Borrow a Little from your Rich Cousin: Using Embeddings and Polarities of English Words for Multilingual Sentiment Classification

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3053–3062, Osaka, Japan, December 11-17 2016. Borrow a Little from your Rich Cousin: Using Embeddings and Polarities of English Words for Multilingual Sentiment Classification Prerana Singhal and Pushpak Bhattacharyya Dept. of Computer Science and Engineering IIT Bombay, Maharashtra, India {singhal.prerana,pushpakbh}@gmail.com Abstract In this paper, we provide a solution to multilingual sentiment classification using deep learning. Given input text in a language, we use word translation into English and then the embeddings of these English words to train a classifier. This projection into the English space plus word embed- dings gives a simple and uniform framework for multilingual sentiment analysis. A novel idea is augmentation of the training data with polar words, appearing in these sentences, along with their polarities. This approach leads to a performance gain of 7-10% over traditional classifiers on many languages, irrespective of text genre, despite the scarcity of resources in most languages. 1 Introduction Sentiment Analysis deals with extraction of opinion polarity from texts. Extensive work has been done in this field over the past years, mainly for English. Subsequently, rich English resources like SentiWord- Net (Esuli and Sebastiani, 2006) and pre-trained word embeddings (Mikolov et al., 2013a) are available publicly. However, lack of rich resources and annotated corpora in other languages like Dutch, Russian or Hindi makes it difficult to analyze texts with as good accuracy as that in English. Deep learning models have achieved astonishing results in several fields like Speech Recognition and Computer vision, and have shown promising results when used for several NLP tasks (like Convolutional Neural Networks for Sentence Classification (Kim, 2014), LSTMs for tweet classification (Wang et al., 2015) and Recursive Deep Models for Sentiment Analysis (Socher et al., 2013)). A key feature of deep learning models, which seems to attract NLP researchers, is their lack of demand for manual feature engineering, unlike other classical machine learning algorithms (SVM, etc.) (Pang et al., 2002). Multilingual Sentiment Analysis has been a challenging, yet an important area of research since a long time, mainly involving other NLP tools like Sentence-level Machine translation (Joshi et al., 2010; Wan, 2009), Machine Translation with SentiWordNet scores (Denecke, 2008), semantic orientation calculator (Brooke et al., 2009), and Wordnets (Balamurali et al., 2012) to serve the purpose. Machine learning methods have been used and evaluated on different sets of languages (Boiy and Moens, 2009; Seki et al., 2010; Seki et al., 2008), which involve many constraints and manual functionalities. In addition, strategies to build a multilingual corpus (Schulz et al., 2010) have been devised which, quite frankly, are not possible for large number of languages, owing to the diversity involved. In this paper, we present a simple approach to multilingual sentiment classification which: (a) poses minimal restrictions on the language or the text genre to be used, (b) has minimal demands for pre- processing tools, and (c) shows no aversion to the small size of datasets or inadequacy of resources in any language. …

Content truncated to 500 words