TRACE: Temporal Rhetorical Analysis and Consistency Evaluation for Legislative Speech

Author

• David HernandezFollow

Available at: https://digitalcommons.calpoly.edu/theses/3354

Date of Award

6-2026

Degree Name

MS in Computer Science

Department/Program

Computer Science

College

College of Engineering

Advisor

Foaad Khosmood

Advisor Department

Computer Science

Advisor College

College of Engineering

Abstract

Legislators frequently discuss the same policy issues across multiple hearings and legislative sessions, sometimes maintaining consistent positions and other times modifying or reframing their stance over time. Understanding how these positions evolve is important for analyzing political discourse and democratic accountability, yet identifying such shifts at scale remains difficult.

We introduce TRACE (Temporal Rhetorical Analysis and Consistency Evaluation), a system built on the Digital Democracy Database (DDDB) for detecting rhetorical inconsistency in California legislative hearing testimony. TRACE organizes utterances into speaker-anchored timelines indexed by bill and session, then applies hybrid semantic retrieval — combining dense BGE embeddings with BM25 lexical search — to surface candidate statement pairs for the same legislator across hearings. Each candidate pair is passed to an LLM adjudicator that returns an explicit verdict and reasoning trace, reframing inconsistency detection from a static pairwise classification task into a longitudinal analysis of how individual legislators’ stated positions evolve over time.

We evaluate TRACE against a curated anchor case set drawn from the California legislative record, a 21-topic unsupervised corpus sweep over 1.25 million utterances, and a cross-speaker session scan of the 2025–2026 legislative session. The system correctly classifies 8 of 10 targeted anchor cases, discovers 11 confirmed instances of temporal rhetorical drift without prior knowledge of which legislators were involved, and surfaces 36 verified cross-speaker contradictions within legislative hearings. Benchmark evaluation on an external political statement dataset demonstrates adjudication performance competitive with GPT-4 Turbo, supporting the use of a locally hosted model without external API dependencies.