Cross Lingual Sentiment Analysis using Modified BRAE

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 159–168, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Cross-Lingual Sentiment Analysis using modified BRAE Sarthak Jain Department of Computer Engineering Delhi Technological University DL, India successar@gmail.com Shashank Batra Department of Computer Engineering Indian Institute of technology, Delhi DL, India shashankg@gmail.com Abstract Cross-Lingual Learning provides a mech- anism to adapt NLP tools available for la- bel rich languages to achieve similar tasks for label-scarce languages. An efficient cross-lingual tool significantly reduces the cost and effort required to manually an- notate data. In this paper, we use the Recursive Autoencoder architecture to de- velop a Cross Lingual Sentiment Analysis (CLSA) tool using sentence aligned cor- pora between a pair of resource rich (En- glish) and resource poor (Hindi) language. The system is based on the assumption that semantic similarity between different phrases also implies sentiment similarity in majority of sentences. The resulting sys- tem is then analyzed on a newly developed Movie Reviews Dataset in Hindi with la- bels given on a rating scale and compare performance of our system against exist- ing systems. It is shown that our approach significantly outperforms state of the art systems for Sentiment Analysis, especially when labeled data is scarce. 1 Introduction Sentiment Analysis is a NLP task that deals with extraction of opinion from a piece of text on a topic. This is used by a large number of advertising and media companies to get a sense of public opin- ion from their reviews. The ever increasing user generated content has always been motivation for sentiment analysis research, but majority of work has been done for English Language. However, in recent years, there has been emergence of increas- ing amount of text in Hindi on electronic sources but NLP Frameworks to process this data is sadly miniscule. A major cause for this is the lack of annotated datasets in Indian Languages. One solution is to create cross lingual tools be- tween a resource rich and resource poor language that exploit large amounts of unlabeled data and sentence aligned corpora that are widely available on web through bilingual newspapers, magazines, etc. Many different approaches have been identi- fied to perform Cross Lingual Tasks but they de- pend on the presence of MT-System or Bilingual Dictionaries between the source and target lan- guage. In this paper, we use Bilingually Constrained Recursive Auto-encoder (BRAE) given by (Zhang et al., 2014) to perform Cross Lingual sentiment analysis. Major Contributions of this paper are as follows: First, We develop a new Rating scale based Movie Review Dataset for Hindi. Second, a general framework to perform Cross Lingual Classification tasks is developed by modifying the architecture and training procedure for BRAE model. This model exploits the fact that phrases in two languages, that share same semantic meaning, can be used to learn language independent seman- tic vector representations. These embeddings can further be fine-tuned using labeled dataset in En- glish to capture enough class information …

Content truncated to 500 words