Author: Min Jiang, MS (2017)
Primary advisor: Hua Xu, PhD
Committee members: Trevor Cohen, MBChB, MD, PhD, Cui Tao, PhD
PhD thesis, The University of Texas School of Health Information Sciences at Houston.
Syntactic parsing is one of the fundamental tasks of Natural Language Processing (NLP). However, few studies have explored syntactic parsing in the medical domain. This dissertation systematically investigated different methods to improve the performance of syntactic parsing of clinical text, including (1) Constructing two clinical treebanks of discharge summaries and progress notes by developing annotation guidelines that handle missing elements in clinical sentences; (2) Retraining four state-of-the-art parsers, including the Stanford parser, Berkeley parser, Charniak parser, and Bikel parser, using clinical treebanks, and comparing their performance to identify better parsing approaches; and (3) Developing new methods to reduce syntactic ambiguity caused by Prepositional Phrase (PP) attachment and coordination using semantic information.
Our evaluation showed that clinical treebanks greatly improved the performance of existing parsers. The Berkeley parser achieved the best F-1 score of 86.39% on the MiPACQ treebank. For PP attachment, our proposed methods improved the accuracies of PP attachment by 2.35% on the MiPACQ corpus and 1.77% on the I2b2 corpus. For coordination, our method achieved a precision of 94.9% and a precision of 90.3% for the MiPACQ and i2b2 corpus, respectively. To further demonstrate the effectiveness of the improved parsing approaches, we applied outputs of our parsers to two external NLP tasks: semantic role labeling and temporal relation extraction. The experimental results showed that performance of both tasks’ was improved by using the parse tree information from our optimized parsers, with an improvement of 3.26% in F-measure for semantic role labelling and an improvement of 1.5% in F-measure for temporal relation extraction.