
Initial extraction	536,373	GLiNER + regex + canonicalization.
Structural/entity cleanup	521,992	Removed extraction markers, malformed standalone URL tokens, generic role nouns and a few systematic type conflicts.
POS/UD-assisted cleanup	510,352	Removed 11,640 single-token false positives using morphology/POS and lexicon rules.
Person-type refinement	510,352	Reclassified 1,282 public/personal names as ATOR_POLITICO and 126 non-political public figures away from ATOR_POLITICO.

Current entities.parquet SHA-256: 6c4f87b6d41d839b610043668d47c5ba95b7a3bf42ebb4b2ba786a0e5be53388.

5. Matrix Factorization scoring

The official Twitter/X communitynotes Matrix Factorization scoring code was run on the PT subset with minimal compatibility patches for the smaller note universe. The release includes coreNoteIntercept and coreNoteFactor1, restricted to Portuguese notes. These fields are null for notes with too few ratings.

6. Syntax and audit layers

The UD syntax layer was generated automatically with UDPipe Bosque over summary_clean. The quality_audit/ tables were added after the post-processing rounds to document offset validity, entity-to-UD alignment and topic/entity referential integrity. These tables are intended to make downstream filtering explicit instead of silently hiding uncertain spans.

---

Source Code, Reproducibility and Exploration

The source_code/ folder contains sanitized notebooks used to construct and analyze the dataset:

Notebook	Purpose
01_etl_snapshot.ipynb	upstream snapshot ETL and typed analytical tables
02_portuguese_filter.ipynb	Portuguese-language filtering and ratings/status joins
03_topic_modeling_macrothemes.ipynb	embeddings, topic modeling, LLM labels and macrothemes
04_named_entity_recognition.ipynb	GLiNER, regex extractors and canonical dictionaries
05_matrix_factorization_scoring.ipynb	reproduction of Matrix Factorization scoring fields
06_helpfulness_classification.ipynb	exploratory helpfulness/consensus classification experiments
07_udpipe_bosque_enrichment_colab.ipynb	optional Colab pipeline for UDPipe Bosque annotations

Notebook outputs and embedded credentials were removed before publication. Full reproduction requires the upstream Community Notes data, local credentials where needed, optional X API access for hydration, and sufficient CPU/GPU resources for the modeling stages. The notebooks are transparent research code, not a one-command production package.

Public exploration interfaces:

• Static project catalog: https://communitynotesbr.netlify.app/
• Argument/entity explorer: https://explorador-argumentos.netlify.app/

---

Quantitative Description

Volume by status

consenso	Source status	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	null or unmapped	5,427	3.8

Topic coverage

Status	n_notes	%
Valid topic and macrotheme	94,889	66.6
BERTopic outlier (topic_id = -1)	47,162	33.1
Missing topic ID	397	0.3

Entities by type

tipo_entidade	n
URL_DOMINIO	174,624
PESSOA	58,543
LOCAL	49,022
ORGANIZACAO	46,835
VEICULO_MIDIA	43,292
ATOR_POLITICO	33,617
PLATAFORMA_DIGITAL	27,889
ORGAO_PUBLICO	16,740
DATA	13,292
PARTIDO	9,016
ESTATISTICA	7,519
EVENTO_POLITICO	7,347
PROGRAMA_PUBLICO	5,732
VALOR_MONETARIO	4,224
ORGAO_SEGURANCA	4,196
ORGAO_JUDICIARIO	2,828
LEI_NORMA	1,789
OPERACAO_POLICIAL	1,170
PESSOA_PUBLICA	1,159
PROCESSO_JUDICIAL	888
FONTE_CITADA	630

126,533 notes (88.83%) have at least one entity. The current release has a mean of 4.03 entities per annotated note and 3.58 entities per note overall.

Missing data

Field family	Missingness	Interpretation
coreNoteIntercept / coreNoteFactor1	27.3%	Notes with too few ratings are not scored by the MF algorithm.
topic_id	397 rows (0.3%)	Notes outside the topic-model run.
topic_name, contexto_central, categoria_ampla	47,559 rows (33.4%)	BERTopic outliers plus missing topic IDs.
macrotheme_id, macrotheme_label	47,559 rows (33.4%)	Same scope as topic labels.
status_final	5,427 rows (3.8%)	Status empty or unmapped in the snapshot.
stratum	138,393 rows (97.2%)	Only 4,055 notes were selected for tweet-hydration sampling.

---

Quality Assessment

No exhaustive human validation was performed. The release instead provides automatic checks and transparent failure-mode reporting.

Deterministic checks

The ETL uses file hashes, schema-drift checks, referential integrity checks and snapshot-to-snapshot comparisons. In the current release, entities.entity_id has zero duplicates and every entity row matches the topic and macrotheme IDs of its source note.

Entity offset audit

quality_audit/entity_offset_audit.parquet checks entities.parquet against notes_pt.summary_clean.

Audit status	n_entities
exact	238,770
normalized_exact	45
mismatch	11,384
out_of_bounds	260,153

Important interpretation: many offsets are not relative to summary_clean. Several URL/evidence spans appear to have been extracted from a richer extraction text where links were inserted into or appended to the note. The audit therefore flags these rows instead of rewriting offsets automatically. For span-level tasks, prefer rows with exact or normalized_exact, or inspect the shifted/repeated-text categories in the audit table.

Entity-to-UD alignment

Alignment status	n_entities
aligned_to_ud_sentence	237,738
entity_offset_not_validated_against_summary_clean	271,537
entity_crosses_or_exceeds_sentence_span	626
no_sentence_covering_offset	257
no_valid_ud_sentence_span	194

This alignment is useful for POS/morphology-aware filtering and syntactic analyses, but it should not be treated as a gold span annotation.

UD sentence/token audit

Audit	Status	n_sentences
Sentence location	exact_in_order	292,149
Sentence location	not_found	126
Token row count	matches_n_conllu_rows	292,275
Token ID sequence	sequential	292,275

Topic integrity

Audit	Status	n
Note topic reference	topic_exists	94,889
Note topic reference	no_topic_cluster	47,162
Note topic reference	missing_topic_id	397
Entity topic reference	topic_matches_note	510,352
Entity macrotheme reference	macrotheme_matches_note	510,352

Cross-modal validation

The cross_modal_validation/ folder contains compact tables comparing macrotheme labels with empirical entity-type and domain signatures. These are diagnostic previews rather than formal validation. They are useful for spotting residual procedural topics, overly broad macrothemes and labels whose entity/domain distribution is weakly concentrated.

---

Limitations

• No human line-by-line adjudication. Topic labels, entity types, offsets and UD parses are automatic. The post-processing rules are conservative but still rule-based.
• Tweet text is not redistributed. This follows X terms and the Community Notes public-data design. Use tweetId for hydration when permitted.
• Snapshot-bound status. Community Notes statuses evolve. The fields here reflect the 2026-04-07 snapshot.
• Portuguese filter tradeoff. The fastText threshold favors precision and may exclude short, code-mixed or unusual Portuguese notes.
• Topic outliers. 33.4% of notes have no valid topic/macrotheme label because they are BERTopic outliers or were outside the topic run.
• Entity offsets require care. Offsets are not uniformly valid against summary_clean; consult quality_audit/ for span-level work.
• Automatic syntax layer. UD annotations may contain tokenization, lemmatization, POS-tagging, morphological or dependency errors, especially in informal text, sarcasm, named entities and short notes.
• NER taxonomy is domain-specific. Types such as ATOR_POLITICO, PESSOA_PUBLICA, ORGAO_PUBLICO and VEICULO_MIDIA are designed for Brazilian Community Notes research and may not transfer directly to generic NER benchmarks.

---

Suggested Use Cases

• Portuguese NER evaluation and error analysis with politically salient entity categories.
• Topic-conditioned studies of source citation, media domains and evidence practices.
• Helpful-note classification using text, rating/status fields and MF scoring features.
• Polarization analysis using coreNoteFactor1, topic labels and macrothemes.
• Syntax-aware studies combining entities with automatic UD POS/morphology features.
• Procedural-content detection and macrotheme-quality diagnostics.
• Public-interest research on Brazilian online discourse across the 2021-2026 Community Notes window.

---

Citation

@misc{communitynotesbr2026,
  title        = {Community Notes BR: An NLP-Enriched Portuguese Dataset for Collaborative Misinformation Moderation},
  author       = {Rocha, Davi Machado da},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/histlearn/notas-comunidade-ptbr},
  note         = {Dataset associated with a manuscript submitted to the XIV Symposium on Knowledge Discovery, Mining and Learning (KDMiLe), BRACIS 2026}
}

License

The original Community Notes public data is released by X under CC0. Derivative annotations in the core release are distributed under CC0 1.0 Universal to match the source license. Tweet text is not included. The optional UDPipe Bosque syntax layer should be interpreted together with the model/license metadata in syntax_ud_bosque/udpipe_bosque_metadata.json.

AI Usage Disclosure

This dataset includes outputs from machine-learning and generative-AI components as part of its construction:

• intfloat/multilingual-e5-large-instruct for embedding generation.
• Marco-Mini-Instruct for structured topic and macrotheme labels, constrained to parseable JSON.
• urchade/gliner_multi-v2.1 for named-entity extraction.
• UDPipe Bosque for automatic tokenization, POS tagging, morphology, lemmatization and dependency parsing.

Generative AI also assisted with parts of the prose in this datacard and the associated manuscript. The main released annotations are exactly the pipeline outputs and post-processing rules documented above; they should be treated as automatic research annotations, not as human gold labels.

---

Maintainer: Davi Machado da Rocha (USP). Upstream source: Community Notes public data, snapshot 2026-04-07, X Corp under CC0.