Cross-Lingual Sentiment Analysis Without (Good) Translation

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 506–515, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Cross-Lingual Sentiment Analysis Without (Good) Translation Mohamed Abdalla and Graeme Hirst msa@ and gh@cs.toronto.edu Department of Computer Science, University of Toronto Toronto, Canada Abstract Current approaches to cross-lingual senti- ment analysis try to leverage the wealth of labeled English data using bilingual lexi- cons, bilingual vector space embeddings, or machine translation systems. Here we show that it is possible to use a sin- gle linear transformation, with as few as 2000 word pairs, to capture fine-grained sentiment relationships between words in a cross-lingual setting. We apply these cross-lingual sentiment models to a di- verse set of tasks to demonstrate their functionality in a non-English context. By effectively leveraging English sentiment knowledge without the need for accurate translation, we can analyze and extract features from other languages with scarce data at a very low cost, thus making sen- timent and related analyses for many lan- guages inexpensive. 1 Introduction Methods for sentiment analysis and classification have largely been limited to English, making use of large amounts of labeled data to produce senti- ment classification. As a consequence, many de- veloped approaches cannot be readily applied to other languages, which usually do not have the wealth of labeled data that is exclusive to English. Therefore many approaches which deal with other languages often: i) experiment with small datasets that are limited in domain or size of training and testing sets (Lee and Renganathan, 2011; Tan and Zhang, 2008), or ii) attempt to elucidate sentiment lexicons for their respective languages (Moham- mad et al., 2016). A growing number of publications attempt to leverage labeled English data to compensate for the relative lack of training material in the other languages. This is usually done through the use of either bilingual lexicons (Balamurali et al., 2012), machine translation (MT) systems (Salameh et al., 2015; Zhou et al., 2016), or more recently, through the use of bilingual vector space embeddings (Chen et al., 2016). Unfortunately, in some cases such data is still expensive to obtain. Many languages do not have good, or sometimes any, MT systems, and the cost of producing word alignments or sen- tence alignments for training bilingual word em- beddings (BWE) (Zou et al., 2013; Bengio and Corrado, 2015) or similar techniques (Jain and Ba- tra, 2015) is prohibitive for data-poor languages. Here we introduce a high-performance, low- cost approach to cross-lingual sentiment classifi- cation, which can be used to benchmark more ex- pensive methods. We demonstrate the utility of this approach by highlighting how very limited training data suffices for effective cross-lingual sentiment analysis in various contexts (both at the word and sentence/document level). Our approach relies on the simple vector space translation matrix method (Mikolov et al., 2013a), which computes a matrix to convert from the vector space of one language to that of another. It hinges on the obser- vation that sentiment is highly “preserved” even …

Content truncated to 500 words