Question Answering Using Deep Neural Networks: Single Turn and Beyond
Join Zoom Meeting
https://nus-sg.zoom.us/j/99835670478?pwd=ajZ4cDh6SlNDTzlVTDFMR0RLWFJzZz09
Meeting ID: 998 3567 0478
Password: 121314
Abstract:
In this thesis, we tackle the task of question answering (QA). Machine comprehension (MC) systems mimic the process of reading comprehension (RC) by answering questions after understanding a natural language text. We first present a deep neural network model for MC-based QA. We develop an end-to-end question-focused multi-factor attention network for answer extraction. Our proposed multi-factor attentive encoding helps to aggregate relevant evidence by using a tensor-based multi-factor attention mechanism. Due to the proposed question-focused attention pointing mechanism, it also learns to focus on the important question words to encode the aggregated question vector. Our proposed model achieves significant improvements over prior work on three large-scale challenging QA datasets.
The second task that we have tackled is nil-aware answer extraction for machine reading comprehension. For a given question, the associated passage might not contain any valid answer. These questions are defined as nil questions. Most of the recently proposed neural models do not consider nil questions, although it is crucial for a practical QA system to be able to determine whether a text passage contains a valid answer for a question. We focus on developing models that extract an answer for a question if and only if the associated text passage contains a valid answer. Otherwise, they are expected to return Nil as answer. We propose a nil-aware answer extraction framework that is capable of returning Nil or a text span from the associated passage as an answer in a single step. In addition to our proposed MC model, we show that our proposed framework can be easily integrated with several other recently proposed QA models developed for reading comprehension and can be trained in an end-to-end fashion. The proposed framework decomposes pieces of evidence into relevant and irrelevant parts and then combines them to infer the existence of any answer. Experiments show that the integration of our proposed framework significantly outperforms several strong baseline systems that use pipeline or threshold-based approaches.
Finally, we focus on developing a system for multi-turn conversational question answering. Recently proposed conversational question answering systems lack the ability to ask a follow-up clarification question when a given question is underspecified. In this work, we focus on developing a conversational question answering system that can predict the answer to a question in a conversation, and has the ability to ask a follow-up clarification question if the question is underspecified. We propose a pipeline approach which consists of an answer prediction model and a follow-up question generation model. The answer prediction model is based on a dual co-attention network while the follow-up question generation model is based on a sequence-to-sequence neural network enhanced with a copying mechanism. Experiments on the ShARC dataset show the effectiveness of the proposed system.