ZHIXIAOBAO at SemEval-2022 Task 10: Apporoaching Structured Sentiment with Graph Parsing

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1343 - 1348 July 14-15, 2022 ©2022 Association for Computational Linguistics ZHIXIAOBAO at SemEval-2022 Task 10: Apporoaching Structured Sentiment with Graph Parsing Yangkun Lin∗, Chen Liang∗, Jing Xu, Chong Yang†, Yongliang Wang Ant Group Hangzhou, China {linyangkun.lyk, liangchen.liangche, jill.xj, yangchong.yang, yongliang.wyl}@antgroup.com Abstract This paper presents our submission to task 10, Structured Sentiment Analysis of the SemEval 2022 competition. The task aims to extract all elements of the fine-grained sentiment in a text. We cast structured sentiment analysis to the prediction of the sentiment graphs following (Barnes et al., 2021), where nodes are spans of sentiment holders, targets and expressions, and directed edges denote the relation types be- tween them. Our approach closely follows that of semantic dependency parsing (Dozat and Manning, 2018). The difference is that we use pre-trained language models (e.g., BERT and RoBERTa) as text encoder to solve the prob- lem of limited annotated data. Additionally, we make improvements on the computation of cross attention and present the suffix mask- ing technique to make further performance im- provement. Substantially, our model achieved the Top-1 average Sentiment Graph F1 score on seven datasets in five different languages in the monolingual subtask. 1 Introduction SemEval 2022 task 10 is a structured sentiment analysis task, aiming to predict all of the opinion tu- ples in a text. Each opinion O is a tuple (t, h, e, p), where h is a holder who expresses a polarity p to- wards a target t through a sentiment expression e. In practical, the task of structured sentiment analysis can help machines understand how people perceive ideas, policy etc. This paper describes the system developed by the team ZHIXIAOBAO for SemEval-2022 Task 10. We follow the work of (Barnes et al., 2021) to cast the task as dependency graph parsing problem. The predicted opinion tuples are denoted by a directed graph, G = (V, E), for each sentence. As shown in Figure 1, all tokens in a sentence are presented as nodes and there are directed edges between the ∗ The first two authors contributed equally to this work. † Corresponding author. nodes to represent their relations. Each node in V can point to multiple nodes, and can have multiple incoming edges too. Sentiment expressions are re- garded as roots in the structured sentiment graph (e.g., the token “5” and “don’t” in Figure 1). Notice that not all nodes connect to the other nodes (e.g., the token “give” in Figure 1). The isolated tokens are the none-sentiment elements in the sentence, thus we should be able to predict the edge type of “null” in our model. Original edge types are de- fined as holder, target, and expression. Following the work of (Barnes et al., 2021), we also tried the “+inlabel” style of definition, where none-“null” edge types consist of holderin, holderout, targetin, targetout, expressionin, expressionout. Foot mark- ers in, out denotes the in-span and out-span edges respectively. For example, the edge from “5” to …

Content truncated to 500 words