HITSZ-HLT at SemEval-2022 Task 10: A Span-Relation Extraction Framework for Structured Sentiment Analysis

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1406 - 1411 July 14-15, 2022 ©2022 Association for Computational Linguistics HITSZ-HLT at SemEval-2022 Task 10: A Span-Relation Extraction Framework for Structured Sentiment Analysis Yihui Li, Yifan Yang, Yice Zhang, Ruifeng Xu Harbin Institute of Technology (Shenzhen), China 20s051013@stu.hit.edu.cn, {evanyfyang,zhangyc_hit}@163.com, xuruifeng@hit.edu.cn Abstract This paper describes our system that partici- pated in the SemEval-2022 Task 10: Structured Sentiment Analysis, which aims to extract opin- ion tuples from texts. A full opinion tuple gen- erally contains an opinion holder, an opinion target, the sentiment expression, and the cor- responding polarity. The complex structure of the opinion tuple makes the task challeng- ing. To address this task, we formalize it as a span-relation extraction problem and propose a two-stage extraction framework accordingly. In the first stage, we employ the span module to enumerate spans and then recognize the type of every span. In the second stage, we employ the relation module to determine the relation be- tween spans. Our system achieves competitive results and ranks among the top-10 systems in almost subtasks. 1 Introduction Sentiment analysis, also called opinion mining, aims to analysis people’s attitudes and emotions towards specific targets, such as products, orga- nizations, events, etc (Liu, 2012). It has become an important research field in natural language pro- cessing (Medhat et al., 2014; Hussein, 2018; Zhang et al., 2018). Structured sentiment analysis. Barnes et al. (2021) formally defines a complete opinion as a quadruple (h, t, e, p) where h is a holder who ex- presses a polarity p towards a target t through a sentiment expression e. Figure 1 presents examples of opinion quadruples. On the basis of this defi- nition, Barnes et al. (2022) formally establishes a benchmark for structured sentiment analysis. This benchmark consists of two tracks, the monolingual track and the crosslingual track, and we participate the monolingual track. In this paper, we cast this task as a span-relation extraction problem (Jiang et al., 2020), which is a formalization that has been widely used in many Figure 1: Examples of opinion quadruples (Barnes et al., 2021). information extraction tasks (Eberts and Ulges, 2019; Xu et al., 2021; Lu and Ng, 2021; Li et al., 2021). With the span-relation formalization, opin- ion quadruple extraction is divided into two stages. • In the first stage, we extract “meaningful” text spans and recognize their types. Specifically for this task, the type space is {h, t, e}. For those spans classified as e, we additionally detect the sentiment polarity they express. • In the second stage, we determine the rela- tions between spans. The relation space is set to {eh, et, ee, none}. eh and eh are used to facilitate the matching of sentiment expres- sions, holders, and targets during the decoding process. ee is used to deal with discontinuous sentiment expressions, which is inspired by (Li et al., 2021). In addition, we employ span pruning (Xu et al., 2021) to reduce the computation of the second stage. Finally, opinion …

Content truncated to 500 words