Year: 2020
Type: dissertation
Abstract: Recently, with the development of deep learning technology, deep learning has achieved great progress, and deep neural network models have even outperformed humans in many tasks, especially in the field of natural language processing (NLP).However, the performance of deep models is still far from human performance in zero-shot learning (ZSL), where the models are asked to deal with targets belonging to unseen classes.ZSL is becoming increasingly important, as it can resolve the lack of labeled data which is a common problem in real-world applications.One approach to ZSL is to acquire shared embedding spaces of the seen and unseen classes.With such shared embedding spaces, the deep models can transfer the knowledge from seen classes to unseen classes, allowing the models to perform the ZSL tasks.This thesis takes bilingual lexicon induction (BLI) and zero-shot relation extraction (RE) in knowledge base question answering (KBQA) as two examples of research subjects in ZSL and investigates various methods utilizing the shared embedding spaces.BLI, generating word translations, is a basic application of bilingual word embedding (BWE) mapping.Research on the BLI can help to improve the understanding of embedding mapping which can be applied to zero-shot RE in KBQA.KBQA is a special type of question answering (QA) tasks, where the questions are based on knowledge bases.KBQA is closely related to real-world applications, as it provides a feasible and practical way to deal with vast real-world knowledge.RE is a key component of KBQA, identifying the relation in the question in order to understand the meaning of the question.Zero-shot RE is the RE where the test relations do not appear in the training data.The research in this thesis pushes the understanding of embedding spaces in zero-shot learning forward and improves the performance of BLI and zero-shot RE.I could not enjoy the meaningful and joyful years in doing my master research without the help of many kind seniors and my lovely friends.First and foremost, I want to thank my advisor, Professor Yoshimasa Tsuruoka.He has encouraged me when I apply for the master's project in Tsuruoka Laboratory, which bred the very beginning of my master's career.He is a successful teacher, providing the most interesting and challenging orientation tutorials.He is also an insightful scholar that he can always point out the covert but fatal problems in my work.He has been caring and supportive about my academic activities and my life and has provided many valuable chances such as a part-time job at the National Institute of Advanced Industrial Science and Technology (AIST).He has been giving suggestions but never giving compulsive orders, which I think has been the reason for the friendly and vigorous atmosphere of our laboratory.I really feel lucky that he is my advisor.I would like to thank our laboratory members.When I came to the laboratory, I could hardly understand Japanese, and I had little practical knowledge of machine learning.My Chinese fellow, Li Tong, has been my "Japanese interpreter" and "technical encyclopedia" in the first semester.Ri Ryoukan, my deskmate, has always been caring about the research of laboratory members including me.We have discussed over many fields, and some discussions were very inspiring.Matiss Ritkers, a researcher, has been warm-hearted and always ready for help.And other members have been sharing their ideas, comments, suggestions, both in academics and in life experience, which are invaluable to me.I want to thank AIST members, especially my direct leader, Hiroshi Noji.He is firm and persevering, leading me to a series of interesting topics, including the KBQA.I have encountered failure in some difficult topics, but he has never been frustrated and has been always encouraging me to look forward.Other AIST members are also kind and helpful.I really enjoy the open and friendly culture of AIST.Besides, I want to thank Lyu Zhaoyang, my former roommate and best friend.We worked together on the BWE mapping, and we have discussed many latest topics in the machine learning field some of which are very inspiring.I also would like to thank other friends who concern my research and gives useful advice .Thanks to all the members I have mentioned or not mentioned.
Author: Yifan Deng
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3217166324