💬 NLP

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

This paper introduces a scalable computational framework combining stylometric analysis, topic modeling, and semantic clustering to analyze over 450,000 speeches in the Brazilian Chamber of Deputies (2003–2025), revealing long-term shifts toward direct speech, crisis-driven agenda reorientations, and discursive alignments where regional and gender identities often supersede formal party affiliation.

Original authors: Flávio Soriano, Victoria F. Mello, Pedro B. Rigueira, Gisele L. Pappa, Wagner Meira Jr., Ana Paula Couto da Silva, Jussara M. Almeida

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Flávio Soriano, Victoria F. Mello, Pedro B. Rigueira, Gisele L. Pappa, Wagner Meira Jr., Ana Paula Couto da Silva, Jussara M. Almeida

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Brazilian Chamber of Deputies as a massive, noisy kitchen where 1,980 chefs (the politicians) are cooking up laws and policies for the country. For decades, researchers trying to understand what's happening in this kitchen have only looked at the final menu (the voting records). They know who voted for the soup and who voted for the stew, but they've completely ignored the cooking process: the arguments, the shouting, the recipes being changed, and the specific flavors the chefs are trying to create.

This paper is like installing a high-tech camera and a super-smart AI chef in that kitchen to watch, listen, and analyze everything that happens over 22 years (2003–2025). The researchers didn't just look at the votes; they analyzed over 450,000 speeches to answer three simple questions: How do they speak? What are they talking about? and Who is speaking in a similar way?

Here is a breakdown of their findings using everyday analogies:

1. How They Speak: The "Soundbite" Shift

The Analogy: Think of a politician's speech like a long, complex novel.
The Finding: Over the last two decades, these novels have been shrinking into short, punchy tweets.

Shorter & Simpler: In 2004, the average speech was a long, winding road (456 words) full of complex legal jargon. By 2020, they had become short, direct paths (269 words).
Why? It's like politicians are adapting to the "TikTok era." They are speaking faster and more directly to grab attention in a busy media world.
The Twist: Even though the speeches got shorter, the chefs didn't run out of ingredients. The vocabulary remained rich and diverse; they just stopped writing long paragraphs and started using more action verbs (like "do," "fix," "build") instead of abstract nouns.

2. What They Talk About: The "Reactive Menu"

The Analogy: Imagine the kitchen menu changes instantly based on what's happening outside the restaurant.
The Finding: The legislative agenda is incredibly reactive. It doesn't stick to a plan; it panics and pivots based on national crises.

The "Infrastructure" Era (2006–2014): For years, the menu was all about building roads and bridges (Development).
The "Corruption" Shock (2005 & 2015): Suddenly, the menu switched entirely to "Politics and Administration" because of massive corruption scandals and the impeachment of a president.
The "Health" Emergency (2020–2023): When the pandemic hit, the entire kitchen stopped talking about everything else and focused 100% on Health.
The "Security" Pivot (2025): Most recently, the focus has sharply swung toward Public Security and Justice.
Takeaway: The politicians aren't ignoring long-term problems; they are just constantly putting out the biggest fire in front of them.

3. Who Agrees With Whom: The "Hidden Alliances"

The Analogy: Imagine you try to sort the chefs into teams based on their apron colors (their political parties). You'd expect all the "Red Apron" chefs to talk the same way.
The Finding: The apron color is a terrible predictor of who actually agrees with whom.

The Map: Instead of 11 big party teams, the AI found 49 distinct "tribes" of politicians who speak similarly.
The Real Dividers:
- Region is King: A chef from the North often speaks more like a chef from the South (even if they are on opposite political teams) than like a chef from the same party in a different state. Local interests matter more than party labels.
- Gender is a Bridge: The researchers found a fascinating "sisterhood" effect. There are small groups where over 90% of the chefs are women. These women come from all different political parties (Red, Blue, Green), yet they speak in a remarkably similar way, focusing on shared concerns like social rights and education.
The "Noise": About 30% of the politicians didn't fit into any specific tribe. They were "generalists" who spoke about everything and nothing in particular, making them hard to categorize.

The Big Picture

This study is like upgrading from a black-and-white photo of a political debate to a 3D, color, surround-sound movie.

Old Way: "Party A voted against Party B." (Boring, incomplete).
New Way: "Party A and Party B are actually arguing about the same thing, but Party A is using short, angry sentences while Party B is using long, legal ones. Meanwhile, the women from both parties are quietly agreeing on a third issue that nobody else is talking about."

Why does this matter?
It shows us that democracy is messier and more interesting than just counting votes. It reveals that politicians are often driven by their local neighborhoods and their identities (like gender) more than by the party logo on their shirts. By using computers to read all these speeches, the researchers gave us a new tool to understand the true, complex heartbeat of a democracy.

1. Problem Statement

Traditional analyses of legislative behavior rely heavily on structured data, specifically roll-call voting records, to infer political alignments and ideological positions. While effective for mapping party cohesion and conflict outcomes, this approach has significant limitations:

Semantic Blindness: It ignores the rich rhetorical, semantic, and argumentative content of political speeches.
Strategic Divergence: In the Brazilian context, there is a known divergence between "saying" (speeches aimed at constituents) and "voting" (actions aligned with party discipline). Voting records alone fail to capture the "cheap talk" or strategic signaling inherent in parliamentary discourse.
Lack of Multidimensionality: Existing computational studies often focus narrowly on specific topics (e.g., immigration) or use limited NLP techniques, failing to treat parliamentary discourse as a multidimensional phenomenon involving form (style), content (themes), and alignment (who speaks similarly).

2. Methodology

The authors propose a scalable, generalizable computational framework that combines three complementary analytical pillars applied to a corpus of 453,280 speeches from the Brazilian Chamber of Deputies (2003–2025).

A. Data Collection & Preprocessing

Corpus: Retrieved via RESTful API from the official Open Data Portal. Includes 1,980 distinct deputies.
Filtering: Speeches were filtered to include only the 11 largest political parties (e.g., PT, MDB, PSDB) to ensure statistical robustness, resulting in a final corpus of 354,719 speeches.
Preprocessing:
- Monologue Isolation: Used regex to truncate transcripts at the point of speaker interruption, ensuring each document represents a single deputy's speech.
- Normalization: Removed parliamentary formalities (e.g., "Mr. President"), standardized party names across historical changes (e.g., PMDB $\to$ MDB), and converted text to lowercase.

B. Three Analytical Pillars

1. Stylometric & Linguistic Analysis (How it is said)

Goal: Track the evolution of rhetorical structure and style over time.
Metrics:
- Readability: Flesch Reading Ease score.
- Lexical Diversity: Type-Token Ratio (TTR).
- Syntactic Structure: Relative frequency of Nouns, Verbs, Adjectives, and Adverbs (via POS tagging).
- Named Entity Recognition (NER): Frequency of People, Organizations, and Locations to track political attention.
Tool: spaCy with the pt_core_news_lg model.

2. Contextual Topic Modeling (What is being said)

Goal: Identify the thematic agenda and its evolution.
Pipeline:
- Bigram Detection: Used Gensim's Phrases to identify multi-word expressions (e.g., "reforma tributária").
- Embedding: Converted speeches to vectors using distiluse-base-multilingual-cased-v2 (chosen for its sentence-level context optimization over Portuguese-specific BERT models).
- Clustering: Applied UMAP for dimensionality reduction and HDBSCAN for clustering.
- Topic Extraction: Used c-TF-IDF to extract representative keywords.
- Macro-Theme Aggregation: Employed an LLM (Claude 3 Sonnet) to aggregate granular topics into 6 macro-themes (e.g., Politics, Health, Public Security) and perform lexical expansion for classification.

3. Semantic Clustering (Who speaks similarly)

Goal: Map discursive alignments that may transcend formal party lines.
Pipeline:
- Embedding: Used Linq-Embed-Mistral (7B parameters, 32k context window) to generate 4096-dimensional vectors for entire speeches without truncation.
- Fingerprinting: Calculated an average embedding vector for each deputy.
- Clustering: Projected vectors via UMAP and clustered using HDBSCAN.
- Optimization: Hyperparameters were tuned via Optuna to maximize the Silhouette Score while penalizing trivial solutions (e.g., simple left-right splits).
- Profiling: Analyzed clusters based on Party Affiliation (using Herfindahl-Hirschman Index for concentration), Region, and Gender.

3. Key Contributions

Multidimensional Framework: A novel pipeline that triangulates stylometry, topic modeling, and semantic clustering to analyze legislative discourse holistically.
Large-Scale Empirical Characterization: The first comprehensive NLP analysis of the Brazilian Chamber of Deputies covering 22 years (450k+ speeches), filling a gap in literature regarding non-US legislative bodies.
Methodological Innovation: Demonstrates the utility of modern sentence-transformers and LLM-assisted thematic aggregation for political science, moving beyond simple word-frequency approaches.
Open Resources: The code and data processing pipeline are made publicly available to foster reproducibility.

4. Key Results

A. Stylistic Evolution (Form)

Conciseness: Speeches have become significantly shorter (avg. 456 words in 2004 $\to$ 269 words in 2020).
Simplification: Readability scores improved (less complex), and the syntactic style shifted from nominal (noun-heavy, abstract) to verbal (verb-heavy, action-oriented).
Implication: This aligns with the "mediatization of politics," where communication adapts to faster media cycles.

B. Thematic Dynamics (Content)

Reactive Agenda: The legislative agenda is highly volatile and driven by national crises rather than stable long-term priorities.
- 2005: "Mensalão" corruption scandal $\to$ Politics/Administration dominance.
- 2015–2017: Impeachment of Dilma Rousseff $\to$ Politics/Administration dominance.
- 2020–2023: COVID-19 Pandemic $\to$ Health dominance.
- 2025: Sharp rise in Public Security and Justice.
Key Entities: NER analysis confirmed the President as the central discursive figure, with rising mentions of the Judiciary (Alexandre de Moraes) and cultural figures linked to social laws.

C. Discursive Alignments (Alignment)

Granularity: The model identified 49 distinct discursive clusters, far exceeding the 11 political parties analyzed.
Party Fragmentation: The average Partisan HHI (pHHI) was low (0.28), indicating that most clusters are not party monoliths. Only extreme ideological groups (e.g., PSOL, PCdoB) showed high party cohesion.
Regional Identity: 23 clusters showed high regional concentration (rHHI > 0.7), with some clusters composed entirely of deputies from specific regions (North, Northeast, South), suggesting "discursive federalism."
Gender Affinity: Several clusters were overwhelmingly female (>90%), spanning multiple parties. This suggests gender acts as a stronger organizing principle for discourse than party affiliation for many deputies.
Noise: ~30% of deputies were classified as "noise" (outliers), representing generalist speakers or those with insufficient speech data.

5. Significance

Beyond Voting Records: The study proves that speech analysis reveals political dynamics (regionalism, gender coalitions, strategic signaling) that are invisible in roll-call vote data.
New Lens on Representation: It challenges the view of parties as monolithic blocks, revealing a complex landscape where local interests and identity politics often supersede formal party lines.
Scalable Methodology: The framework offers a robust template for analyzing parliamentary discourse in other democracies, particularly those with large, unstructured textual archives.
Civic Impact: By mapping the "inner workings" of deliberation, this approach provides tools for greater transparency, allowing journalists and the public to understand the true drivers of political behavior beyond election cycles.

Limitations & Future Work

Context: The current model analyzes monologues without linking them to specific bills or interactive debate dynamics (interruptions/dialogue).
Bias: There is a risk of misinterpreting group patterns as essentialist stereotypes; the authors emphasize these are dynamic groupings, not fixed labels.
Future Directions: Linking speeches to specific legislative outcomes, modeling deliberative dialogue, and testing alternative language models pre-trained specifically on Brazilian Portuguese.