Guided text summarization with limited supervision

Mao, Yuning

Permalink

https://hdl.handle.net/2142/115460 Copy

Description

Title

Guided text summarization with limited supervision

Author(s)

Mao, Yuning

Issue Date

2022-04-15

Director of Research (if dissertation) or Advisor (if thesis)

Han, Jiawei

Doctoral Committee Chair(s)

Han, Jiawei

Committee Member(s)

• Ji, Heng
• Abdelzaher, Tarek
• Yih, Wen-tau

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• text summarization
• natural language processing
• text generation

Abstract

In the information age, people are surrounded by a massive amount of texts in daily life. It has become increasingly difficult to acquire desirable, salient, and non-redundant information from unstructured text data. Given the abundance of textual information, automatic text summarization rises as a critical means to provide people with salient and adequate information in a timely and manageable fashion. Text summarization has seen great progress in recent years, yet many issues remain or emerge, such as the lack of high-quality training data and model hallucination, due to the data-driven nature of mainstream summarization methods without necessary guidance. In light of the aforementioned background and challenges, my research aims to improve text summarization with more explicit guidance. In contrast to most prior studies on guided text summarization, I propose to conduct guided summarization with limited supervision – I design guided text summarization methods that either require limited labeled data or are totally unsupervised, which is critical for the development of text summarization as human annotations are typically expensive and time-consuming to obtain while the backbone of modern text summarizers (i.e., data-hungry neural models) require increasingly more training data. I have worked on guided text summarization from two aspects – the summarizer’s aspect and the user’s aspect, with various subjects as guidance (e.g., background documents and keyphrases). Summarizer-centric guidance directs summarizer with weak supervision, while user-centric guidance grants more control of the summarizer to a user with limited but valuable supervision from the user. From the summarizer’s perspective, mainstream summarization methods are typically data-driven without explicit guidance, which might be effective with abundant labeled data but fall short when the available training data is limited. I thus design methods that direct summarizer with weak supervision and improve summary quality without human annotation. The guidance received by a summarizer can be from internal sources, e.g., by leveraging content redundancy [1] and document structure [2]. The guidance can also be external, e.g., by making use of surrounding documents, which contain useful background information [3] or even provide direct supervision signals [4]. From the user’s perspective, mainstream summarizers are not customizable and generally offer the same output to different users. However, different users have various preferences and information needs in nature. I thus design methods that grant more control of the summarizer to the user so that a user can directly guide the summarization process and obtain tailored summaries suited to their own preferences, such as covering specific entities or events and generating summaries with different granularity [5, 6, 7].

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Yuning Mao

Owning Collections