Date of Award

6-19-2019

Document Type

Thesis (Undergraduate)

Department

Department of Computer Science

First Advisor

Saeed Hassanpour

Abstract

Objective: Currently, a major limitation for natural language processing (NLP) analyses in clinical applications is that a concept can be referenced in various forms across different texts. This paper introduces Multi-Ontology Refined Embeddings (MORE), a novel hybrid framework for incorporating domain knowledge from various ontologies into a distributional semantic model, learned from a corpus of clinical text. This approach generates word embeddings that are more accurate and extensible for computing the semantic similarity of biomedical concepts than previous methods. Materials and Methods: We use the RadCore and MIMIC-III free-text datasets for the corpus-based component of MORE. For the ontology-based component, we use the Medical Subject Headings (MeSH) ontology and two state-of-the-art ontology-based similarity measures. In our approach, we propose a new learning objective, modified from the Sigmoid cross-entropy objective function, to incorporate domain knowledge into the process for generating the word embeddings. Results and Discussion: We evaluate the quality of the generated word embeddings using an established dataset of semantic similarities among biomedical concept pairs. We show that the similarity scores produced by MORE have the highest average correlation (60.2%), with the similarity scores being established by multiple physicians and domain experts, which is 4.3% higher than that of the word2vec baseline model and 6.8% higher than that of the best ontology-based similarity measure. Conclusion: MORE incorporates knowledge from biomedical ontologies into an existing distributional semantics model (i.e. word2vec), improving both the flexibility and accuracy of the learned word embeddings. We demonstrate that MORE outperforms the baseline word2vec model, as well as the individual UMLS-Similarity ontology similarity measures.

Comments

Originally posted in the Dartmouth College Computer Science Technical Report Series, number TR2019-873.

Share

COinS