Query-induced multi-task decomposition and enhanced learning for aspect-based sentiment quadruple prediction (MRC-CLRI)

✨ Introduction

This repository contains the code and data for the paper titled "Query-induced multi-task decomposition and enhanced learning for aspect-based sentiment quadruple prediction". The paper introduces a novel end-to-end non-generative model for ASQP involving multi-task decomposition within machine reading comprehension (MRC) framework. This README provides an overview of the repository and instructions for running the code and using the data.

📃 Data

The ACOS dataset is sourced from ACOS, while the ASQP dataset is sourced from ABSA-QUAD.

🚀 Quick Start

⚙️ Setup

To run the code in this repository, you'll need the following dependencies:

Python 3.9
PyTorch 2.2
transformers

Install these dependencies using pip:

conda create -n MRC-CLRI python=3.9
conda activate MRC-CLRI
pip install -r requirements.txt

🤖 Download Pre-trained Model

Before executing the code, you need to download the pre-trained model SentiWSP.

⚡️ Running the Code

Model Training:

python run.py \
  --train_batch_size 4 \
  --data_path ./data/ACOS/v2/rest/ \
  --task ACOS \
  --data_type rest \
  --model_path ../pretrained-models/SentiWSP \
  --learning_rate1 3e-5 \
  --learning_rate2 1e-5 \
  --use_category_SCL \
  --use_sentiment_SCL \
  --contrastive_lr1 3e-5 \
  --contrastive_lr2 1e-5 \
  --do_train

Model Testing:

We release the ACOS-Rest MRC-CLRI model (one seed): rest_test_model.pkl [Google Drive]. You can run it with the following command:

# without Refined Inference
python run.py \
  --eval_batch_size 8 \
  --data_path ./data/ACOS/v2/rest/ \
  --task ACOS \
  --data_type rest \
  --model_path ../pretrained-models/SentiWSP \
  --checkpoint_path ./outputs/saves/ACOS/rest/rest_test_model.pkl \
  --do_test
# 'f1': 0.6201716738197425

# with Refined Inference (Use the hyperparameters from our paper)
python run.py \
  --eval_batch_size 8 \
  --data_path ./data/ACOS/v2/rest/ \
  --task ACOS \
  --data_type rest \
  --model_path ../pretrained-models/SentiWSP \
  --checkpoint_path ./outputs/saves/ACOS/rest/rest_test_model.pkl \
  --beta 25 \
  --alpha 0.98 \
  --do_test
# 'f1': 0.6271186440677967

Model Inference:

python run.py \
  --do_inference \
  --load_ckpt_name ./outputs/saves/ACOS/rest/rest_test_model.pkl