Towards a Unified End-to-End Approach for Fully Unsupervised Cross-Lingual Sentiment Analysis

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 1035–1044 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 1035 Towards a Unified End-to-End Approach for Fully Unsupervised Cross-lingual Sentiment Analysis Yanlin Feng and Xiaojun Wan Wangxuan Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {fengyanlin,wanxiaojun}@pku.edu.cn Abstract Sentiment analysis in low-resource languages suffers from the lack of training data. Cross- lingual sentiment analysis (CLSA) aims to improve the performance on these languages by leveraging annotated data from other lan- guages. Recent studies have shown that CLSA can be performed in a fully unsupervised manner, without exploiting either target lan- guage supervision or cross-lingual supervi- sion. However, these methods rely heavily on unsupervised cross-lingual word embeddings (CLWE), which has been shown to have seri- ous drawbacks on distant language pairs (e.g. English - Japanese). In this paper, we propose an end-to-end CLSA model by leveraging un- labeled data in multiple languages and mul- tiple domains and eliminate the need for un- supervised CLWE. Our model applies to two CLSA settings: the traditional cross-lingual in-domain setting and the more challenging cross-lingual cross-domain setting. We em- pirically evaluate our approach on the multi- lingual multi-domain Amazon review dataset. Experimental results show that our model out- performs the baselines by a large margin de- spite its minimal resource requirement. 1 1 Introduction While English sentiment analysis has achieved great success with the help of large-scale anno- tated corpus, this is not the case for most of languages where only limited data is available. Cross-lingual sentiment analysis (CLSA) tackles this problem by adapting the sentiment resource in a source language to a poor-resource language (the target language). Current state-of-the-art CLSA methods rely heavily on cross-lingual word embeddings (CLWE) to transfer sentiment information from 1The source code is available at https://github. com/Evan-Feng/UXSenti the source language to the target language. CLWE encodes words from multiple languages in a common space, thus making it possible to share a classifier across languages. Recent studies have shown that CLWE can be obtained in an unsupervised way, i.e., without any cross-lingual resources (Zhang et al., 2017; Conneau et al., 2017; Artetxe et al., 2018). This motivates fully unsupervised CLSA approaches (Chen et al., 2018a) that do not rely on either target language supervision or cross-lingual supervision. These methods generally involve the following steps: 1. Train monolingual embeddings separately on multiple languages using monolingual unla- beled data. 2. Map the monolingual embeddings to a shared space using unsupervised CLWE methods, either adversarial training methods (Con- neau et al., 2017) or non-adversarial methods (Artetxe et al., 2018; Xu et al., 2018). 3. Train a sentiment classifier using the anno- tated corpus in the source language. However, it has been shown that the quality of unsupervised CLWE is highly sensitive to the choice of language pairs and the comparability of the monolingual data (Søgaard et al., 2018). Therefore, these methods fail when the source lan- guage and the target language have …

Content truncated to 500 words