ETMS@IITKGP at SemEval-2022 Task 10: Structured Sentiment Analysis Using A Generative Approach

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1373 - 1381 July 14-15, 2022 ©2022 Association for Computational Linguistics ETMS@IITKGP at SemEval-2022 Task 10: Structured Sentiment Analysis Using A Generative Approach R Raghav∗and Adarsh Vemali∗and Rajdeep Mukherjee Indian Institute of Technology Kharagpur, India {rraghav5600, adarsh.vemali, rajdeep1989}@iitkgp.ac.in Abstract Structured Sentiment Analysis (SSA) deals with extracting opinion tuples in a text, where each tuple (h, e, t, p) consists of h, the holder, who expresses a sentiment polarity p towards a target t through a sentiment expression e. While prior works explore graph-based or sequence labeling-based approaches for the task, we in this paper present a novel unified generative method to solve SSA, a SemEval-2022 shared task. We leverage a BART-based encoder- decoder architecture and suitably modify it to generate, given a sentence, a sequence of opin- ion tuples. Each generated tuple consists of seven integers respectively representing the in- dices corresponding to the start and end posi- tions of the holder, target, and expression spans, followed by the sentiment polarity class asso- ciated between the target and the sentiment ex- pression. We perform rigorous experiments for both Monolingual and Cross-lingual subtasks, and achieve competitive Sentiment F1 scores on the leaderboard in both settings. 1 Introduction Structured Sentiment Analysis (SSA) is the task of extracting structured information around sentiment expressions present in text in the form of opinion tuples O = {O1, O2, ..., On}, where each opinion tuple Oi = (h, t, e, p) consists of h, the holder (or source, used interchangeably) who expresses a sentiment polarity p towards a target (or aspect) t using an opinion or sentiment expression e (Barnes et al., 2021a). Prior works (Liu, 2012; Peng et al., 2020) have highlighted the importance of address- ing sentiment analysis as a structured prediction problem in order to capture the complete informa- tion around various opinions expressed in the text. The task of SSA thus expects to exploit the pair- wise interactions between the members of the same opinion tuple during the extraction process. ∗Equal contribution With the exponential growth of online market- places and user-generated content therein, SSA or near similar tasks of aspect-sentiment-opinion triplet extraction (Peng et al., 2020; Mukherjee et al., 2021a; Yan et al., 2021), and aspect-category- sentiment-opinion quad extraction (Cai et al., 2021), the newest additions under the broader um- brella of aspect-based sentiment analysis (ABSA) (Pontiki et al., 2014a,b) have become more impor- tant than ever (Mukherjee et al., 2021b). In the face of ever-expanding choices, it becomes a challeng- ing necessity to take educated explainable decisions from past user reviews. SSA guides the learning in the proper direction by facilitating an automated way to focus on major sentiment or opinion indi- cators. As a result, the task has wide applications in various market segments, such as e-commerce, food delivery, healthcare, ride sharing, travel and hospitality, to name a few. Previous efforts on SSA have primarily fo- cused on two approaches: sequence labeling-based (He et al., 2019), and graph-based …

Content truncated to 500 words