Fair and Argumentative Language Modeling for Computational Argumentation

Paper Abstract

Although much work in NLP has focused on measuring and mitigating stereotypical bias in semantic spaces, research addressing bias in computational argumentation is still in its infancy. In this paper, we address this research gap and conduct a thorough investigation of bias in argumentative language models. To this end, we introduce ABBA, a novel resource for bias measurement specifically tailored to argumentation. We employ our resource to assess the effect of argumentative fine-tuning and debiasing on the intrinsic bias found in transformer-based language models using a lightweight adapter-based approach that is more sustainable and parameter-efficient than full fine-tuning. Finally, we analyze the potential impact of language model debiasing on the performance in argument quality prediction, a downstream task of computational argumentation. Our results show that we are able to successfully and sustainably remove bias in general and argumentative language models while preserving (and sometimes improving) model performance in downstream tasks.

---

Repository Description

This repository contains all code and data needed to reproduce the experiments and results reported in our paper.

Data

• ABBA
  • This folder contains the annotated CSV files of the Queerphobia and Islamophobia bias dimension
• target_term_pairs
  • This folder contains the target term pairs used for CDA and creation of the ABBA test set
• argument_quality
  • This folder contains the CSV files of the IBM Rank and GAQ Corpus

Additional Data sources used:

• Wikipedia Dump
• Args.me Corpus
• Debate.org Corpus
• CMV Corpus

Note: The computed language adapters could not be uploaded to GitHub due to size constraints. You can find them on MADATA: https://madata.bib.uni-mannheim.de/id/eprint/390

Code

Includes all python files and notebooks subject to this paper.

A brief description of the files in code/bias_evaluation is:

• calculate_language_model_bias.ipynb
  • This notebook can be used to evaluate the language model bias of different model architectures using the ABBA annotation corpus as a test set. This is done by calculating the models perplexity for stereotypically and anti-stereotypically biased sentences and performing a paired t-test on the results.

A brief description of the files in code/adapter_based_debiasing is:

• cda_dataframe_creation.py
  • This script creates the datasets augmented with counterfactually biased sentences using different CDA strategies. Using the predefined target terms, it creates the CDA datasets from a certain input data set.
• cda_train_test_split.py
  • This script splits the previously created CDA dataset into a train and a test portion without reordering the sentences.
• debias_clm.py
  • This script is based on the run_clm.py script of the Adapter Hub, which can be found here. It is used to train a debiased language model adapter using a causal language modeling loss.
• debias_mlm.py
  • This script is based on the run_mlm.py script of the Adapter Hub, which can be found here. It is used to train a debiased language model adapter using a masked language modeling loss.

A brief description of the files in code/annotation_study is:

• create_annotation_df.ipynb
  • This notebook includes the steps performed to create the annotation dataframe from the arguments of the debate.org data set.
• analyze_ABBA_corpus.ipynb
  • This notebook provides some insights in the distribution of biased and unbiased sentences of the ABBA corpus as well as information about the debaters that wrote the arguments.

A brief description of the files in code/argumentative_language_modeling is:

• fine_tune_clm_adapter.py
  • This script is based on the run_clm.py script of the Adapter Hub, which can be found here. It is used to train an argumentative language model adapter using a causal language modeling loss.
• fine_tune_mlm_adapter.py
  • This script is based on the run_mlm.py script of the Adapter Hub, which can be found here. It is used to train an argumentative debiased language model adapter using a masked language modeling loss.

A brief description of the files in code/argument_quality_prediction is:

• prepare_aq_dataframes.py
  • This script prepares the datasets used to train and test the models on the downstream task of argument quality prediction.
• argument_quality_prediction.py
  • This script is based on the run_glue.py script of the Adapter Hub, which can be found here. It is used to train a regression adapter on top of a language adapter to predict the argument quality of a text as a value between [0,1].
• argument_quality_prediction_fusion.py
  • This script is based on the run_glue.py script of the Adapter Hub, which can be found here. It is used to train a regression adapter on top of a fused layer consisting of two language adapters in this case the argumentative and debiased adapters to predict the argument quality of a text as a value between [0,1].
• argument_quality_prediction_stacking.py
  • This script is based on the run_glue.py script of the Adapter Hub, which can be found here. It is used to train a regression adapter on top of two stacked language adapters (argumentative & debiased) to predict the argument quality of a text as a value between [0,1].
• hyperparameter_evaluation.py
  • This script is used to access all evaluation files of the argument quality training and return the hyperparameters of the best run.
• calculate_average_result.py
  • This script is used to calculate the mean result of the runs using different initial seeds, as well as the confidence interval of the mean result.

A brief description of the files in code/utils is:

• helper_functions.py
  • This script contains helper functions that are used multiple times in different other scripts related to this paper. It contains e.g., a function that creates the anti-sterotypically biased sentences based on the stereotypical biased sentences of the ABBA test set.
• target_terms.py
  • This script contains functions that return the target and attribute terms as arrays, which are used as vocabulary to identify stereotypically biased sentences for the two evaluated bias types.

Shell Files

Includes example shell files to run the python code

---

Citation

@inproceedings{FairAndArgLM,
  title={Fair and Argumentative Language Modeling for Computational Argumentation},
  author={Holtermann, Carolin and Lauscher, Anne and Ponzetto, Simone Paolo},
  booktitle={...},
  year={2022}
}

---

Author contact information:

cholterm@mail.uni-mannheim.de
anne.lauscher@unibocconi.it
simone@informatik.uni-mannheim.de