Span-level Aspect-based Sentiment Analysis via Table Filling

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9273–9284 July 9-14, 2023 ©2023 Association for Computational Linguistics Span-level Aspect-based Sentiment Analysis via Table Filling Mao Zhang1,2, Yongxin Zhu1,2, Zhen Liu1,2, Zhimin Bao3, Yunfei Wu3 Xing Sun3, Linli Xu1,2 1School of Computer Science and Technology, University of Science and Technology of China 2State Key Laboratory of Cognitive Intelligence 3Tencent YouTu Lab {zmyyy,zyx2016,liuzhenz}@mail.ustc.edu.cn {zhiminbao,marcowu,winfredsun}@tencent.com, linlixu@ustc.edu.cn Abstract In this paper, we propose a novel span-level model for Aspect-Based Sentiment Analysis (ABSA), which aims at identifying the senti- ment polarity of the given aspect. In contrast to conventional ABSA models that focus on mod- eling the word-level dependencies between an aspect and its corresponding opinion expres- sions, in this paper, we propose Table Filling BERT (TF-BERT), which considers the consis- tency of multi-word opinion expressions at the span-level. Specially, we learn the span repre- sentations with a table filling method, by con- structing an upper triangular table for each sen- timent polarity, of which the elements represent the sentiment intensities of the specific senti- ment polarity for all spans in the sentence. Two methods are then proposed, including table- decoding and table-aggregation, to filter out tar- get spans or aggregate each table for sentiment polarity classification. In addition, we design a sentiment consistency regularizer to guaran- tee the sentiment consistency of each span for different sentiment polarities. Experimental results on three benchmarks demonstrate the effectiveness of our proposed model. 1 Introduction Aspect-based sentiment analysis (Pontiki et al., 2014) ABSA is a fine-grained branch of sentiment analysis, which aims at recognizing the sentiment polarity of the given aspect in the sentence. For example, given the sentence “Boot time is super fast, around anywhere from 35 seconds to 1 minute” and the aspect “Boot time”, the opinion expression corresponding to the aspect is “super fast” so that the sentiment polarity of the aspect "Boot time" is positive. Recently, several methods (Tang et al., 2015; Wang et al., 2016; Ma et al., 2017; Huang et al., 2018; Sun et al., 2019a; Chen et al., 2020; Zhang and Qian, 2020; Xiao et al., 2021; Li et al., 2021; Zhou et al., 2021) have been proposed to exploit the connections between the given aspect and its corresponding opinion expressions in the task of ABSA. Tang et al. (2015) introduces recurrent neu- ral networks (RNNs) to retrieve the aspect-related information by fusing the aspect with its contextual- ized information in the sentence. Furthermore, Ma et al. (2017); Huang et al. (2018) propose to model the distance dependency between the aspect and the distant opinion expressions with attention mecha- nisms (Vaswani et al., 2017). To better leverage the syntax information in the ABSA task, some recent studies (Sun et al., 2019a; Xiao et al., 2021; Li et al., 2021) adopt graph neural networks (GNNs) over the dependency trees. Moreover, Chen et al. (2020); Zhou et al. (2021) generate dynamic aspect-specific trees for every sentence-aspect pair to learn the re- lationships between the …

Content truncated to 500 words