基于擦除和生成式模型的情感可解释性分析

Interpretability of Sentiment Based on Erasure and Generative Models

情感可解释性分析是近年来比较新颖的研究方向,其目标是在预测文本的情感极性的同时给出决定情感极性的证据片段。该文在仅有情感分类任务数据集的基础上,提出了基于擦除的情感可解释性片段抽取方法,通过被擦除单词对情感极性逻辑判断的波动影响来决定证据的抽取。随后,利用擦除的方法使用模型对公开情感分析数据集中的部分数据进行片段抽取并人工过滤得到有监督数据,再使用T5序列生成式模型进行有监督训练,从而进一步提升证据抽取的性能。最终在“百度2022语言与智能技术竞赛:情感可解释评测”中获得第三名的成绩。

Related studies pay more and more attentions to sentiment interpretability analysis,which aims to predict the text's emotional polarity and the evidence segments that determine the emotional polarity.Based on the dataset of sentiment classification tasks,this study proposes an erasure based method for extracting evidence segments,which analysizes the logits influence of masked words on the emotional polarity.Subsequently,the model is used to extract the evidence of partial data in the public sentiment analysis dataset and the supervised data is manually filtered form the auto-extracted dataset.To further improve the performance of evidence extraction,this study fine-tunes the T5 sequence generative model on supervised data.This study won the third place in the Baidu 2022 Language and Intelligent Technology Competition:Sentimental Interpretable Assessment.