Leveraging Multilingual Resources for Language Invariant Sentiment Analysis

Leveraging Multilingual Resources for Language Invariant Sentiment Analysis Allen J. Antony Arghya Bhattacharya Jaipal Singh Goud Radhika Mamidi Language Technologies Research Centre International Institute of Information Technology - Hyderabad Gachibowli, Hyderabad, Telangana-500032 {allen.antony@research., arghya.b@research., jaipal.singh@research.,radhika.mamidi@}iiit.ac.in Abstract Sentiment analysis is a widely researched NLP problem with state-of-the-art solu- tions capable of attaining human-like ac- curacies for various languages. How- ever, these methods rely heavily on large amounts of labelled data or sentiment weighted language specific lexical re- sources that are unavailable for low- resource languages. Our work attempts to tackle this data scarcity issue by in- troducing a neural architecture for lan- guage invariant sentiment analysis capable of leveraging various monolingual datasets for training without any kind of cross- lingual supervision. The proposed archi- tecture attempts to learn language agnostic sentiment features via adversarial training on multiple resource-rich languages which can then be leveraged for inferring senti- ment information at a sentence level on a low resource language. Our model out- performs the current state-of-the-art meth- ods on the Multilingual Amazon Review Text Classification dataset (Prettenhofer and Stein, 2010) and achieves significant performance gains over prior work on the low resource Sentiraama corpus (Gangula and Mamidi, 2018). A detailed analy- sis of our research highlights the ability of our architecture to perform significantly well in the presence of minimal amounts of training data for low resource languages. c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1 Introduction Sentiment analysis refers to a series of methods, techniques, and tools aimed at extracting the in- tended sentiment from a written opinion. Tradi- tional sentiment analysis techniques have relied on using supervised term weighting methods includ- ing terms’ distribution of classes, word-level po- larity scoring and using SVMs (Durant and Smith, 2006) and Naive Bayes classifiers (Prasad, 2010) for pattern extraction using hand-crafted features. The advent of deep learning techniques for senti- ment analysis has now enabled the extraction of high quality sentiment data from written texts. One majorly overlooked factor in the performance of these neoteric approaches is their dependency on large annotated datasets compiled from multiple data sources related to or sourced from newspa- pers, tweets, photos and product reviews. (Socher et al., 2013; Kim, 2014; Tai et al., 2015; Iyyer et al., 2015; Wang et al., 2016). Given global nature of the current information sharing infrastructure, most data generated be- longs to one of the three languages : English, Man- darin or Spanish. This abundance of raw data aids and motivates the creation of annotated resources in these languages. Conversely, the paucity of an- notated data in most languages makes it a challeng- ing task to develop deep learning based solutions for them. Hence there is a pressing need to pay special attention to developing solutions capable of sentiment analysis in a low resource setting. Some of the initial methods that attempt to tackle this problem of data scarcity using transfer learning (training a neural model on one language …

Content truncated to 500 words