Cross-Lingual Sentiment Analysis with Machine Translation

Recent advancements in machine translation foster an interest of its use in sentiment analysis. This thesis investigates prospects and limitations of using machine translation in cross-lingual sentiment analysis. To perform a sentiment analysis we need to learn linguistic features by either using tools such as part-of-speech taggers, parsers, or basic resources such as annotated corpora or sentiment lexica. We are motivated to study the translation of existing resources in English simply because building such tools and resources for each language requires considerable human effort. This severely limits the implementation of language specific sentiment analysis techniques similar to those developed for English. Labeled corpora and sentiment lexica are two main resources in the application of sentiment analysis. We translate them to a language with limited resources where we opt to focus on improving classification accuracy when (labeled or raw) training instances are available. In some cases, however, we may not have access to any training data. To address this scenario we explore methods to translate sentiment lexica to a target language as we also try to improve machine translation performance by generating additional context. For all experiments we work on English and Turkish data which consist of movie and product reviews and we perform two-class (positive-negative) classification -polarity detection in which we discard the neutral class. Consequently, we obtain promising results in polarity detection experiments where we use general-purpose classifiers trained on translated corpora while in this point we remark that dissimilarities between two corpora in different languages should be further studied for better integration of resources. We also find quantitative evidences to suggest that lexica translation is more troublesome since the inherit differences of expressing sentiment between two languages make it harder to preserve the sentiment of words/phrases when translating them from one language to another.