Learnable Dependency-based Double Graph Structure for Aspect-based Sentiment Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 7086–7092 October 12–17, 2022. 7086 Abstract Dependency tree-based methods might be susceptible to the dependency tree due to that they inevitably introduce noisy infor- mation and neglect the rich relation information between words. In this paper, we propose a learnable dependency-based double graph (LD2G) model for aspect- based sentiment classification. We use multi-task learning for domain adaptive pretraining, which combines Biaffine At- tention and Mask Language Model by incorporating features such as structure, re- lations and linguistic features in the sentiment text. Then we utilize the depend- ency enhanced double graph-based MPNN to deeply fuse structure features and rela- tion features that are affected with each other for ASC. Experiment on four bench- mark datasets shows that our model is superior to the state-of-the-art approaches. 1 Introduction Aspect-based sentiment analysis (ABSA) is a fine- grained sentiment analysis task for extracting the emotional orientation of the given one or more as- pects in a sentence. ABSA is generally divided into two parts: aspect extraction (AE) and aspect senti- ment classification (ASC). We will focus on ASC task in this paper. Dependency-based approaches have illustrated promising performance by utilizing the spatial structure of the dependency tree (DT). Early stud- ies focus on the syntax feature enhanced aspect- based sentiment analysis (ABSA) models (Brun et al., 2014;Kiritchenko et al., 2014). Dependency- based analysis currently tends to use graph neural networks to learn representation over DT due to that graph neural network can directly learn on graph data including DTs (Kipf et al., 2016). DT is regarded as an undirected tree structure to produce a symmetry adjacency matrix for Graph Convolu- tional Networks (GCN) (Sun et al., 2019; Zhang et al., 2019). (Wang et al., 2020) use GNNs to capture implicit relation between aspect words and others and achieved encouraged results. (Wang et al., 2021) utilized reinforcement learning for aspect- based sentiment classification (ASC). In this paper, we consider how to effectively fuse structure and relation features for ASC. the follow- ing two questions: 1) How to utilize relations of edges between words for improving ASC? Most of existing approaches consider the structure infor- mation and neglected the rich edge relations between words with exception to the works (Chen et al., 2022; Tian et al., 2021a; Tian et al., 2021b). However, these works consider graph features (i.e., structure, edge relations, tree distance) as different input sources, and do not explore how to effectively fuse various features because these features are af- fected with each other. 2) How to effectively handle the multi-feature noisy information from DT for improving ASC? Existing approaches directly uti- lize the DT generated by a tool like spaCy. Direct application of DT is susceptible to noisy infor- mation in the dependency tree (Tang et al., 2020), which may achieve lower performance compared to using only the flat structure. DT includes 45 kinds of relations such as nsubj, amod and conj, and so on. Dependency parsing requires effective way to reduce …

Content truncated to 500 words