Multi-Supervised LSTM for Bengali Text Sentiment Analysis
Syed Muaz Ali, Afif Ibna Kadir Khan Turja, Sabiha Tahseen, Md Humaion Kabir Mehedi, Md Sabbir Hossain, Annajiat Alim Rasel · 2023
Bangla text sentiment analysis is one of the challenges of recent years. Various machine learning and deep learning approaches have been made to further improve the classification models’ performance to achieve state-of-the art results. Since text sentiment analysis is a sequential task, Long Short-term memory have provided great results in these cases. In this paper, we propose a methodology to tweak an existing LSTM model to gain further improvement in results. As Bengali Text Sentiment analysis is a sequence-to-one task, we discuss how using the features of the entire sequence generated from LSTM along with the last hidden state’s output can improve the results. We compare with a single LSTM layer where the last hidden state is used for producing the final result with our methodology where also a single LSTM layer is used but with additional feature engineering and multi-supervision is applied on the generated sequence and final hidden state to produce the results. Our final results show that adding these tweaks to an existing LSTM layer can increase the overall performance of the model.