MvP: Multi-view Prompting Improves Aspect Sentiment Tuple Prediction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4380–4397 July 9-14, 2023 ©2023 Association for Computational Linguistics MVP: Multi-view Prompting Improves Aspect Sentiment Tuple Prediction Zhibin Gou∗, Qingyan Guo∗, Yujiu Yang† Tsinghua University zebgou@gmail.com gqy22@mails.tsinghua.edu.cn yang.yujiu@sz.tsinghua.edu.cn Abstract Generative methods greatly promote aspect- based sentiment analysis via generating a se- quence of sentiment elements in a specified format. However, existing studies usually pre- dict sentiment elements in a fixed order, which ignores the effect of the interdependence of the elements in a sentiment tuple and the diver- sity of language expression on the results. In this work, we propose Multi-view Prompting (MVP) that aggregates sentiment elements gen- erated in different orders, leveraging the intu- ition of human-like problem-solving processes from different views. Specifically, MVP in- troduces element order prompts to guide the language model to generate multiple sentiment tuples, each with a different element order, and then selects the most reasonable tuples by vot- ing. MVP can naturally model multi-view and multi-task as permutations and combinations of elements, respectively, outperforming previ- ous task-specific designed methods on multiple ABSA tasks with a single model. Extensive experiments show that MVP significantly ad- vances the state-of-the-art performance on 10 datasets of 4 benchmark tasks, and performs quite effectively in low-resource settings. De- tailed evaluation verified the effectiveness, flex- ibility, and cross-task transferability of MVP.1 1 Introduction Aspect-based sentiment analysis (ABSA) aims to predict tuples of sentiment elements of interest for a given text. There are four sentiment elements that constitute the main line of ABSA research: aspect term (a), aspect category (c), opinion term (o) and sentiment polarity (s) (Zhang et al., 2022). Given an example sentence, “I love the sushi badly!”, the corresponding elements are “sushi”, “food quality”, ∗Equal contribution. †Corresponding author. 1Code and data released at https://github.com/ ZubinGou/multi-view-prompting Task Output Aspect Category Opinion Sentiment (ACOS) (a, c, o, s) Aspect Sentiment Quad Prediction (ASQP) (a, c, o, s) Aspect Sentiment Triplet Extraction (ASTE) (a, o, s) Target Aspect Sentiment Detection (TASD) (a, c, s) Table 1: Aspect sentiment tuple prediction tasks with their corresponding outputs. Notably, although both ACOS and ASQP are the most complex quadratic pre- diction tasks, ACOS focuses on implicit aspects and opinions compared to ASQP. Detailed tasks and dataset statistics are shown in Appendix A. “love” and “positive”, respectively. Early studies fo- cus on a single sentiment element like aspect term (Liu et al., 2015; Ma et al., 2019), aspect category (Zhou et al., 2015) or sentiment polarity (Wang et al., 2016; Chen et al., 2017). Recent works propose compound ABSA tasks involving multi- ple associated elements, such as aspect sentiment triplet extraction (ASTE) (Peng et al., 2020), tar- get aspect sentiment detection (TASD) (Wan et al., 2020), aspect sentiment quad prediction (ASQP) (Zhang et al., 2021a) and aspect category opinion sentiment (ACOS) (Cai et al., 2020a). Their target formats are shown in Table 1. Recently, generative methods have been used to handle various ABSA tasks uniformly and achieved …

Content truncated to 500 words