A Generative Model for Structured Sentiment Analysis

Yihui Li, Yice Zhang, Yifan Yang, Ruifeng Xu

Abstract: Structured Sentiment Analysis (SSA) aims to extract the complete sentiment structure from a given text. Existing approaches predominantly rely on the interactions of words to predict the relationships between sentiment elements. While these methods have shown effectiveness, they overlook the rich label semantics associated with SSA tasks and necessitate extensive task-specific designs. In order to address the above problems, we propose a generative framework for tackling the SSA task. We designed two templates to transform the SSA task into a text generation problem, which facilitate the training process by formulating the SSA task as a text generation problem. Through experiments conducted on three SSA datasets, we demonstrate that our proposed generative approach outperforms all existing methods, thereby highlighting the advantages of employing the generative model for SSA.