DiaASQ

This repository contains data and code for our paper "DiaASQ: A Benchmark of Conversational Aspect-based Sentiment Quadruple Analysis](https://arxiv.org/abs/2211.05705)"

To clone the repository, please run the following command:

git clone https://github.com/unikcc/DiaASQ

News

✨ 2023-05-10: Released training code.
📢 2023-05-10: Released the train and valid dataset.
⚡ 2022-12-10: Created repository.

Quick Links

• Overview
• Requirements
• Data Preparation
• Model Usage
• Citation

Overview

In this work, we propose a new task named DiaASQ, which aims to extract Target-Aspect-Opinion-Sentiment quadruples from the given dialogue. More details about the task can be found in our paper.

Requirements

The model is implemented using PyTorch. The versions of the main packages:

• python>=3.7
• torch>=1.8.1

Install the other required packages:

pip install -r requirements.txt

Data Preparation

Parsed data

Download the parsed data in JSON format from Google Drive Link. Unzip the files and place them under the data directory like the following:

data/dataset/jsons_zh
data/dataset/jsons_en

The dataset currently only includes the train and valid sets. The test set will be released at a later date; refer to this issue for more information.

Model Usage

• Train && Evaluate for Chinese dataset

  bash scripts/train_zh.sh

• Train && Evaluate for English dataset

  bash scripts/train_en.sh

• If you do not have a test set yet, you can run the following command to train and evaluate the model on the valid set.

  bash scripts/train_zh_notest.sh
  bash scripts/train_en_notest.sh

• GPU memory requirements

Dataset	Batch size	GPU Memory
Chinese	2	8GB.
English	2	16GB.

• Customized hyperparameters:
  You can set hyperparameters in main.py or src/config.yaml, and the former has a higher priority.

Citation

If you use our dataset, please cite the following paper:

@article{lietal2022arxiv,
  title={DiaASQ: A Benchmark of Conversational Aspect-based Sentiment Quadruple Analysis},
  author={Bobo Li, Hao Fei, Fei Li, Yuhan Wu, Jinsong Zhang, Shengqiong Wu, Jingye Li, Yijiang Liu, Lizi Liao, Tat-Seng Chua, Donghong Ji}
  journal={arXiv preprint arXiv:2211.05705},
  year={2022}
}