OPI at SemEval-2022 Task 10: Transformer-based Sequence Tagging with Relation Classification for Structured Sentiment Analysis

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1366 - 1372 July 14-15, 2022 ©2022 Association for Computational Linguistics OPI at SemEval-2022 Task 10: Transformer-based Sequence Tagging with Relation Classification for Structured Sentiment Analysis Rafał Po´swiata National Information Processing Institute, 00-608 Warsaw, Poland rposwiata@opi.org.pl Abstract This paper presents our solution for SemEval- 2022 Task 10: Structured Sentiment Analy- sis. The solution consisted of two modules: the first for sequence tagging and the second for relation classification. In both modules we used transformer-based language models. In addition to utilizing language models spe- cific to each of the five competition languages, we also adopted multilingual models. This approach allowed us to apply the solution to both monolingual and cross-lingual sub-tasks, where we obtained average Sentiment Graph F1 of 54.5% and 53.1%, respectively. The source code of the prepared solution is available at https://github.com/rafalposwiata/structured- sentiment-analysis. 1 Introduction Structured Sentiment Analysis (SSA) can be formulated as an information extraction task in which one attempts to find all of the opinion tu- ples O = Oi, ..., On in a text. Each opinion Oi is a tuple (h, t, e, p) where h is a holder who expresses a polarity p towards a target t through a senti- ment expression e, implicitly defining pairwise relationships between elements of the same tuple (Barnes et al., 2021). An example of such tuples as a structure sentiment graph was shown in Figure 1. This problem is relatively new and there has been little work published on the subject to date. To stimulate interest in this issue among the NLP community the SemEval-2022 Task 10: Structured Sentiment Analysis (Barnes et al., 2022) competi- tion was organized. The contest consisted of two sub-tasks: monolingual and cross-lingual. In the monolingual sub-task, the systems were trained and then tested on the datasets in the same languages. In the cross-lingual sub-task, systems had to be prepared for Catalan, Basque and Spanish datasets, while data in these languages could not be used for training. This setup is often known as zero-shot cross-lingual transfer (Hu et al., 2020). In this paper we present our system for this com- petition. We mainly focused on the solution for the monolingual track, however, it has also been successfully applied to the cross-lingual. The rest of the paper is organized as follows. Section 2 briefly describes related work. Section 3 shows an overview of used datasets. Section 4 elabo- rates on our solution. Experiments showing the effectiveness of the created system performed on development and test sets are presented in Section 5. The next section briefly describes the mistakes and limitations of our system. Finally, Section 7 concludes this paper. 2 Related Work Structured Sentiment Analysis can be broken down into five sub-tasks: a) expression (opinion) extrac- tion, b) target (aspect) extraction, c) holder extrac- tion, d) defining the relationship between these elements, and e) assigning polarity (Barnes et al., 2021).2 A few years ago, the main focus was on Aspect- Based Sentiment Analysis (ABSA), which …

Content truncated to 500 words