A silver-tongued writer: a deep learning and NLP based Bangla sentence composer with proper linguistic affection

Abstract

Communication is one of the most essential parts of human life. With the advent of the 21st century, the usage of online interaction has expanded significantly. Text-based communication is prevalent in many aspects of our lives. For example, social media, email, and business invitations. It has also allowed us to maintain relationships with others, irrespective of time or geographic location. While communicating, people typically overlook the text-style representation. This establishes a barrier between the sender and receiver, preventing the message from being correctly interpreted by the receiver. Due to a lack of proper selection of words, one might face challenges in getting the expected reply from the receiver. In addition, erroneous interpretations of these writings might have detrimental psychological impacts on individuals. The fundamental objective of this paper is to develop a Bangla sentence composer model capable of detecting impolite statements in written text and transferring the style to politeness with proper affection. We begin by discussing several classification approaches. For this purpose, we used a combination of deep learning and transformer-based techniques to identify linguistic politeness from Bangla sentences. Simple RNN, CNN BiLSTM, LSTM, GRU, and various pre-trained versions of BERT were the algorithms we employed. Among all models, BERT produced much superior outcomes than the others. BanglaBERT Showed the best result with an accuracy of 84%. Also, we constructed our own dictionary to replace the impolite and harsh words. After correctly recognizing the expression, we intend to replace harsh impolite words from input sentences using our custommade dictionary. The dictionary consists of 410 impolite, harsh words and their corresponding polite version. After that, the result is modified using the BanglaT5 BanglaParaphrase, a Bangla text-to-text generation model. With the aid of this combined deep learning approach, we attempt to identify the optimum sentence variant that does not negatively affect human emotion. Using these techniques, we can achieve preferable results that will earn the recipient’s appreciation through our preferred sentence composer approach.