Joint document-level information extraction

Constructing knowledge graphs from unstructured text is an important task that is relevant to many domains. Recently neural models have been used to great effect in order to solve many information extraction tasks. However, there are still many challenges that need to be solved before our models can achieve a level of natural language understanding that could be comparable to human. In order to accomplish that, we need to create models that can be optimized to jointly perform various IE tasks on large volumes of text, while properly utilizing all of the available information. In addition, as our models get more complex it is important to focus on producing explainable predictions, so that the reasoning behind a specific extracted fact can be understood by human users of the model. As a step towards solving these challenges, we introduce two new document-level IE models. The first model is trained to jointly perform identification, coreference, and classification of entities and events within a document by utilizing aggregated contextual information from each relevant mention. The second model builds on the first to extract relations with evidence between the entities in a document. We evaluate our models on the ACE-05+ and DocRed datasets respectively, and find improvements over the current SOTA in terms of F-score on entity, event, and evidence extraction.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Kriman, Samuel

Contributors dc:contributor

• Ji, Heng

• Information Extraction
• Entity
• Relation
• Event
• Evidence
• Joint Model

Statement dc:rights

• Copyright 2021 Samuel Kriman

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110851

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110851