A FRAMEWORK FOR QUESTION ANSWERING SYSTEM USING DYNAMIC CO-ATTENTION NETWORKS

Abstract

Question answering (QA) systems have evolved exponentially over the past few years and have reached a reliable human standard. Attention mechanisms, as well as other methods of deep learning, paved the way for this development. But, because of their single-pass nature, they are incapable of recovering from local maxima matching to incorrect answers. Dynamic coattention network (DCN) is used to answer this issue. But as it has only one layer, the ability of the DCN to write diverse input representations is limited. We proposed a few modifications to DCN to overcome these findings. First, we used a bidirectional long short-term memory network (biLSTM) to encode the question and document. Next, we applied the concept of self-attention to DCN by using multiple coattention layers. This helps the encoder to generate more profuse input representations. Lastly, we combine outputs from these layers; this improves the long-range dependencies. We built a question answering system based on this multiattention DCN and tested on one of our course documents. On Stanford question answering dataset (SQuAD), this system improves the F1 mean on validation to 79.9% from its previous state of art at 75.6%.