Open domain question answering system

• Main Author: Hoang, Nghia Tuyen
• Other Authors: Joty Shafiq Rayhan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Information systems::Information storage and retrieval; Science::Mathematics
• Online Access: https://hdl.handle.net/10356/156955
• Institution: Nanyang Technological University

Deep learning methods have drawn tremendous attention from both the research community and the industrial practitioners thanks to their undeniable power in learning feature representation in higher dimensions without manual, handcrafting features. An application of deep learning that arises naturally is question answering, in which a question answering system must answer questions posed by humans. One of its sub-fields, opendomain question answering, attempts to answer questions about nearly anything, without being given relevant reference texts. Despite its impactful applications in search engines, chatbots and factual correction, research work in open-domain question answering is relatively under-explored due to its complex and large-scale nature. In this work, we aim to advance the progress of recent open-domain question answering systems by developing various mathematical-driven methods. More specifically, in the first part of this thesis, we introduce the widely adopted two-stage paradigm in opendomain question answering and perform comprehensive error analysis on state-of-the-art models. Based on this, we are then able to formulate and develop methods aiming specifically at overcoming these weaknesses in the second part of the thesis. These approaches range from simple methods such as parameter sharing and data augmentation to more sophisticated methods such as designing new objective functions or pseudo data synthesis and semi-supervised learning. Finally, we unify these developed methods into a single framework that outperforms state-of-the-art models by a significant margin on common benchmarking datasets. The code to reproduce our experiments is released at https://github.com/hnt4499/DPR.