Introduction
Predicting a politician's vote based solely on their speech is a significant challenge in natural language processing. Political discourse is often characterized by complex jargon and strict adherence to "party lines," making individual transcripts poor predictors of final voting behavior. This paper introduces the Time Evolving Contextual Graph (TEC) model, which addresses these challenges by integrating BERT-based text analysis with a dynamic graph framework. TEC tracks a speaker’s historical stances and their relationships with party members to model how a politician’s "state" evolves over time, leading to significantly higher accuracy in predicting parliamentary votes.
In parliamentary systems, votes are influenced by three primary factors:
- Intra-party Cohesion: The tendency of politicians to vote in blocks ("herd mentality").
- Temporal Consistency: The rooting of current stances in historical ideologies and past debates.
- Motion Context: The specific framing and linguistic cues of the bill being debated.
Methodology
The TEC model employs a multi-layered architecture designed to capture the evolving state of a speaker across multiple dimensions.
Text and Cohesion Encoding
The process begins with BERT, fine-tuned on the ParlVote dataset, to convert speech transcripts and motion descriptions into high-dimensional embeddings. These embeddings serve as node features in a heterogeneous graph where nodes represent speakers and motions. Edges represent intra-party relationships and speaker-motion interactions. A Graph Attention Network (GAT) is then utilized to learn weights for these relationships, capturing the social and topical context of the debate.
Temporal Modeling
To capture the evolution of a speaker's stance, TEC processes a sequence of these graphs over a defined "lookback window." The GAT-processed features are fed into a Gated Recurrent Unit (GRU), which acts as a Temporal Context Encoder. Finally, a temporal attention mechanism identifies the most relevant past debates that influence the current vote prediction.
Experimental Setup
The model was evaluated using the ParlVote dataset, comprising UK House of Commons transcripts from 1997 to 2019. The dataset includes 33,461 transcripts from 1,346 speakers across 1,995 motions. The task is a binary classification of votes ('Aye' vs. 'No'), with a relatively balanced distribution (53.57% 'Aye' / 46.43% 'No').
Results and Discussion
TEC demonstrated superior performance over several baselines, achieving an accuracy of 75.07% and a Matthews Correlation Coefficient (MCC) of 0.47. This represents a substantial improvement over standard BERT+MLP models (67.31% accuracy) and state-of-the-art static graph models like GPolS (73.69% accuracy).
The Impact of Temporal Context
A key finding of the study is the importance of the historical "lookback window." As shown in the performance analysis, the model reaches peak effectiveness when considering approximately six months of historical data.
Using too little history fails to provide sufficient context, while excessive history (spanning years) introduces noise from outdated political stances that may no longer reflect current realities. Furthermore, ablation studies confirmed that intra-party context (who a speaker's allies are) is a more powerful predictor than the specific content of the motion, validating political theories regarding the power of party discipline.
Conclusion
The TEC model proves that temporal and social contexts are essential for accurate speaker state analysis in political debates. By modeling the interconnected ecosystem of parties and time, TEC significantly outperforms text-only and static graph approaches. Future work may involve incorporating additional metadata such as constituency demographics or external signals from social media to further refine the predictive capabilities of the model.
Summary updated