Legislative hearings are distinct communication events with their own human recognizable structure and logic, yet they have been analyzed as distinct phenomena beyond their transcripts. Understanding legislative hearings at a deeper level is a major step in analyzing the work and power relationships within democratic bodies. In this work, we present a novel end-to-end pipeline for parsing rhetorical structures from legislative committee hearings with the goal of improving legislative discourse analysis and democratic accountability. We also propose a rhetorical structure grammar and find that 74.71% of California state bill discussions in our dataset can be parsed with this grammar. We evaluate our pipeline with two utterance section classifiers on a random subset of over 2619 annotated utterances corresponding to 100 hearings. We find that a logistic regression-based classifier with grammar-based output masking and utterance tag smoothing outperforms more complex models, achieving an F1 score of 59.0% against our ground truth discourse section labels.
Figures & tables
Figure 1: Analysis of the ground-truth labeled sample showed the following occurrence frequencies for different speaker roles, section types, and utterance lengths. The outlier utterance at around 104 tokens was due to a segmentation error for bill ‘CA_201720180AB10‘, hearing ‘52708‘.
Referent Type
Example
Replacement
Speaker
Debra Bowen
SpeakerPositionEnum associated with the individual referenced
Bill
AB 420
‘BILL’
Organization
California Polytechnic State University
‘ORG’
Table 1: Different types of named entities’ in-text mentions were replaced with more abstract, 1-token strings described in the ‘Replacement’ column. Examples of mentions that would be replaced are shown in the ‘Example’ column.
Order
Total
Kept
Dropped
Dropped (%)
2
9890
8971
919
9.3
3
5858
5208
650
11.1
4
2216
1894
322
14.5
5
950
775
175
18.4
6
455
366
89
19.6
7
239
198
41
17.2
Table 2: Rates at which shorter token ngrams were dropped in favor of longer ones for TF-IDF vectorization.
Figure 2: T-SNE clusters associated with the dimensionality-reduced ngram TF-IDF embeddings.
Figure 3: The final data model used for serializing, parsing, and restructuring hearing transcripts. Grey blocks denote the fields associated with different dataclasses. Blue blocks denote dataclasses. The core enumerables are defined in more depth in later sections.
Classifier
Parameter
Value
Logistic Regression
class_weight
balanced
Logistic Regression
max_iter
2000
Logistic Regression
solver
lbfgs
MLP Classifier
hidden_layer_sizes
(10, 10)
MLP Classifier
max_iter
2000
MLP Classifier
solver
adam
Table 3: Classifier hyperparameters.
Figure 4: The pipeline used in our utterance section classifier. Green blocks shapes represent processes. Blue cylinders represent data stores. Grey blocks denote classes. Red parallelograms denote input/output data.
Figure 5: The confusion matrix of our utterance classifier. Our classifier categorizes each utterance into one of ten possible categories: author introduction, bill presentation, legislator discussion, expert testimony, public comments, closing remarks, voting, and 3 fallback sections. These confusion matrices are based on evaluation on a validation set of bill discussions randomly selected from the manually labeled ground truth samples.
Model
Mask
Accuracy
Precision
Recall
F1
LR
True
0.701
0.587
0.624
0.590
LR
False
0.665
0.581
0.613
0.575
MLP
True
0.685
0.546
0.535
0.533
MLP
False
0.668
0.516
0.520
0.510
Table 4: Classifier evaluation macro metrics with and without masking.
Class
Precision
Recall
F1-score
Support
S0
0.792
0.704
0.745
27
S1
0.690
0.714
0.702
28
S2
0.921
0.671
0.777
207
S3
0.568
0.700
0.627
60
S4
0.868
0.861
0.864
137
S5
0.767
0.717
0.742
46
Table 5: Logistic regression classifier evaluation metrics. For readability purposes, section names have been replaced with their corresponding grammar symbols as described in 8 .
Figure 6: Speaker roles and positions associated with different sections. Voting subsections’ speaker role requirements may vary, but they represent subsets of the roles attached in the diagram. These represent the section-terminal relations given in Appendix B
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Enum Symbol Type
Symbol Name
Notation
TOP
START
A
TOP
MIDDLE
B
TOP
UPPER_MIDDLE
B1
TOP
LOWER_MIDDLE
B2
TOP
END
C
SectionEnum
INTRO
S0
Appendix
Table 8: Top-level, section, and speaker enumerables and their corresponding grammar symbols.
Analyses of legislative behavior often rely on voting records, overlooking the rich semantic and rhetorical content of political speech. In this paper, we ask three complementary questions about parliamentary discourse: how things are said, what is being said, and who is speaking in discursively similar ways. To answer these questions, we introduce a scalable and generalizable computational framework that combines diachronic stylometric analysis, contextual topic modeling, and semantic clustering of deputies' speeches. We apply this framework to a large-scale case study of the Brazilian Chamber of Deputies, using a corpus of over 450,000 speeches from 2003 to 2025. Our results show a long-term stylistic shift toward shorter and more direct speeches, a legislative agenda that reorients sharply in response to national crises, and a granular map of discursive alignments in which regional and gender identities often prove more salient than formal party affiliation. More broadly, this work offers a robust methodology for analyzing parliamentary discourse as a multidimensional phenomenon that complements traditional vote-based approaches.
Flávio Soriano, Victoria F. Mello, Pedro B. Rigueira +4
Computer Science Department, Universidade Federal de Minas Gerais, Belo Horizonte, Minas Gerais, Brazil
Self-introductions are common in legislative committee testimonies. Successfully detecting them and extracting the speaker's name is enormously helpful in the task of speaker identification in the context of government meetings. In this paper, we present a pipeline for detection of self-introductions in legislative committee testimony using machine learning. We construct a training dataset from 1.54 million utterances spanning five state legislative sessions, apply a name-matching heuristic to generate automatic labels, and train three classifiers: a decision tree, random forest, and XGBoost to find self-introductions and extract the speaker's name. We construct a feature set combining bag-of-words, positional context, structural signals, introductory phrase indicators, and discourse context features. Among the three classifiers, XGBoost achieves the best performance with an F1 score of 0.9747 and the fewest total errors; adding fine-tuned BERT probability features improves this further. As an extension, we score the full candidate dataset with a fine-tuned BERT classifier and add BERT probability outputs as features. This BERT-augmented XGBoost model improves F1 from 0.9747 to 0.9782 and reduces total test errors from 241 to 207. The primary gain over the decision tree baseline (F1 0.9323) is driven by discourse context features and the boosting ensemble strategy; BERT provides a modest complementary signal. Analysis of false positives reveals that a minority are genuine self-introductions mislabeled due to name inconsistencies in the source data, indicating that measured metrics modestly understate true performance.
This paper introduces GELATO (Government, Executive, Legislative, and Treaty Ontology), a dataset of U.S. House and Senate bills from the 118th Congress annotated using a novel two-level named entity recognition ontology designed for U.S. legislative texts. We fine-tune transformer-based models (BERT, RoBERTa) of different architectures and sizes on this dataset for first-level prediction. We then use LLMs with optimized prompts to complete the second level prediction. The strong performance of RoBERTa and relatively weak performance of BERT models, as well as the application of LLMs as second-level predictors, support future research in legislative NER or downstream tasks using these model combinations as extraction tools.