Cross-Lingual Unsupervised Sentiment Classification with Multi-View Transfer Learning

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5759–5771 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5759 Cross-Lingual Unsupervised Sentiment Classification with Multi-View Transfer Learning Hongliang Fei, Ping Li Cognitive Computing Lab Baidu Research 1195 Bordeaux Dr, Sunnyvale, CA 94089, USA 10900 NE 8th St, Bellevue, WA 98004, USA {hongliangfei,liping11}@baidu.com Abstract Recent neural network models have achieved impressive performance on sentiment classifi- cation in English as well as other languages. Their success heavily depends on the availabil- ity of a large amount of labeled data or paral- lel corpus. In this paper, we investigate an ex- treme scenario of cross-lingual sentiment clas- sification, in which the low-resource language does not have any labels or parallel corpus. We propose an unsupervised cross-lingual sen- timent classification model named multi-view encoder-classifier (MVEC) that leverages an unsupervised machine translation (UMT) sys- tem and a language discriminator. Unlike pre- vious language model (LM) based fine-tuning approaches that adjust parameters solely based on the classification error on training data, we employ the encoder-decoder framework of a UMT as a regularization component on the shared network parameters. In particular, the cross-lingual encoder of our model learns a shared representation, which is effective for both reconstructing input sentences of two languages and generating more representative views from the input for classification. Exten- sive experiments on five language pairs verify that our model significantly outperforms other models for 8/11 sentiment classification tasks. 1 Introduction Recent neural network models have achieved re- markable performance on sentiment classification in English and other languages (Conneau et al., 2017; Chen et al., 2018; He et al., 2019; Chen and Qian, 2019). However, their success heavily depends on the availability of a large amount of labeled data or parallel corpus. In reality, some low-resource languages or applications have lim- ited labeled data or even without any labels or par- allel corpus, which may hinder us from training a robust and accurate sentiment classifier. To build sentiment classification models for low-resource languages, recent researchers de- veloped cross-lingual text classification (CLTC) models (Xu and Yang, 2017; Eriguchi et al., 2018), which transfers knowledge from a resource-rich (source) language to a low-resource (target) lan- guage. The core of those models is to learn a shared language-invariant feature space that is indicative of classification for both languages. Therefore a model trained from the source lan- guage can be applied to the target language. Based on how the shared feature space is learned, there are three categories, namely word-level align- ments (Andrade et al., 2015), sentence-level align- ments (Eriguchi et al., 2018) and document level alignments (Zhou et al., 2016). Those models can well capture the semantic similarity between two languages. They, however, require parallel resources such as a bilingual dictionary, parallel sentences, and parallel Wikipedia articles. Such a limitation may prevent these models from be- ing applicable in languages without any parallel resources. Recently, there have been several attempts at de- veloping “zero-resource” models (Ziser and Re- ichart, 2018; Chen …

Content truncated to 500 words