Solving Data Sparsity for Aspect Based Sentiment Analysis Using Cross-Linguality and Multi-Linguality

Proceedings of NAACL-HLT 2018, pages 572–582 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Solving Data Sparsity for Aspect based Sentiment Analysis using Cross-linguality and Multi-linguality Md Shad Akhtar∗, Palaash Sawant+, Sukanta Sen∗, Asif Ekbal∗ and Pushpak Bhattacharyya∗ ∗ Department of Computer Science & Engineering Indian Institute of Technology Patna, India {shad.pcs15,sukanta.pcs15,asif,pb}@iitp.ac.in + Goa University, palaash77@gmail.com Abstract Efficient word representations play an impor- tant role in solving various problems related to Natural Language Processing (NLP), data min- ing, text mining etc. The issue of data spar- sity poses a great challenge in creating effi- cient word representation model for solving the underlying problem. The problem is more intensified in resource-poor scenario due to the absence of sufficient amount of corpus. In this work, we propose to minimize the effect of data sparsity by leveraging bilingual word embeddings learned through a parallel corpus. We train and evaluate Long Short Term Mem- ory (LSTM) based architecture for aspect level sentiment classification. The neural network architecture is further assisted by the hand- crafted features for the prediction. We show the efficacy of the proposed model against state- of-the-art methods in two experimental setups i.e. multi-lingual and cross-lingual. 1 Introduction Sentiment analysis (Pang and Lee, 2005) tries to automatically extract the subjective information from a user written textual content and classifies it into one of the predefined set of classes, e.g. posi- tive, negative, neutral or conflict. Sentiment anal- ysis performed on coarser level (i.e. document or sentence level) does not provide enough informa- tion for a user who is critical of finer details such as battery life of a laptop or service of a restau- rant etc. Aspect level sentiment analysis (ABSA) (Pontiki et al., 2014) serves such a purpose, which first identifies the features (or aspects) mentioned in the text and then classifies it into one of the tar- get classes. For example, the following review is for a restaurant where the writer shares her/his ex- perience. Though s/he likes the food but certainly not happy with the service. One of the best food we had in a while but the service was very disappointing. Analyzing such reviews on sentence level will re- flect only an overall sentiment (i.e. conflict) of the sentence ignoring critical information such as food and service qualities. However, ABSA will first identify all the aspects in the text (i.e. food and ser- vice) and then associate positive with food and neg- ative with service. Identification of aspect terms is also known as aspect term extraction or opin- ion target extraction. In this work, we focus on the second problem i.e. aspect level sentiment classi- fication. Literature survey suggests a wide range of re- search on sentiment analysis (at the document or sentence level) is being carried out in recent years (Turney, 2002; Kim and Hovy, 2004; Jagtap and Pawar, 2013; Poria et al., 2016; Kaljahi and Foster, 2016; Gupta et al., 2015). However, most of these researches are focused on resource-rich language …

Content truncated to 500 words