UFRGSent at SemEval-2022 Task 10: Structured Sentiment Analysis using a Question Answering Model

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1360 - 1365 July 14-15, 2022 ©2022 Association for Computational Linguistics UFRGSent at SemEval-2022 Task 10: Structured Sentiment Analysis using a Question Answering Model Lucas Rafael Costella Pessutto Institute of Informatics UFRGS – Brazil lrcpessutto@inf.ufrgs.br Viviane P. Moreira Institute of Informatics UFRGS – Brazil viviane@inf.ufrgs.br Abstract This paper describes the system submitted by our team (UFRGSent) to SemEval-2022 Task 10: Structured Sentiment Analysis. We propose a multilingual approach that relies on a Question Answering model to find tuples con- sisting of holder, target, and opinion expression. The approach starts from general questions and uses the extracted tuple elements to find the remaining components. Finally, we employ an aspect sentiment classification model to clas- sify the polarity of the entire tuple. Despite our method being in a mid-rank position in the Se- mEval competition, we show that the question- answering approach can achieve good coverage retrieving sentiment tuples, allowing room for improvements in the technique. 1 Introduction Opinions abound on the Internet nowadays. They are a valuable source of information since people often rely on them for making purchases. Com- panies can also benefit from this vast amount of opinions, as they do not need to conduct opinion polls or focus groups to measure the acceptance of a particular product (Liu, 2011). The large volume of opinions available becomes hard for humans to process. This leads to the study of ways of au- tomating the processing of opinions, in order to summarize them. Sentiment Analysis is the field of study which aims at processing the information conveyed by un- structured texts, providing structured information that facilitates the understanding of the opinions, at- titudes, or emotions towards a particular entity (Liu, 2011). Sentiment Analysis can be performed at different levels of granularity (entire review, sen- tence, or aspect). Aspect-Based Sentiment Analy- sis (ABSA) aims to identify and rate the features (or aspects) of the entity being evaluated. Typically, ABSA involves the following phases: (i) identify and extract entities in reviews; (ii) identify and extract the aspects of an entity; (iii) cluster simi- lar aspects; and (iv) determine the polarity of the sentiment over the entities and the aspects. Most of the research in Sentiment Analysis focuses on solving only one of these phases at a time. Task 10 in SemEval 2022 – Structured Senti- ment Analysis (Barnes et al., 2022) proposes a new approach to tackle the Sentiment Analysis problem, where the elements that constitute an opinion are identified together, in a structured way through a graph. Thus, Structured Sentiment Analysis can be seen as an information extraction task since we want to find the text spans where opinions about a particular feature are expressed (Barnes et al., 2021). In this paper, we describe UFRGSent, a multilin- gual approach that relies on a question answering system to find the elements of an opinion present in review texts and a fine-tuned model to classify the polarity of the sentiment tuple. Our average …

Content truncated to 500 words