ATN

Codes and datasets for our COLING 2022 paper: "Attention Transfer Network for Aspect-level Sentiment Classification"

Overview

• In this paper, we transter attention knowledge from resource-rich document-level sentiment classification to resource-poor aspect-level sentiment classification.
• We design an Attention Transfer Framework (ATN) to enhance the attention process of resource-poor aspect-level sentiment classification.

Dependencies

• python=3.5
• numpy=1.14.2
• tensorflow=1.9

Usage

Step1: pretrained (skip this step)

• train(for example, you can use the folowing command to pre-train DSC model)

python pre_train.py

• eval (at inference time, you can get attention scores)

python pre_train_eval.py

step2: transfer

• Attention Guide

sh run_atn_guide.sh

• Attention Fusion

sh run_atn_fusion.sh

Citation

@inproceedings{zhao-etal-2020-attention,
    title = "Attention Transfer Network for Aspect-level Sentiment Classification",
    author = "Zhao, Fei and Wu, Zhen and Dai, Xinyu",
    booktitle = "Proceedings of the 28th International Conference on Computational Linguistics",
    year = "2020",
    pages = "811--821"
}

if you have any questions, please contact me zhaof@smail.nju.edu.cn.