[Paper] Mapping the Political Discourse in the Brazilian Chamber of Deputies: A Multi-Faceted Computational Approach

Source: arXiv - 2604.21897v1

Overview

This paper presents a scalable, data‑driven framework for dissecting parliamentary speeches, applied to more than 450 k remarks from Brazil’s Chamber of Deputies (2003‑2025). By looking beyond roll‑call votes, the authors reveal how legislators talk, what they talk about, and who talks alike, offering a richer picture of political dynamics that can be leveraged by developers building civic‑tech, NLP, or analytics tools.

Key Contributions

• Multi‑dimensional analysis pipeline that fuses (i) diachronic stylometry, (ii) contextual topic modeling, and (iii) semantic clustering of speakers.
• Large‑scale empirical study on a 22‑year corpus of Brazilian legislative speeches, demonstrating the pipeline’s scalability.
• Empirical insights:
  • A clear stylistic drift toward shorter, more direct utterances over time.
  • Rapid agenda reshaping in response to national crises (e.g., economic shocks, pandemics).
  • Discursive alignments driven more by region and gender than by party affiliation.
• Open‑source toolkit (or at least a reproducible workflow) that can be adapted to other parliaments or deliberative bodies.

Methodology

1. Data Collection & Pre‑processing

   • Scraped official transcripts, cleaned HTML, removed stop‑words, and lemmatized Portuguese text.
   • Aligned each speech with metadata: deputy ID, party, state, gender, and timestamp.

2. Diachronic Stylometric Analysis

   • Computed classic style metrics (sentence length, lexical richness, use of passive voice) per year.
   • Tracked trends with simple time‑series models to spot long‑term shifts.

3. Contextual Topic Modeling

   • Trained a multilingual BERT‑based encoder (e.g., bert-base-portuguese-cased) to obtain sentence embeddings.
   • Applied a dynamic topic model (BERTopic) that clusters embeddings while allowing topics to evolve across years.

4. Semantic Speaker Clustering

   • Aggregated each deputy’s speech embeddings into a single representation (average or attention‑weighted).
   • Performed hierarchical clustering (e.g., HDBSCAN) to discover groups of deputies with similar rhetorical footprints.

5. Evaluation & Validation

   • Compared stylistic trends against external events (e.g., 2014 economic recession, 2020 COVID‑19).
   • Validated speaker clusters with known demographic attributes (region, gender) and party lines.

The pipeline is modular: swap the embedding model, change the clustering algorithm, or add a sentiment layer without breaking the overall workflow.

Results & Findings

Overall, the study demonstrates that how deputies speak can be as informative as what they vote for, opening a new analytical dimension for political scientists and technologists alike.

Practical Implications

• Civic‑Tech Platforms: Real‑time monitoring dashboards can flag emerging topics or stylistic shifts, alerting NGOs, journalists, and the public to policy pivots before votes occur.
• Legislative Analytics SaaS: Companies can enrich vote‑based scoring systems with speech‑based similarity scores, offering clients a more nuanced risk assessment of legislative outcomes.
• Bias & Representation Audits: The framework can surface under‑represented voices (e.g., women from certain regions) by quantifying discursive participation, supporting diversity initiatives.
• NLP Model Benchmarking: The Brazilian parliamentary corpus is a valuable multilingual, domain‑specific dataset for testing language models on long‑form political text.
• Policy Forecasting: Topic‑trend detection can feed into predictive models that anticipate budget allocations or regulatory focus, aiding strategic planning for businesses.

Developers can plug the open‑source pipeline into existing data pipelines (e.g., Apache Beam, Airflow) to process new legislative sessions automatically.

Limitations & Future Work

• Language Specificity: The current implementation is tuned for Portuguese; cross‑lingual transfer may need additional tokenization and cultural adaptation.
• Speaker Metadata Gaps: Missing or inconsistent demographic data can bias clustering results.
• Causality vs. Correlation: While topics align with crises, the model does not prove that speeches drive policy changes.
• Scalability to Real‑Time: Processing 450 k speeches took several hours on a modest GPU cluster; optimizing for streaming ingestion remains an open challenge.

Future research could integrate sentiment analysis, network‑based interaction graphs (who replies to whom), and multimodal data (e.g., video transcripts) to build an even richer portrait of parliamentary discourse.

Authors

• Flávio Soriano
• Victoria F. Mello
• Pedro B. Rigueira
• Gisele L. Pappa
• Wagner Meira
• Ana Paula Couto da Silva
• Jussara M. Almeida

Paper Information

• arXiv ID: 2604.21897v1
• Categories: cs.CL, cs.CY
• Published: April 23, 2026
• PDF: Download PDF