Generation of a Large Gene/Protein Lexicon by Morphological Pattern Analysis

Lorrie Tanabe

11:00 am
Tuesday September 30
5th Floor Conference Room

The identification of gene/protein names in biomedical text is an important 
named entity recognition (NER) problem. It is a necessary first step towards 
semantic analysis and text mining, and can also aid in information retrieval.  
Dictionaries can improve the performance of gene/protein NER, however, currently 
available lexicons are incomplete. In previous work we have processed MEDLINE 
documents to obtain a collection of over 2 million names of which we estimate 
two thirds are valid gene/proteins.  

This talk focuses on an approach to purify this large set of names based on the 
generation of classes of names with common morphological features.  Within each 
class, inductive logic programming (ILP) is applied to learn which features are 
most predictive of gene/protein names.  193 classes and ILP theories were 
generated and applied to the 2 million names defining a subset of 1.2 million 
names.  A simple false positive filter eliminated 8% of these, leaving 1.14 
million entries in the final lexicon, which is composed of 82% (+/- 3%) complete 
and accurate gene/protein names, 12% names related to genes/proteins (generic 
names, partial names, etc.), and 6% names unrelated to genes/proteins.