Abstract:
This has driven artificial intelligence (AI) development to grow in rapidly, and large
amounts of language resources are being developed for an array of languages. But Bangla
has a long way to go in terms of the contributing field of AI. Different categories of Bangla
Facebook comments, such as Not Bully, Troll, Sexual and Religious are reviewed in this
study specifically for the Bangla language resources. Using more than twenty-five
thousand comments, we experimented and optimized various models such as Bangla
BERT, GRU, LSTM and CNN. In our experimental results, the best performing Bangla
BERT model reached an accuracy of 80% on the test dataset, whereas GRU achieved 70%, LSTM with 65% and finally CNN achieved just around 66%. We noticed ingrained biases in the dataset too. This can be useful for Bangla AI Resources which can be further utilized in sentiment analysis and content moderation systems to aid Bangla Speakers both domestically as well as globally.
