ISCAS at SemEval-2022 Task 10: An Extraction-Validation Pipeline for Structured Sentiment Analysis

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1305 - 1312 July 14-15, 2022 ©2022 Association for Computational Linguistics ISCAS at SemEval-2022 Task 10: An Extraction-Validation Pipeline for Structured Sentiment Analysis Xinyu Lu∗, Mengjie Ren∗, Yaojie Lu, Hongyu Lin Chinese Information Processing Laboratory Institute of Software, Chinese Academy of Sciences, Beijing, China 181303216@yzu.edu.cn, mjren@bupt.edu.cn, {yaojie2017,hongyu}@iscas.ac.cn Abstract ISCAS participated in both sub-tasks in SemEval-2022 Task 10: Structured Senti- ment competition. We design an extraction- validation pipeline architecture to tackle both monolingual and cross-lingual sub-tasks. Ex- perimental results show the multilingual ef- fectiveness and cross-lingual robustness of our system. Our system is openly released on: https://github.com/luxinyu1/ SemEval2022-Task10/. 1 Introduction Aspect-based Sentiment Analysis (ABSA) aims to detect the fine-grained sentiment tendency ly- ing underneath texts. After decades of develop- ment, this area has formed a large family of tasks. Nevertheless, many of them are too simple or overlap with each other. Meanwhile, the popu- lar evaluation resources are limited both in number and linguistic diversity. SemEval-2022 Task 10 (Barnes et al., 2022) is proposed to unify differ- ent sub-tasks in ABSA and introduces new met- rics, new datasets on different languages to bet- ter evaluate methods in this area. Task 10 chal- lenges its participants to extract opinion quadru- ple (holder, target, expression, polarity) from texts across English, Spanish, Basque, Catalan and Norwegian in monolingual (Sub-task 1) or cross- lingual (Sub-task 2) manners. Figure 1 provides two aspect-level annotations in a same sentence. We applied an extraction-validation pipeline sys- tem and participated in both sub-task. Our system ranked at 10th in 32 teams on the monolingual task without using extra data, and achieved competitive performance on the cross-lingual task. Besides, the proposed pipeline can be employed universally in monolingual and cross-lingual scenarios. ∗This work was finished during their internship at ISCAS- CIP Lab. 2 Background 2.1 Task Definition Task 10 is formalized as detecting all opinion tuples O = Oi, ..., On in given text s. Con- cretely, each opinion Oi is a quadruple (h, t, e, p), denoting a holder who expresses a polarity ∈ {Positive, Neutral, Negative} towards a target through a sentiment expression. It’s worthy to note that, h, t, e can be empty in this task. Following Cai et al. (2021), tuples with empty values are re- gard as implicit opinions in this system description paper. For the example in Figure 1, the quadruple "(–, them, don’t believe negative)" is an implicit opinion. 2.2 Related Work Aspect-based Sentiment Analysis Recently, there has been a large body of work focusing on different sub-tasks of ABSA. Generally we divide these sub-tasks into two categories: atomic and compound. Atomic ones take single element (e.g., t, e, or p) as the output and most of them can be treated as a sequence tagging problem (Li and Lam, 2017; Xu et al., 2018; Li et al., 2018; Wu et al., 2020b; Pouran Ben Veyseh et al., 2020). The com- pound ones need to find pairs (e.g., (t, p)), triplets (e.g., (t, …

Content truncated to 500 words