PITA: Prompting Task Interaction for Argumentation Mining

This repository implements a prompt tuning model for Argumentation Mining (AM), unifying AM's 3 subtasks in a generative way.

Requirements

allennlp==2.8.0 allennlp_models==2.10.1 easydict==1.9 nltk==3.8.1 numpy==1.21.6 pandas==1.3.5 scikit_learn==1.0.2 scipy==1.7.3 tensorboardX==2.5.1 tensorboardX==2.6.2.2 torch==1.9.0+cu111 torch_geometric==2.1.0.post1 tqdm==4.62.3 transformers==4.26.1 ujson==5.5.0

Preprocess

BART-base

Datasets

The PE and CDCP datasets have been preprocessed to .csv file

Task Interaction Graphs Construction: (1) w/o task tokens

python ./data/pe/construct_graphs_2.py

(2) w/ task tokens

python ./data/pe/construct_graphs_6.py

Train & Test

(1) w/o task tokens in prompts, run python scripts with suffix 3

python run_pe3.py --config ./configs/reproduce/pe_bartbase_graph1.json

(2) w/ task tokens in prompts, run python scripts with suffix 7_1

python run_pe7_1.py --config ./configs/reproduce/pe_bartbase_graph5.json

Reproducibility

We experiment on one Nvidia A100(40G) GPU with CUDA version $11.1$. All hyperparameters are in json files.