Contextualizing Argument Quality Assessment with Relevant Knowledge

This repository contains the data and additional materials for the paper "Contextualizing Argument Quality Assessment with Relevant Knowledge" accepted at the NAACL 2024 conference.

Overview

In this paper, we present a novel framework, SPARK, for assessing the quality of arguments by contextualizing them with relevant external knowledge. SPARK automatically produces four types of knowledge: feedback, assumptions, counter-arguments, and similar quality arguments to help improve the performance of existing argument quality ranking models. SPARK beats all baselines both in-domain on the GAQCorpus and out-of-domain on IBM-Rank-30K.

Install requirements

To install all requirements, run the following script:

pip install -r requirements.txt

Reproducing results

1. Augmenting the dataset The augmenting script, augment.py requires three input arguments: the input csv file, the output directory and an optional boolean argument which determines if a similar quality argument is to be sampled. Sampling a similar quality argument requires labels so this must not be used for testing splits.

python3 augment.py --input_file="path/to/csv" \
                    --output_dir="path/to/output_dir" \
                    --add_similar

For the DPR augmentation, there are three arguments that the pyserini_augment script takes: dataset_path (path to input csv), out_path (path to output csv) and col_name (Optional text column name. Default is "title")

python3 pyserini_augment.py --dataset_path="path/to/csv" \
                    --out_path="path/to/output_csv" \
                    --col_name="title"

Similarly, the llama_augment.py and flan_t5_augment.py scripts can be run by passing the input_file and output_dir to create those augmented datasets

2. Running the DPR training

python3 train_dpr.py --train_dataset_path="path1/" "path2/" "path3/" \
                                        --val_dataset_path="path1/" "path2/" "path3/" \
                                        --test_dataset_path="path1/" "path2/" "path3/" \
                                        --test_ibm_dataset_path="path/" \
                                        --model_name="bert-base-uncased" \
                                        --output_dir="output_dir/" \
                                        --train_batch_size=16 \
                                        --eval_batch_size=16 \
                                        --learning_rate=5e-5 \
                                        --epochs=5

3. Running the custom augmentation training

python3 train_dual_encoder.py --train_dataset_path="path1/" "path2/" "path3/" \
                                        --val_dataset_path="path1/" "path2/" "path3/" \
                                        --test_dataset_path="path1/" "path2/" "path3/" \
                                        --test_ibm_dataset_path="path/" \
                                        --model_name="bert-base-uncased" \
                                        --output_dir="output_dir/" \
                                        --train_batch_size=16 \
                                        --eval_batch_size=16 \
                                        --learning_rate=5e-5 \
                                        --epochs=5

The test set evaluation results will be displayed after the training is complete.

Citation

If you use the data or any materials from this repository, please cite our paper:

@misc{deshpande2023contextualizing,
      title={Contextualizing Argument Quality Assessment with Relevant Knowledge},
      author={Darshan Deshpande and Zhivar Sourati and Filip Ilievski and Fred Morstatter},
      year={2023},
      eprint={2305.12280},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

For any questions or issues, please contact Darshan Deshpande at [email protected].