Toward a Multi-domain Marathi SentiWordNet: Word Polarity Transfer and Ensemble-Based Evaluation

Construction of sentiment lexicon for morphologically rich low resource Indian language like Marathi is a challenging task. An approach is proposed with TF-IDF word embedding and ensemble-based machine learning. Language specific sentiment analysis is essential due to diversity respective to grammar, vocabulary, and location. For Marathi language there is no Sentiment WordNet available. Based on the fact that polarity of word remains same across languages, the seed Marathi SentiWordNet is derived from existing English and Hindi SentiWordNet. Polarity corpora of different domain are processed individually and the seed Marathi lexicon is evaluated using different features based on sentiment polarity score and positive and negative word count. The proposed ensemble-based model performed very well for Bag of Word feature with up to 90% accuracy. When TF-IDF feature is used the performance is ranging from 63 to 76% and polar words emerged with top TF-IDF score. Contextual sentiment analysis for Marathi text will be possible with the help of this Marathi SentiWordNet.