DOI
https://doi.org/10.25772/8JJX-VX13
Author ORCID Identifier
0000-0002-4950-6876
Defense Date
2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
First Advisor
Dr.Bridget T. McInnes
Abstract
Relation Extraction (RE) is a task of Natural Language Processing (NLP) to detect and classify the relations between two entities. Relation extraction in the biomedical and scientific literature domain is challenging as text can contain multiple pairs of entities in the same instance. During the course of this research, we developed an RE framework (RelEx), which consists of five main RE paradigms: rule-based, machine learning-based, Convolutional Neural Network (CNN)-based, Bidirectional Encoder Representations from Transformers (BERT)-based, and Graph Convolutional Networks (GCNs)-based approaches. RelEx's rule-based approach uses co-location information of the entities to determine whether a relation exists between a selected entity and the other entities. RelEx's machine learning-based approach consists of traditional feature representations into traditional machine learning algorithms. RelEx's CNN-based approach consists of three CNN architectures: Segment-CNN, single-label Sentence-CNN, and multi-label Sentence-CNN. RelEx's BERT-based approach utilizes BERT's contextualized word embeddings into a feed-forward neural network. Finally, RelEx's GCN-based approach consists of two GCN-based architectures: GCN-Vanilla, GCN-BERT. We evaluated variations of these approaches in two different domains across four distinct relation types.
Overall our findings showed that the rule-based approach is applicable for data with fewer instances in the training data. In contrast, the CNN-based, BERT-based, and GCN-based approaches perform better with labeled data with many training instances. These approaches automatically identify patterns in the data efficiently, whereas rule-based approaches require expert knowledge to generate rules. The CNN-based, BERT-based approaches capture the local contextual information within a sentence or document by embedding both semantic and syntactic information in a learned representation. However, their ability to capture the long-range dependency global information in a text is limited. GCN-based approaches capture the global association information by performing convolution operations on neighbor nodes in a graph and incorporating information from neighbors. Combining GCN with BERT integrates the local contextual and global association information of the words and generates better representations for the words.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
5-10-2022