Improving named entity recognition accuracy of gene and protein in biomedical text

• Main Author: Tohidi, Hossein
• Format: Thesis
• Language: English; English
• Published: 2011
• Online Access: http://psasir.upm.edu.my/id/eprint/27703/1/FSKTM%202011%2026R.pdf; http://psasir.upm.edu.my/id/eprint/27703/
• Institution: Universiti Putra Malaysia

The plethora of biomedical material on the WWW is one of the factors that have sustained interest in automatic methods for extracting information from biomedical document, which can help biologists in their research. To extract useful knowledge from the biomedical literature, we must be able to recognize names of biomedical entities, such as genes, proteins, cells, and diseases which are called Named Entity. The task of recognizing entity-denoting expressions, or named entities (NE), in natural language documents is called Named Entity Recognition (NER). Among the biomedical types such as gene, protein, virus, cells, and etc, the most important biomedical types for recognition are gene and protein, which is the scope of this research. The most important reason why most researchers focus on the gene and protein named entities is due to the complexity nature of such types. This complexity includes the issues of character-level variation, word-level variation, and word order variation in biomedical text literature. Typically there are four approaches for Named Entity Recognition, namely: Dictionary-Based, Rule-Based, Statistical and Machine Learning, and Hybrid approaches. In this study, to handle the above issues in recognizing gene and protein names, a statistical similarity measurement as a pattern matching function is proposed. Our approach is based on an assumption that a named entity occurs among a noun group which is extracted using Brill Part of Speech tagger. The strength of our proposed approach for recognizing biomedical named entity is based on a Statistical Character-Based Syntax Similarity (SCSS) algorithm which measured similarity between all extracted candidates and the well-known biomedical named entities from a corpus. For this study, we have used the GENIA V3.0 corpus, which is the largest annotated corpus in the molecular and biology domain. The proposed approach is evaluated based on two measures: recall and precision which are used to calculate a balanced F-test. We have compared our pattern matching function with the other methods and result is satisfied as precision is 98.5% and recall is 96.4%, while the F-test is 97.5 for both gene and protein names recognizing and precision is 99.3% and recall is 99.1%, while the F-test is 99.1 for protein names recognizing.