This is the official repository of the paper "Multimodal Argument Mining: A Case Study in Political Debates".
The project is organized as follows:
- deasy-learning: a simple python library for performing deep learning experiments (closed beta version).
- deasy-speech: the code regarding the experiments described in the paper.
We provide few important details for reproducing our experiments.
The corpora can be retrieved from the following links:
- UKDebates (in this repo is named as
arg_aaai2016): http://argumentationmining.disi.unibo.it/aaai2016.html - M-Arg: https://github.com/rafamestre/m-arg_multimodal-argumentation-dataset
- MM-USElecDeb60to16: the code to reproduce the dataset can be found in the multimodal-dataset folder. Information on the dataset and how to use the code to reproduce it can be found in the GUIDELINES.md file.
Deasy-learning works by saving a task configuration via a unique name (string).
A task is a complete experimental test, i.e. training a model on a particular corpus.
Run deasy-speech/runnables/list_tasks.py to compute all the available tasks.
A deasy-speech/registrations/ folder is created at the end of the process containing JSON files with all the existing configurations.
We report down below useful tags that are used to define tasks:
use_audio_features=True # Use feature-based audio representation
use_audio_features=False # Use embedding-based audio representation
ablation_study=None # Perform TA input configuration without any modality removal
ablation_study=text # Ablation study w/o text
ablation_study=audio # Ablation study w/o audio
The deasy-speech/runnables/task_train.py offers the following API:
# Run training
task_train(
task_config_name="", # name of the task to run
test_name='', # folder name to use where to save test results
save_results=True, # whether to save test results in a folder or not
framework_config_name=None, # a framework configuration name (optional)
debug=False # whether to run the task in debug mode (e.g., batch_size=1, epochs=1, etc..) or not
)
As an example, if you want to train the BiLSTM model calibrated for the TO input configuration on M-Arg, the script is defined as follows:
# Run training
task_train(
task_config_name="internal_key:instance--flag:task--framework:tf--tags:['annotation_confidence=0.85', 'calibrated', 'lstm', 'text_only']--namespace:m-arg",
test_name='bilstm_to',
save_results=True,
framework_config_name=None,
debug=False
)
Run the script to train the model.
You can inspect all calibrated model configurations in deasy-speech/configurations/models.py.
Run deasy-speech/runnables/task_forward.py to loaded a trained model and perform an inference step.
The script offers the following API:
# Run inference
task_inference(test_name='', # folder name containing a trained model
task_config_registration_info="", # a task configuration to use for inference (optional). If omitted, the same configuration used for training is selected.
task_folder='', # folder name of the task (parent folder of test_name)
save_results=False, # whether to save inference results or not
framework_config_name=None, # a framework configuration name (optional)
debug=False # whether to run the task in debug mode or not
)
An ablation study can be performed by selecting a specific task_config_registration_info value. In particular, look for task registration names with the ablation_study tag.
We report down below an example showing how to remove the text modality from the calibrated BiLSTM model on the ACD task concerning MM-USElecDeb60to16 corpus.
# Run inference
task_inference(test_name='bilstm_acd_ta_wav2vec',
task_config_registration_info="internal_key:instance--flag:task--framework:tf--tags:['ablation_study=text', 'lstm', 'calibrated', 'task_type=acd', 'text_audio', 'use_audio_features=False']--namespace:us_elec",
task_folder='us_elec',
save_results=False,
framework_config_name=None,
debug=False
)
SVM classifier require a dedicated script and are found in deasy-speech/runnables/other/ folder.
The following arguments are generally shared among SVM scripts:
is_calibrating=True # enables calibration
use_text_features=True # use text modality
use_audio_features=True # use audio modality
use_audio_data=True # use embedding-based audio representation (feature-based if set to False)
test_name='' # folder where to save test results
save_results=True # whether to save test results or not
Similarly to SVM, we provide dedicated scripts for running baselines like the random baseline reported in the paper.
These scripts are found in deasy-speech/runnables/other/
This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.
If using this dataset, please cite the following publication:
Eleonora Mancini, Federico Ruggeri, Andrea Galassi, and Paolo Torroni. 2022. Multimodal Argument Mining: A Case Study in Political Debates. In Proceedings of the 9th Workshop on Argument Mining, pages 158–170, Online and in Gyeongju, Republic of Korea. International Conference on Computational Linguistics.
@inproceedings{mancini-etal-2022-multimodal,
title = "Multimodal Argument Mining: A Case Study in Political Debates",
author = "Mancini, Eleonora and Ruggeri, Federico and Galassi, Andrea and Torroni, Paolo",
booktitle = "Proceedings of the 9th Workshop on Argument Mining",
month = oct,
year = "2022",
address = "Online and in Gyeongju, Republic of Korea",
publisher = "International Conference on Computational Linguistics",
url = "https://aclanthology.org/2022.argmining-1.15",
pages = "158--170"
}