Author ORCID Identifier

0000-0001-8064-521X

Defense Date

2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Bridget T. McInnes

Abstract

Temporal reasoning is the ability to extract and assimilate temporal information to reconstruct a series of events such that they can be reasoned over to answer questions involving time. Temporal reasoning in the clinical domain is challenging due to specialized medical terms and nomenclature, shorthand notation, fragmented text, a variety of writing styles used by different medical units, redundancy of information that has to be reconciled, and an increased number of temporal references as compared to general domain texts. Work in the area of clinical temporal reasoning has progressed, but the current state-of-the-art still has a ways to go before practical application in the clinical setting will be possible. Much of the current work in this field is focused on direct and explicit temporal expressions and identifying temporal relations. However, there is little work focused on relative temporal expressions, which can be difficult to normalize, but are vital to ordering events on a timeline. This work introduces a new temporal expression recognition and normalization tool, Chrono, that normalizes temporal expressions into both SCATE and TimeML schemes. Chrono advances clinical timeline extraction as it is capable of identifying more vague and relative temporal expressions than the current state-of-the-art and utilizes contextualized word embeddings from fine-tuned BERT models to disambiguate temporal types, which achieves state-of-the-art performance on relative temporal expressions. In addition, this work shows that fine-tuning BERT models on temporal tasks modifies the contextualized embeddings so that they achieve improved performance in classical SVM and CNN classifiers. Finally, this works provides a new tool for linking temporal expressions to events or other entities by introducing a novel method to identify which tokens an entire temporal expression is paying the most attention to by summarizing the attention weight matrices output by BERT models.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

5-10-2022

Share

COinS