Generative Cross-Domain Data Augmentation for Aspect and Opinion Co-Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4219 - 4229 July 10-15, 2022 ©2022 Association for Computational Linguistics Generative Cross-Domain Data Augmentation for Aspect and Opinion Co-Extraction Junjie Li, Jianfei Yu∗, and Rui Xia∗ School of Computer Science and Engineering, Nanjing University of Science and Technology, China {jj_li, jfyu, rxia}@njust.edu.cn Abstract As a fundamental task in opinion mining, as- pect and opinion co-extraction aims to identify the aspect terms and opinion terms in reviews. However, due to the lack of fine-grained an- notated resources, it is hard to train a robust model for many domains. To alleviate this is- sue, unsupervised domain adaptation is pro- posed to transfer knowledge from a labeled source domain to an unlabeled target domain. In this paper, we propose a new Generative Cross-Domain Data Augmentation framework for unsupervised domain adaptation. The pro- posed framework is aimed to generate target- domain data with fine-grained annotation by ex- ploiting the labeled data in the source domain. Specifically, we remove the domain-specific segments in a source-domain labeled sentence, and then use this as input to a pre-trained sequence-to-sequence model BART to simul- taneously generate a target-domain sentence and predict the corresponding label for each word. Experimental results on three datasets demonstrate that our approach is more effec- tive than previous domain adaptation meth- ods. The source code is publicly released at https://github.com/NUSTM/GCDDA. 1 Introduction Aspect and opinion co-extraction is a fundamental task in opinion mining. It aims to extract aspect terms and opinion terms from review sentences. Given a review “the keyboard is great”, the aspect term is keyboard and the opinion term is great. Most existing works formulate aspect and opinion co-extraction as a sequence labeling task. With the support of deep learning, many supervised methods have achieved remarkable results (Liu et al., 2015; Yin et al., 2016; Wang et al., 2017; Wu et al., 2020). However, owing to the high cost of fine-grained annotations, the scarcity of labeled data in many new domains makes them fail to obtain a robust performance. ∗Corresponding authors. To alleviate the deficiency of labeled data, unsu- pervised domain adaptation is proposed to transfer knowledge from a labeled source domain to an un- labeled target domain. The main difficulty of unsu- pervised domain adaptation lies in the distribution discrepancy between data from different domains. Specifically, reviews from different domains may have distinct aspect terms, opinion terms, and ex- pression patterns. To tackle the distribution discrep- ancy problem, many domain adaptation methods were proposed for coarse-grained sentiment classi- fication (Blitzer et al., 2007; Yu and Jiang, 2016; Ziser and Reichart, 2018; Ghosal et al., 2020). However, due to the complexity of fine-grained domain adaptation for aspect and opinion extrac- tion, only a handful of studies attempt to address this issue. Most of them are based on the following three paradigms: • rule-based adaptation, which iteratively ex- tracts aspects and opinions in sentences based on domain-independent syntactic rules and opinion dictionary …

Content truncated to 500 words