Learning Bilingual Sentiment Word Embeddings for Cross-language Sentiment Classification

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 430–440, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Learning Bilingual Sentiment Word Embeddings for Cross-language Sentiment Classification Huiwei Zhou, Long Chen, Fulin Shi, and Degen Huang School of Computer Science and Technology Dalian University of Technology, Dalian, P.R. China {zhouhuiwei,huangdg}@dlut.edu.cn {chenlong.415,shi fl}@mail.dlut.edu.cn Abstract The sentiment classification performance relies on high-quality sentiment resources. However, these resources are imbalanced in different languages. Cross-language sentiment classification (CLSC) can lever- age the rich resources in one language (source language) for sentiment classifica- tion in a resource-scarce language (target language). Bilingual embeddings could eliminate the semantic gap between two languages for CLSC, but ignore the senti- ment information of text. This paper pro- poses an approach to learning bilingual sentiment word embeddings (BSWE) for English-Chinese CLSC. The proposed B- SWE incorporate sentiment information of text into bilingual embeddings. Further- more, we can learn high-quality BSWE by simply employing labeled corpora and their translations, without relying on large- scale parallel corpora. Experiments on NLP&CC 2013 CLSC dataset show that our approach outperforms the state-of-the- art systems. 1 Introduction Sentiment classification is a task of predicting sen- timent polarity of text, which has attracted consid- erable interest in the NLP field. To date, a num- ber of corpus-based approaches (Pang et al., 2002; Pang and Lee, 2004; Kennedy and Inkpen, 2006) have been developed for sentiment classification. The approaches heavily rely on quality and quan- tity of the labeled corpora, which are considered as the most valuable resources in sentiment classi- fication task. However, such sentiment resources are imbalanced in different languages. To leverage resources in the source language to improve the sentiment classification performance in the target language, cross-language sentiment classification (CLSC) approaches have been investigated. The traditional CLSC approaches employ ma- chine translation (MT) systems to translate corpo- ra in the source language into the target language, and train the sentiment classifiers in the target lan- guage (Banea et al., 2008). Directly employing the translated resources for sentiment classifica- tion in the target language is simple and could get acceptable results. However, the gap between the source language and target language inevitably impacts the performance of sentiment classifica- tion. To improve the classification accuracy, multi- view approaches have been proposed. In these ap- proaches, the resources in the source language and their translations in the target language are both used to train sentiment classifiers in two indepen- dent views (Wan, 2009; Gui et al., 2013; Zhou et al., 2014a). The final results are determined by en- semble classifiers in these two views to overcome the weakness of monolingual classifiers. However, learning language-specific classifiers in each view fails to capture the common sentiment information of two languages during training process. With the revival of interest in deep learning (Hinton and Salakhutdinov, 2006), shared deep representations (or embeddings) (Bengio et al., 2013) are employed for CLSC (Chandar A P et al., …

Content truncated to 500 words