Bangla E-Mail Body to Subject Generation Using Sequence to Sequence RNNs

Abstract

The development of subjects has become one of the major problems facing deep learning and natural language processing in recent years. A brief comment on a lengthy email body is condensed in the subject generation. Our goal is to develop a Bengali subject generator that is effective and efficient and can produce a clear and insightful subject from a given Bengali email body. To do this, we have gathered a variety of emails body, including educational, commercial, etc. and will use our model to generate subject from those texts. In the encoding layer of our model, bi-directional RNNs are employed, while the decoding layer makes use of LSTMs and an attention model. Our model generates subject using a sequence-to-sequence model. While developing this model, we encountered difficulties with text pre-processing, vocabulary and missing words counting, word embedding, detecting new terms and other tasks. Our primary objectives in this model were to generate a subject and lessen its train loss. In our study, we successfully reduced the train loss to 0.001 by producing a smooth concise subject from a provided email body.