Work

Information Extraction Beyond Sentence Boundary

Year: 2024
Type: dissertation
Abstract: Information Extraction (IE) is the task of extracting structured information from unstructured texts.The collected structured information can be transformed into a Knowledge Base (KB), serving as a valuable assistant for human decision-making.Relation Extraction (RE), an important subfield of IE aiming at extracting relation triples in the form of (subject, relation, object), is closely related to the automatic construction of graph-shaped KB.While classical RE is a sentence-level task, recent studies have pointed out that sentence-level RE is impractical, as relations can hold document-wise, i.e., beyond sentence boundaries.The task, Document-level Relation Extraction (DocRE), is thus proposed to encourage extracting both intra-and inter-sentence relation triples.However, due to the complexity of DocRE, human-annotated supervisory signals of high quality are limited and difficult to expand.This results in two challenges of DocRE regarding human annotations.Firstly, the limited human annotations are not fully utilized to train a better DocRE model.Specifically, evidence annotations -a set of sentences necessary to identify the relation between an entity pair -are provided alongside relation annotations but are not used to train a DocRE model.Secondly, there is no methodology that enables efficient construction of a DocRE dataset in a new language.While there is a demand for automatically populating KB for each language, datasets supporting the training of DocRE models are limited to only two or three languages.This study explores ways to address the aforementioned challenges.The core idea is to better utilize existing language resources with human annotations to help model construction and dataset construction.
Authors: MA, Youmi, Ma, Youmi
Language: English
Open Access status: green
OpenAlex ID: W7019724782
Cites: 117
Cited by:
Related to: 9