Development of Sentiment Lexicon in Bengali utilizing Corpus and Cross-lingual Resources

Bengali, one of the most spoken languages, lacks tools and resources for sentiment analysis. To date, the Bengali language does not have any sentiment lexicon of its own; only the translated versions of English lexica are available. Therefore, in this work, we focus on developing a Bengali sentiment lexicon from a large Bengali review corpus utilizing a cross-lingual approach. To build the sentiment dictionary, we first created a Bengali corpus of around 42000 drama reviews; among them, we manually annotated around 12000 reviews. Utilizing a machine translation system, labeled and unlabeled Bengali review corpus, English sentiment lexica, pointwise mutual information (PMI), and supervised machine learning (ML) classifiers in different phases, we develop a Bengali sentiment lexicon of around 1000 sentiment words. We compare the coverage of our lexicon with the translated English lexica in two evaluation datasets. The proposed lexicon achieves 70%-74% coverage in document-level and around 65% coverage in word-level, which is approximately 30%-100% improvement over the translated lexica in word-level and 30%-50% in document-level. The results demonstrate that our developed lexicon is highly effective in recognizing sentiments in the Bengali text.