Computer Laboratory

Technical reports

Semi-supervised learning for biomedical information extraction

Andreas Vlachos

November 2010, 113 pages

This technical report is based on a dissertation submitted December 2009 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Peterhouse College.

Abstract

This thesis explores the application of semi-supervised learning to biomedical information extraction. The latter has emerged in recent years as a challenging application domain for natural language processing techniques. The challenge stems partly from the lack of appropriate resources that can be used as labeled training data. Therefore, we choose to focus on semi-supervised learning techniques which enable us to take advantage of human supervision combined with unlabeled data.

We begin with a short introduction to biomedical information extraction and semi-supervised learning in Chapter 1. Chapter 2 focuses on the task of biomedical named entity recognition. Using raw abstracts and a dictionary of gene names we develop two systems for this task. Furthermore, we discuss annotation issues and demonstrate how the performance can be improved using user feedback in realistic conditions. In Chapter 3 we develop two biomedical event extraction systems: a rule-based one and a machine learning based one. The former needs only an annotated dictionary and syntactic parsing as input, while the latter requires partial event annotation additionally. Both systems achieve performances comparable to systems utilizing fully annotated training data. Chapter 4 discusses the task of lexical-semantic clustering using Dirichlet process mixture models. We review the unsupervised learning method used, which allows the number of clusters discovered to be determined by the data. Furthermore, we introduce a new clustering evaluation measure that addresses some shortcomings of the existing measures. Chapter 5 introduces a method of guiding the clustering solution using pairwise links between instances. Furthermore, we present a method of selecting these pairwise links actively in order to decrease the amount of supervision required. Finally, Chapter 6 assesses the contributions of this thesis and highlights directions for future work.

Full text

PDF (1.0 MB)

BibTeX record

@TechReport{UCAM-CL-TR-791,
  author =	 {Vlachos, Andreas},
  title = 	 {{Semi-supervised learning for biomedical information
         	   extraction}},
  year = 	 2010,
  month = 	 nov,
  url = 	 {http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-791.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  number = 	 {UCAM-CL-TR-791}
}