Analysis of Parliamentary Debate Transcripts Using Community-Based Graphical Approaches (Student Abstract)

Anjali Bhavan, Mohit Sharma, Ramit Sawhney, Rajiv Ratn ShahAAAI 2020 · DOI

Introduction

Analyzing the stances of elected officials is vital for a healthy democracy, but the sheer volume of parliamentary transcripts makes manual oversight impossible. Traditional sentiment analysis often treats speeches as isolated pieces of text, ignoring the social and political context—the "who" and "with whom"—that defines political life.

This research proposes a hybrid sentiment analysis model for UK Parliamentary debates that combines traditional text features with graph-based "community" embeddings. By mapping how politicians vote and interact, the model aims to bridge the gap between what a politician says and their underlying political affiliations, testing the hypothesis that parliamentarians with similar voting patterns tend to deliver similar speeches.

Methodology

The authors propose a two-step classification approach using a Multi-Layer Perceptron (MLP):

  1. Graph Construction: Two distinct graphs represent the political landscape: simGraph (connecting politicians by shared stances) and oppGraph (connecting them based on differences in affiliations and voting patterns).
  2. Feature Extraction:
    • Graph Embeddings: The DeepWalk algorithm generates 128-dimensional embeddings for each politician from the constructed graphs.
  3. Textual Features: TF-IDF scores (up to tri-grams) and subjectivity scores from the Harvard General Inquirer (HI) lexicon are extracted.
  4. Classification: These features are concatenated and fed into an MLP to classify the polarity of motions and speeches.

To validate the link between voting and speaking, the authors also calculated the Pearson correlation between voting similarity and speech similarity (using Word2Vec cosine similarity).

Dataset

The study utilizes the Hansard parliamentary debates dataset, which includes speeches from 607 politicians. The content spans diverse topics such as healthcare, security, and general governance. The dataset provides labels for speaker party affiliations, motion polarities, and speech polarities.

Key Insights

  • Political Homophily: The research leverages the principle that individuals associate with similar others. Encoding "political neighborhoods" into graph embeddings provides context that text alone cannot capture.
  • Voting Predicts Speaking: The authors found a Pearson correlation of 0.697 between voting similarity and speech similarity, confirming that political blocs share a "linguistic DNA" when debating motions.
  • Graph vs. Text: While text features are powerful, adding graph-based community insights consistently improved performance, suggesting a politician's network is a strong signal for their sentiment.

Results

The model was evaluated using ten-fold cross-validation. The combination of all features (HI + TF-IDF + DeepWalk) yielded the best results, achieving a high F1-score of 0.93.

Performance comparison of different feature sets in classifying speech polarity.

Using DeepWalk to capture community structure allowed for a more compact feature vector compared to previous methods while maintaining or improving accuracy.

Limitations and Future Work

A primary challenge identified is the "Rebel" problem: most misclassified data points belonged to politicians who vote against their party's typical habits. These outliers are naturally harder for a community-based model to predict.

Future research aims to investigate which specific types of motions (e.g., ethical or national security issues) elicit the most deviation from standard party rhetoric. Additionally, the researchers suggest that incorporating multimodal data, such as video or audio of debates, could further enhance sentiment detection by capturing tone and body language.

Summary updated

BibTeX
@inproceedings{bhavan2020analysis,
  author = {Anjali Bhavan and Mohit Sharma and Ramit Sawhney and Rajiv Ratn Shah},
  title = {{Analysis of Parliamentary Debate Transcripts Using Community-Based Graphical Approaches (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7148},
}
Figure