This repository provides the DisCOC model, which is used in the argument impact classification.

Figure 1: Architecture

• We use a dataset crawled from kialo.com to evaluate our model. The dataset is provided by Esin Durmus, and you can contact her via email ([email protected]). If you get the data, please put argument_impact.pkl in data/arg_impact and put argument_specificity_stance.pkl in arg_spec_stance.
• We use the pretrained BMGF-RoBERTa to predict the implicit discourse relations between two arguments. Please refer BMGF-RoBERTa to get the code and pretrained model. Alternatively, you can use the predictions in data/arg_impact directly.

Table 1: Performance of DisCOC and baselines with different attention mechanisms.

Figure 2: F1 scores of different models on varying the maximum path length.

Please refer to Exploring Discourse Structures for Argument Impact Classification to see another results.

• spacy
• torch
• tensorboardX
• sentencepiece
• transformers
• numpy
• pandas
• tqdm
• scikit-learn

We need to merge the stance and discourse relations with text features. Please refer to src/preprocess_argimpact.ipynb to preprocess data and check correlations.

To train models with text features with stances and discourse relations, you need to run train_chain.py. You can use different encoders (e.g., lstm, bert, roberta, xlnet, and albert) and different encoding modes (e.g., han, flat, interval, segmented, contextualized, and discoc). Another hyper-parameters can also be passed by command line arguments.

python train_chain.py \
    --train_dataset_path ../data/arg_impact/train.jsonl \
    --valid_dataset_path ../data/arg_impact/valid.jsonl \
    --test_dataset_path ../data/arg_impact/test.jsonl \
    --save_model_dir ../dumps/argimpact \
    --label_map argument_impact_label_map_3 \
    --max_len 100 \
    --max_num_text 1 \
    --max_num_context 5 \
    --mode discoc \
    --encoder roberta \
    --gpu_ids 0,1 \
    --epochs 10 \
    --train_batch_size 32 \
    --eval_batch_size 10 \
    --lr 2e-5 \
    --weight_decay 0.0 \
    --max_grad_norm 2.0 \
    --activation leaky_relu \
    --dropout 0.1 \
    --finetune full \
    --add_trans \
    --save_best f1 \
    --seed 0

We also provide the script to predict the argument persuasion only by discourse relations. You can run train_disco.py to reproduce the experiments.

python train_disco.py \
    --train_dataset_path ../data/arg_impact/train_discourse.jsonl \
    --valid_dataset_path ../data/arg_impact/valid_discourse.jsonl \
    --test_dataset_path ../data/arg_impact/test_discourse.jsonl \
    --save_model_dir ../dumps/argimpact \
    --label_map argument_impact_label_map_3 \
    --max_num_pairs 5 \
    --predictor lstm \
    --hidden_dim 128 \
    --gpu_ids 0 \
    --epochs 100 \
    --train_batch_size 32 \
    --eval_batch_size 10 \
    --lr 1e-3 \
    --weight_decay 5e-4 \
    --max_grad_norm 2.0 \
    --activation leaky_relu \
    --dropout 0.0 \
    --save_best f1 \
    --seed 0

The details of this architecture are described in the following paper. If you use this model in your work, please kindly cite it.

@inproceedings{LiuOSJ21,
  author    = {Xin Liu and
               Jiefu Ou and
               Yangqiu Song and
               Xin Jiang},
  title     = {Exploring Discourse Structures for Argument Impact Classification},
  booktitle = {Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing (ACL-IJCNLP) 2021},
  year      = {2021},
}