Towards Sub-Word Level Compositions for Sentiment Analysis of Hindi-English Code Mixed Text

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2482–2491, Osaka, Japan, December 11-17 2016. Towards Sub-Word Level Compositions for Sentiment Analysis of Hindi-English Code Mixed Text Aditya Joshi1∗, Ameya Prabhu†2∗, Manish Shrivastava3 and Vasudeva Varma1 1Search and Information Extraction Lab 2Centre for Visual Information Technology 3Language Technologies Research Center International Institute of Information Technology, Hyderabad (India) {aditya.joshi, ameya.prabhu} @research.iiit.ac.in {m.shrivastava, vv} @.iiit.ac.in Abstract Sentiment analysis (SA) using code-mixed data from social media has several applications in opinion mining ranging from customer satisfaction to social campaign analysis in multilingual societies. Advances in this area are impeded by the lack of a suitable annotated dataset. We introduce a Hindi-English (Hi-En) code-mixed dataset for sentiment analysis and perform em- pirical analysis comparing the suitability and performance of various state-of-the-art SA methods in social media. In this paper, we introduce learning sub-word level representations in LSTM (Subword-LSTM) architecture instead of character-level or word-level representations. This linguistic prior in our architecture enables us to learn the information about sentiment value of important morphemes. This also seems to work well in highly noisy text containing misspellings as shown in our exper- iments which is demonstrated in morpheme-level feature maps learned by our model. Also, we hypothesize that encoding this linguistic prior in the Subword-LSTM architecture leads to the su- perior performance. Our system attains accuracy 4-5% greater than traditional approaches on our dataset, and also outperforms the available system for sentiment analysis in Hi-En code-mixed text by 18%. 1 Introduction Code Mixing is a natural phenomenon of embedding linguistic units such as phrases, words or mor- phemes of one language into an utterance of another (Muysken, 2000; Duran, 1994; Gysels, 1992). Code-mixing is widely observed in multilingual societies like India, which has 22 official languages most popular of which are Hindi and English. With over 375 million Indian population online, usage of Hindi has been steadily increasing on the internet. This opens up tremendous potential for research in sentiment and opinion analysis community for studying trends, reviews, events, human behaviour as well as linguistic analysis. Most of the current research works have involved sentiment polarity detection (Feldman, 2013; Liu, 2012; Pang and Lee, 2008) where the aim is to identify whether a given sentence or document is (usually) positive, negative or neutral. Due to availability of large-scale monolingual corpora, resources and widespread use of the language, English has attracted the most attention. Seminal work in sentiment analysis of Hindi text was done by Joshi et al. (2010) in which the authors built three step fallback model based on classification, machine translation and sentiment lexicons. They also observed that their system performed best with unigram features without stemming. Bakliwal et al. (2012) generated a sentiment lexicon for Hindi and validated the results on translated form of Amazon Product Dataset Blitzer et al. (2007). Das and Bandyopadhyay (2010) created Hindi SentiWordNet, a sentiment lexicon for Hindi. ∗* indicates these authors contributed equally to this work. †Corresponding Author This work is licensed under …

Content truncated to 500 words