Affective Knowledge Enhanced Multiple-Graph Fusion Networks for Aspect-based Sentiment Analysis

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5352 - 5362 December 7-11, 2022 ©2022 Association for Computational Linguistics Affective Knowledge Enhanced Multiple-Graph Fusion Networks for Aspect-based Sentiment Analysis Siyu Tang1∗, Heyan Chai1∗, Ziyi Yao1, Ye Ding2†, Cuiyun Gao1, Binxing Fang1,3, Qing Liao1,3† 1 Harbin Institute of Technology, Shenzhen, China 2 Dongguan University of Technology, China 3 Peng Cheng Laboratory, Shenzhen, China {tangsiyu999, chaiheyan, yaoziyi}@stu.hit.edu.cn, dingye@dgut.edu.cn, gaocuiyun@hit.edu.cn, fangbx@cae.cn, liaoqing@hit.edu.cn Abstract Aspect-based sentiment analysis aims to iden- tify sentiment polarity of social media users toward different aspects. Most recent meth- ods adopt the aspect-centric latent tree to con- nect aspects and their corresponding opinion words, thinking that would facilitate establish- ing the relationship between aspects and opin- ion words. However, these methods ignore the roles of syntax dependency relation labels and affective semantic information in determining the sentiment polarity, resulting in the wrong prediction. In this paper, we propose a novel multi-graph fusion network (MGFN) based on latent graph to leverage the richer syntax depen- dency relation label information and affective semantic information of words. Specifically, we construct a novel syntax-aware latent graph (SaLG) to fully leverage the syntax dependency relation label information to facilitate the learn- ing of sentiment representations. Subsequently, a multi-graph fusion module is proposed to fuse semantic information of surrounding contexts of aspects adaptively. Furthermore, we design an affective refinement strategy to guide the MGFN to capture significant affective clues. Extensive experiments on three datasets demon- strate that our MGFN model outperforms all state-of-the-art methods and verify the effec- tiveness of our model. 1 Introduction Sentiment analysis has been a popular research sub- ject in natural language processing. Aspect-based sentiment analysis (ABSA) (Birjali et al., 2021) is a fine-grained sentiment analysis task. For ex- ample, given a sentence “The menu is limited but the dishes are excellent.”, there are two aspects mentioned in the sentence and the sentiment po- larity of aspects “menu” and “dishes” are negative and positive, respectively. Generally, ABSA task is formulated as predicting the polarity of a given ∗ The first two authors contributed equally to this work. † Corresponding Authors In the that just ended, Amy defeated Chelsea election In the that just ended, Chelseadefeated election Amy det advmod det advmod case case nsubj nsubj amod (a) The dependency parse tree from spaCy. 0.14 0.17 took look at the I chicken one appalled and was I 0.03 0.26 0.19 0.07 0.06 0.09 0.13 0.09 (b) The dependency tree derived by ACLT. Figure 1: (a) Two similar sentences with aspect "Amy", each with its own dependency tree. (b) An example,the numbers in arcs denote the weight of edge between aspect word and its contextual words, derived from ACLT (Zhou et al., 2021). sentence-aspect pair. The main challenge of ABSA is to precisely capture the relationship between the aspect and its corresponding opinion expressions. Many existing graph-based methods(Sun et al., 2019a; Zhao et al., 2020; Wang et al., 2020; Li et al., 2021b) have been devoted to obtaining promising …

Content truncated to 500 words