CCE Theses and Dissertations

Date of Award

2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

College of Computing and Engineering

Advisor

Sumitra Mukherjee

Committee Member

Michael J. Laszlo

Committee Member

Francisco J. Mitropoulos

Abstract

Relation extraction and classification represents a fundamental and challenging aspect of Natural Language Processing (NLP) research which depends on other tasks such as entity detection and word sense disambiguation. Traditional relation extraction methods based on pattern-matching using regular expressions grammars and lexico-syntactic pattern rules suffer from several drawbacks including the labor involved in handcrafting and maintaining large number of rules that are difficult to reuse. Current research has focused on using Neural Networks to help improve the accuracy of relation extraction tasks using a specific type of Recurrent Neural Network (RNN). A promising approach for relation classification uses an RNN that incorporates an ontology-based concept embedding layer in addition to word embeddings. This dissertation presents several improvements to this approach by addressing its main limitations. First, several different types of semantic relationships between concepts are incorporated into the model; prior work has only considered is-a hierarchical relationships. Secondly, a significantly larger vocabulary of concepts is used. Thirdly, an improved method for concept matching was devised. The results of adding these improvements to two state-of-the-art baseline models demonstrated an improvement to accuracy when evaluated on benchmark data used in prior studies.

Share

COinS