Classification Benchmarks for Under-resourced Bengali Language based on Multichannel Convolutional-LSTM Network

Exponential growths of social media and micro-blogging sites not only provide\nplatforms for empowering freedom of expressions and individual voices but also\nenables people to express anti-social behaviour like online harassment,\ncyberbullying, and hate speech. Numerous works have been proposed to utilize\nthese data for social and anti-social behaviours analysis, document\ncharacterization, and sentiment analysis by predicting the contexts mostly for\nhighly resourced languages such as English. However, there are languages that\nare under-resources, e.g., South Asian languages like Bengali, Tamil, Assamese,\nTelugu that lack of computational resources for the NLP tasks. In this paper,\nwe provide several classification benchmarks for Bengali, an under-resourced\nlanguage. We prepared three datasets of expressing hate, commonly used topics,\nand opinions for hate speech detection, document classification, and sentiment\nanalysis, respectively. We built the largest Bengali word embedding models to\ndate based on 250 million articles, which we call BengFastText. We perform\nthree different experiments, covering document classification, sentiment\nanalysis, and hate speech detection. We incorporate word embeddings into a\nMultichannel Convolutional-LSTM (MConv-LSTM) network for predicting different\ntypes of hate speech, document classification, and sentiment analysis.\nExperiments demonstrate that BengFastText can capture the semantics of words\nfrom respective contexts correctly. Evaluations against several baseline\nembedding models, e.g., Word2Vec and GloVe yield up to 92.30%, 82.25%, and\n90.45% F1-scores in case of document classification, sentiment analysis, and\nhate speech detection, respectively during 5-fold cross-validation tests.\n