A Hierarchical Interactive Network for Joint Span-based Aspect-Sentiment Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 7013–7019 October 12–17, 2022. 7013 A Hierarchical Interactive Network for Joint Span-based Aspect-Sentiment Analysis Wei Chen1,2, Jinglong Du3, Zhao Zhang4, Fuzhen Zhuang2,5∗, Zhongshi He1∗ 1College of Computer Science, Chongqing University, Chongqing, China 2Institute of Artificial Intelligence, Beihang University, Beijing 100191, China 3College of Medical Informatics, Chongqing Medical University, Chongqing, China 4Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 5SKLSDE, School of Computer Science, Beihang University, Beijing 100191, China {mlg_cwei,zshe}@cqu.edu.cn, zhuangfuzhen@buaa.edu.cn Abstract Recently, some span-based methods have achieved encouraging performances for joint aspect-sentiment analysis, which first extract aspects (aspect extraction) by detecting aspect boundaries and then classify the span-level sen- timents (sentiment classification). However, most existing approaches either sequentially extract task-specific features, leading to insuf- ficient feature interactions, or they encode as- pect features and sentiment features in a par- allel manner, implying that feature represen- tation in each task is largely independent of each other except for input sharing. Both of them ignore the internal correlations between the aspect extraction and sentiment classifica- tion. To solve this problem, we novelly propose a hierarchical interactive network (HI-ASA) to model two-way interactions between two tasks appropriately, where the hierarchical in- teractions involve two steps: shallow-level in- teraction and deep-level interaction. First, we utilize cross-stitch mechanism to combine the different task-specific features selectively as the input to ensure proper two-way interactions. Second, the mutual information technique is applied to mutually constrain learning between two tasks in the output layer, thus the aspect input and the sentiment input are capable of encoding features of the other task via back- propagation. Extensive experiments on three real-world datasets demonstrate HI-ASA’s su- periority over baselines1. 1 Introduction Aspect-sentiment analysis (ASA) (Yan et al., 2021; Birjali et al., 2021) aims at extracting all the as- pects and their corresponding sentiments within the text simultaneously. And it can be divided into two tasks, i.e., aspect extraction (AE) and senti- ment classification (SC). AE is to extract the as- pects (Jakob and Gurevych, 2010; Poria et al., 2016; ∗Corresponding authors. 1The source codes are released in https://github. com/cwei01/HI-ASA. Karimi et al., 2021) in the sentence, and SC aims to predict the sentiments (Jiang et al., 2011; Lin et al., 2019; Karimi et al., 2021) for the extracted aspects. In recent years, the span-based models (Hu et al., 2019; Zhou et al., 2019; Lin and Yang, 2020; Lv et al., 2021) are increasingly becoming an alterna- tive for ASA because of their inherent advantages (e.g., they can avoid sentiment inconsistency and huge search space problems in tagging-based mod- els (Luo et al., 2019; Wang et al., 2021)), where the aspects are extracted by directly predicting the boundary distributions, and the sentiment polarities are classified based on the aspect-level words. For example, in the sentence “The screen size is satis- factory but the phone battery capacity is limited.", the aspect spans are “screen size” and “the phone battery”, and span-level sentiments are positive and negative, respectively. Formally, most of the …

Content truncated to 500 words