Cross-lingual Aspect-based Sentiment Analysis with Aspect Term Code-Switching

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9220–9230 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 9220 Cross-lingual Aspect-based Sentiment Analysis with Aspect Term Code-Switching ∗ Wenxuan Zhang1, Ruidan He2, Haiyun Peng2, Lidong Bing2 and Wai Lam1 1The Chinese University of Hong Kong 2DAMO Academy, Alibaba Group {wxzhang,wlam}@se.cuhk.edu.hk {ruidan.he,haiyun.p,l.bing}@alibaba-inc.com Abstract Many efforts have been made in solving the Aspect-based sentiment analysis (ABSA) task. While most existing studies focus on English texts, handling ABSA in resource-poor lan- guages remains a challenging problem. In this paper, we consider the unsupervised cross- lingual transfer for the ABSA task, where only labeled data in the source language is available and we aim at transferring its knowledge to the target language having no labeled data. To this end, we propose an alignment-free label pro- jection method to obtain high-quality pseudo- labeled data of the target language with the help of the translation system, which could pre- serve more accurate task-specific knowledge in the target language. For better utilizing the source and translated data, as well as enhanc- ing the cross-lingual alignment, we design an aspect code-switching mechanism to augment the training data with code-switched bilingual sentences. To further investigate the impor- tance of language-specific knowledge in solv- ing the ABSA problem, we distill the above model on the unlabeled target language data which improves the performance to the same level of the supervised method. 1 Introduction Aspect-based Sentiment Analysis (ABSA) is the task of extracting mentioned aspects from a given sentence and predicting their corresponding senti- ment polarities1 (Liu, 2012; Pontiki et al., 2014). Consider the following example, “The food is great, but the service is kinda disappointing”, we can de- tect two mentioned aspect terms “food” and “ser- vice”, and judge their corresponding sentiments as positive and negative, respectively. Given its ∗Work done when Wenxuan Zhang was an intern at Al- ibaba. This work was supported by Alibaba Group through Alibaba Research Intern Program, and a grant from the Re- search Grant Council of the Hong Kong Special Administra- tive Region, China (Project Codes: 14204418). 1Also called End-to-End ABSA or Unified ABSA wide application scenarios, it has attracted lots of attention in the NLP community in recent years (Li et al., 2019a; He et al., 2019; Hu et al., 2019; Chen and Qian, 2020; Liang et al., 2020; Mao et al., 2021; Zhang et al., 2021). The majority of existing ABSA studies are con- ducted on English texts. However, in real-world scenarios such as the E-commerce website, users’ opinions are usually expressed in different lan- guages (Pontiki et al., 2016; Keung et al., 2020). Manually annotating a large quantity of ABSA data for every language can be extremely costly. In this work, we investigate the unsupervised cross-lingual transfer for the ABSA task, where we only have labeled data in the source language and aim to transfer the knowledge to target languages whose labeled ABSA data is unavailable. Existing works on cross-lingual ABSA mainly focus on its subtasks, including …

Content truncated to 500 words