USSA: A Unified Table Filling Scheme for Structured Sentiment Analysis

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14340–14353 July 9-14, 2023 ©2023 Association for Computational Linguistics USSA: A Unified Table Filling Scheme for Structured Sentiment Analysis Zepeng Zhai1, Hao Chen2, Ruifan Li1,3,4∗, Xiaojie Wang1,3,4 1School of Artificial Intelligence, Beijing University of Posts and Telecommunications, China 2STCA, Microsoft, China 3Engineering Research Center of Information Networks, Ministry of Education, China 4Key Laboratory of Interactive Technology and Experience System, Ministry of Culture and Tourism, China {zepeng, rfli, xjwang}@bupt.edu.cn and hche@microsoft.com Abstract Most previous studies on Structured Sentiment Analysis (SSA) have cast it as a problem of bi-lexical dependency parsing, which cannot address issues of overlap and discontinuity si- multaneously. In this paper, we propose a niche- targeting and effective solution. Our approach involves creating a novel bi-lexical dependency parsing graph, which is then converted to a unified 2D table-filling scheme, namely USSA. The proposed scheme resolves the kernel bot- tleneck of previous SSA methods by utilizing 13 different types of relations. In addition, to closely collaborate with the USSA scheme, we have developed a model that includes a pro- posed bi-axial attention module to effectively capture the correlations among relations in the rows and columns of the table. Extensive exper- imental results on benchmark datasets demon- strate the effectiveness and robustness of our proposed framework, outperforming state-of- the-art methods consistently1. 1 Introduction Structured Sentiment Analysis (SSA) aims to iden- tify all opinion tuples within a given sentence. An opinion tuple (h, t, e, p) denotes a group of four el- ements: the holder h expresses a sentiment polarity p towards an opinion target t through a sentiment expression e. As shown in Figure 1(a), an example involving two opinion tuples illustrates the defini- tion of SSA. SSA is more challenging than other related tasks because it requires identifying all four elements of the tuple and may involve overlapping or discontinuous elements. For example, aspect- based sentiment analysis (Pontiki et al., 2014, 2015, 2016; Li et al., 2021b) mostly identifies flat aspect and opinion terms, and opinion mining (Katiyar and Cardie, 2016; Xia et al., 2021) identifies opin- ion tuples without the sentiment polarity. *Corresponding author. 1Code and datasets are available at https://github.com/ zzp-seeker/USSA. I found your words to be really rude and long winded . My employer Holder your words Target really rude Expression Holder My employer Target your words Negative Negative My employer found your words to be really rude and long winded . Exp: Neg Holder Target Target E-NEG Exp: Neg Exp: Neg Exp: Neg My employer found your words to be really rude and long winded . H-S H-E T-S H-E T-E T-E T-NW Holder H-NW E-NW E-NW (a) (b) (c) long winded really Expression E-NEG E-NW Figure 1: (a) An example of structured sentiment analy- sis, which contains both overlapping and discontinuous situations about two expressions. (b) Ambiguous ex- pressions when using head-first bi-lexical dependency parsing method (Barnes et al., 2021). (c) Our proposed unified dependency parsing formulation. E-NEG …

Content truncated to 500 words