Files in this item

FilesDescriptionFormat

application/pdf

application/pdfHUANG-DISSERTATION-2020.pdf (3MB)
(no description provided)PDF

Description

Title:Cold-start universal information extraction
Author(s):Huang, Lifu
Director of Research:Ji, Heng
Doctoral Committee Chair(s):Ji, Heng
Doctoral Committee Member(s):Han, Jiawei; Zhai, Chengxiang; Cho, Kyunghyun
Department / Program:Computer Science Department
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):Information Extraction
Event Extraction
Zero-Shot Learning
Common Semantic Space
Abstract:Who? What? When? Where? Why? are fundamental questions asked when gathering knowledge about and understanding a concept, topic, or event. The answers to these questions underpin the key information conveyed in the overwhelming majority, if not all, of language-based communication. At the core of my research in Information Extraction (IE) is the desire to endow machines with the ability to automatically extract, assess, and understand text in order to answer these fundamental questions. IE has been serving as one of the most important components for many downstream natural language processing (NLP) tasks, such as knowledge base completion, machine reading comprehension, machine translation and so on. The proliferation of the Web also intensifies the need of dealing with enormous amount of unstructured data from various sources, such as languages, genres and domains. When building an IE system, the conventional pipeline is to (1) ask expert linguists to rigorously define a target set of knowledge types we wish to extract by examining a large data set, (2) collect resources and human annotations for each type, and (3) design features and train machine learning models to extract knowledge elements. In practice, this process is very expensive as each step involves extensive human effort which is not always available, for example, to specify the knowledge types for a particular scenario, both consumers and expert linguists need to examine a lot of data from that domain and write detailed annotation guidelines for each type. Hand-crafted schemas, which define the types and complex templates of the expected knowledge elements, often provide low coverage and fail to generalize to new domains. For example, none of the traditional event extraction programs, such as ACE (Automatic Content Extraction) and TAC-KBP, include "donation'' and "evacuation'' in their schemas in spite of their potential relevance to natural disaster management users. Additionally, these approaches are highly dependent on linguistic resources and human labeled data tuned to pre-defined types, so they suffer from poor scalability and portability when moving to a new language, domain, or genre. The focus of this thesis is to develop effective theories and algorithms for IE which not only yield satisfactory quality by incorporating prior linguistic and semantic knowledge, but also greater portability and scalability by moving away from the high cost and narrow focus of large-scale manual annotation. This thesis opens up a new research direction called Cold-Start Universal Information Extraction, where the full extraction and analysis starts from scratch and requires little or no prior manual annotation or pre-defined type schema. In addition to this new research paradigm, we also contribute effective algorithms and models towards resolving the following three challenges: How can machines extract knowledge without any pre-defined types or any human annotated data? We develop an effective bottom-up and unsupervised Liberal Information Extraction framework based on the hypothesis that the meaning and underlying knowledge conveyed by linguistic expressions is usually embodied by their usages in language, which makes it possible to automatically induces a type schema based on rich contextual representations of all knowledge elements by combining their symbolic and distributional semantics using unsupervised hierarchical clustering. How can machines benefit from available resources, e.g., large-scale ontologies or existing human annotations? My research has shown that pre-defined types can also be encoded by rich contextual or structured representations, through which knowledge elements can be mapped to their appropriate types. Therefore, we design a weakly supervised Zero-shot Learning and a Semi-Supervised Vector Quantized Variational Auto-Encoder approach that frames IE as a grounding problem instead of classification, where knowledge elements are grounded into any types from an extensible and large-scale target ontology or induced from the corpora, with available annotations for a few types. How can IE approaches be extent to low-resource languages without any extra human effort? There are more than 6000 living languages in the real world while public gold-standard annotations are only available for a few dominant languages. To facilitate the adaptation of these IE frameworks to other languages, especially low resource languages, a Multilingual Common Semantic Space is further proposed to serve as a bridge for transferring existing resources and annotated data from dominant languages to more than 300 low resource languages. Moreover, a Multi-Level Adversarial Transfer framework is also designed to learn language-agnostic features across various languages.
Issue Date:2020-11-04
Type:Thesis
URI:http://hdl.handle.net/2142/109351
Rights Information:Copyright 2020 Lifu Huang
Date Available in IDEALS:2021-03-05
Date Deposited:2020-12


This item appears in the following Collection(s)

Item Statistics