SnTiEmd: Sentiment Specific Embedding Model Generation and Evaluation for a Resource Constraint Language

Abstract

Sentiment analysis is primitive natural language processing (NLP) research for resource-constrained languages where feature extraction in a specific domain is a challenging issue. The word embeddings are good intermediate text to feature extraction methods where capturing semantic regularities between words. However, the performance of general word embeddings is limited in capturing domain-specific semantics or knowledge in sentiment analysis. Moreover, for low-resource languages like Bengali, no sentiment-specific word embedding research has been conducted to date. This study developed a domain-based, e.g., sentiment-specific embedding corpus (SeC) and built an intrinsic evaluation dataset. The three embedding methods (i.e., GloVe, FastText, Word2Vec) are investigated to develop the sentiment-based embedding model (SnTiEmd). The SnTiEmd (i.e., GloVe, fastText, Word2Vec) models are evaluated using an intrinsic evaluation dataset (i.e., semantic and syntactic). The highest accuracy of Pearson correlation for syntactic similarity is ($55.66\%$) and semantic similarity ($52.97\%$), whereas the maximum accuracy for spearman correlation is ($52.28\%$) and ($55.19\%$) for syntactic and semantic word similarity using GloVe-based SnTiEmd, respectively.