Date Available

12-7-2016

Year of Publication

2016

Degree Name

Master of Science (MS)

Document Type

Master's Thesis

College

Engineering

Department/School/Program

Computer Science

First Advisor

Dr. Ramakanth Kavuluru

Abstract

Addressing ambiguity issues is an important step in natural language processing (NLP) pipelines designed for information extraction and knowledge discovery. This problem is also common in biomedicine where NLP applications have become indispensable to exploit latent information from biomedical literature and clinical narratives from electronic medical records. In this thesis, we propose an ensemble model that employs recent advances in neural word embeddings along with knowledge based approaches to build a biomedical word sense disambiguation (WSD) system. Specifically, our system identities the correct sense from a given set of candidates for each ambiguous word when presented in its context (surrounding words). We use the MSH WSD dataset, a well known public dataset consisting of 203 ambiguous terms each with nearly 200 different instances and an average of two candidate senses represented by concepts in the unified medical language system (UMLS). We employ a popular biomedical concept, Our linear time (in terms of number of senses and context length) unsupervised and knowledge based approach improves over the state-of-the-art methods by over 3% in accuracy. A more expensive approach based on the k-nearest neighbor framework improves over prior best results by 5% in accuracy. Our results demonstrate that recent advances in neural dense word vector representations offer excellent potential for solving biomedical WSD.

Digital Object Identifier (DOI)

https://doi.org/10.13023/ETD.2016.490

Share

COinS