Modeling Aspect Correlation for Aspect-based Sentiment Analysis via Recurrent Inverse Learning Guidance

Proceedings of the 29th International Conference on Computational Linguistics, pages 6887–6896 October 12–17, 2022. 6887 Modeling Aspect Correlation for Aspect-based Sentiment Analysis via Recurrent Inverse Learning Guidance Longfeng Li1, Haifeng Sun1, Qi Qi1, Jingyu Wang1, Jing Wang1∗and Jianxin Liao1 1State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China {lilongfeng,hfsun,qiqi8266,wangjingyu,wangjing,liaojx}@bupt.edu.cn Abstract Aspect-based sentiment analysis (ABSA) aims to distinguish sentiment polarity of every spe- cific aspect in a given sentence. Previous re- searches have realized the importance of inter- active learning with context and aspects. How- ever, these methods are ill-studied to learn complex sentence with multiple aspects due to overlapped polarity feature. And they do not consider the correlation between aspects to distinguish overlapped feature. In order to solve this problem, we propose a new method called Recurrent Inverse Learning Guided Net- work (RILGNet). Our RILGNet has two points to improve the modeling of aspect cor- relation and the selecting of aspect feature. First, we use Recurrent Mechanism to improve the joint representation of aspects, which en- hances the aspect correlation modeling iter- atively. Second, we propose Inverse Learn- ing Guidance to improve the selection of as- pect feature by considering aspect correlation, which provides more useful information to de- termine polarity. Experimental results on Se- mEval 2014 Datasets demonstrate the effec- tiveness of RILGNet, and we further prove that RILGNet is state-of-the-art method in multi- aspect scenarios. 1 Introduction Aspect-based sentiment analysis (ABSA) aims to distinguish the orientation of sentiment existed on every aspect (Liu, 2012). For example, in the sentence: "Food is usually very good, though I wonder about freshness of raw vegetables", as- pect "raw vegetables" has negative polarity and "food" has positive polarity. ABSA task consists of two subtasks called aspect extraction and as- pect sentiment classification (Zhang et al., 2018). In this paper, we focus on aspect sentiment classi- fication task and assume that aspects are known (Majumder et al., 2018; Jiang et al., 2020). As ∗* Corresponding author. shown in above case, multiple sentiment polari- ties could exist in one sentence, which leads to the overlapping use of multiple aspects when predict- ing the polarity of current aspect (Du et al., 2019; Jiang et al., 2020). We can see that sentiment po- larity of "food" comes from word "good" which implies positive. "raw vegetables" does not have any words linked in above case. However, con- junction "though" determines sentiment polarity of "raw vegetables" which implies negative. Fea- ture about "good" is overlapped by these two as- pects and conjunction "though" leads to different sentiment polarity between "food" and "raw veg- etables". In this case, two aspects exhibit high cor- relation. Therefore, modeling the correlation be- tween multiple aspects is a suitable way to distin- guish overlapped feature. From the above case, it is not difficult to see that aspect correlation comes from these sources: aspects, their sentiment polari- ties and conjunctions. Conjunction establishes cor- relation between aspects. If model learns correla- tion between aspects, it can make …

Content truncated to 500 words