Skip to content

umanlp/speechact

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Speech act classification

Repository description

This repository contains the code for our paper "How to do politics with words: Investigating speech acts in parliamentary debates". The code can be used to replicate results from the paper and access model predictions.

If you use the code or materials in this repository, please cite our paper How to Do Politics with Words: Investigating Speech Acts in Parliamentary Debates:

@inproceedings{reinig-etal-2024-politics-words,
    title = "How to Do Politics with Words: Investigating Speech Acts in Parliamentary Debates",
    author = "Reinig, Ines  and
      Rehbein, Ines  and
      Ponzetto, Simone Paolo",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italy",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.727",
    pages = "8287--8300",
    abstract = "This paper presents a new perspective on framing through the lens of speech acts and investigates how politicians make use of different pragmatic speech act functions in political debates. To that end, we created a new resource of German parliamentary debates, annotated with fine-grained speech act types. Our hierarchical annotation scheme distinguishes between cooperation and conflict communication, further structured into six subtypes, such as informative, declarative or argumentative-critical speech acts, with 14 fine-grained classes at the lowest level. We present classification baselines on our new data and show that the fine-grained classes in our schema can be predicted with an avg. F1 of around 82.0{\%}. We then use our classifier to analyse the use of speech acts in a large corpus of parliamentary debates over a time span from 2003{--}2023.",
}

Annotation guidelines

The file Guidelines_speechacts.en.pdf contains the annotation guidelines that we used to annotate speech acts in parliamentary debates.

Speech act classification

Setup

This project was tested with python 3.11.

The requirements for speech act classification can be found in the file classification/requirements.txt.

Usage

In general, the following script can be used to train or evaluate a speech act classification model:

python3 speech-act-classification-BERT.py [your-configuration-file.conf]

The behaviour of the script depends on the contents of the provided .conf file. Different .conf files for evaluation and training are presented below.

The script always generates a logfile with information about the settings that you defined in the .conf file.

We also provide the code that we used for our SVM-based baseline:

python3 speech-act-classification-SVM.py config/svm-speech-act-classification.conf

Model and prediction files

We make one run of our best speech act classification model, BERT-large-context-run-x (Micro-F1 reported in table 5 of the paper: 84.01%/81.96% dev/test), publicly available. Due to space constraints in this repository, the model has to be downloaded from the following source: TODO: link/to/model

Please place the model's directory directly in this classification folder or change the path accordingly in the .conf file.

We make BERT-based model predictions on the dev and test set available in the folder predictions:

  • BERT-base w/o context (79.99% Micro-F1 on the dev set): BERT-base-all-runs/
  • BERT-base w/ context (80.18% Micro-F1 on the dev set): BERT-base-context-all-runs/
  • BERT-large w/ context (84.01% Micro-F1 on the dev set): BERT-large-context-all-runs/

The dataset used in classification experiments is available as dataset.pkl or in TSV-format (unzip dataset.zip).

Train a BERT-based model

First, modify train-BERT_context.conf to your liking. Change the following variables:

  • pretrained_bert_model: the name of the underlying pre-trained BERT model in the transformers library
  • output_dir: path where the model will be saved
  • context_mode: boolean indicating whether to use the BERT_context setting, as described in the paper.
  • the hyperparameters in the section [PARAM], if needed

Then, run the script:

python3 speech-act-classification-BERT.py train-BERT_context.conf

This trains a single run and saves it in the path provided in output_dir.

Evaluate a BERT-based model / Generate predictions

To evaluate one specific run of a model, proceed as follows:

  • Change the model path in the variable fine_tuned_model_path in the .conf file.
  • Run the script: python3 speech-act-classification-BERT.py config/eval_BERT_single_run.conf

To evaluate several runs of one model, proceed as follows:

  • Change the model path in the variable runs_paths in the .conf file.
  • Run the script: python3 speech-act-classification-BERT.py config/eval_BERT_multiple_runs_context.conf

To run a model without context, set the variable context_mode in the .conf file to False.

By default, the script evaluates on the dev set. To evaluate on the test set instead, change the variable do_test to True in the .conf file.

The script writes predictions and the logfile to the output folder provided in the variable output_dir. You can change this variable to a desired output folder. If the folder does not exist, it will be created. If it does exist, its contents with identical filenames will be overwritten.

More info about conf settings

The following table explains most of the variables in the .conf files:

Variable Explanation
do_train whether to train a model
do_save_model whether to save the model that was trained
do_hyperparameter_search whether to conduct hyperparameter search
do_test whether to evaluate on the test set instead of the dev set (default)
do_toy whether to limit the train set to few instances (for debugging)
do_majority_baseline whether to also evaluate a majority baseline on the provided dataset
do_average_runs whether to evaluate and average multiple runs provided in runs_paths below
do_consistency_check_mode Caution: this trains and evaluates on the entire dataset
do_merge_questions whether rhetorical_question and question are merged to a question_all class
do_add_punc whether end of sentence punctuation is to be added (ignore this if you use input_dataframe below)
splits_directory path to directory containing information about splits (ignore this if you use input_dataframe below)
annotations_directory path to directory containing the inception files (ignore this if you use input_dataframe below)
input_dataframe path to the pickle dataframe containing the entire dataset
gpu_node which GPU node to use on the cluster
pretrained_bert_model the underlying pre-trained BERT model
fine_tuned_model_path path to the folder containing the single fine-tuned model that you want to evaluate
runs_paths paths to all the runs that you want to evaluate
output_dir path to the directory where logfile and predictions will be written (will be created if does not exist, does overwrite)
log_filename name of your logfile to be saved in output_dir
seed seed for all things randomness (only relevant for training)
context_mode whether to use the BERT_context setting: this gives the previous sentence, the sentence containing the utterance and the next sentence as input to BERT. The utterance to be classified is marked by setting the token_type_ids for corresponding to the value 1, and 0 for remaining context tokens.

Speech act segmentation

Setup

This project was tested with python 3.10.

The requirements for the speech act segmentation code can be found in the file segmentation/README.md.

Usage

Follow the instructions in segmentation/README.md to create a virtual environment and install the required packages. Then run the code with:

python src/segmenter.py config/segment.conf

Analysis

We make the code to generate the figures in Section 6 of the paper available in the folder analysis/.

Setup

This project was tested with python 3.11.

Requirements to run the code are provided in analysis/analysis_requirements.txt.

We provide two scripts:

  • speech_acts_stats.py: This script generates plots for conflict, demand and request proportions for a given party (see Figures 3 and 4 in the paper).
  • speech_acts_all_parties_stats.py: This scripts aggregates information about conflict and cooperative communications across all parties and generates the bar plot for Figure 2 in the paper.

Usage

First, download the 6 input files for the scripts using the link provided in analysis/party_pickles/link_to_files.txt and place them in the folder party_pickles/. The files are too large to be stored in this GitHub repository.

Then, run one of both scripts:

python speech_acts_stats.py party_pickles/[party].pkl

and

python speech_acts_all_parties_stats.py party_pickles/

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages