A Machine Learning Approach for Phenotype Name Recognition

Proceedings of COLING 2012: Technical Papers, pages 1425–1440, COLING 2012, Mumbai, December 2012. A machine learning approach for phenotype name recognition Mar yam Khordad1 Robert E. Mercer1 Peter Rogan1,2 (1)Department of Computer Science (2)Department of Biochemistry The University of Western Ontario, London, ON, Canada mkhordad@csd.uwo.ca, mercer@csd.uwo.ca, progan@uwo.ca ABSTRACT Extracting biomedical named entities is one of the major challenges in automatic processing of biomedical literature. This paper proposes a machine learning approach for finding phenotype names in text. Features are included in a machine learning infrastructure to implement the rules found in our previously developed rule-based system. The system also uses two available resources: MetaMap and HPO. As we are not aware of any available corpus for phenotype names, a corpus has been constructed. Since manual tagging of the corpus was not possible for us, we started tagging only HPO phenotypes in the corpus and then using a semi-supervised learning method, the tagging process improved. The evaluation results (F-Score 92.25) suggest that the system achieved good performance and it outperforms the rule-based system. KEYWORDS: Phenotype, Named Entity Recognition, MetaMap, Human Phenotype Ontology. 1425 1 Introduction A large amount of biomedical knowledge is available in the biomedical literature. So, automatic processing of biomedical literature to capture and formalize this embedded information is very demanding. Current Natural Language Processing (NLP) systems try to extract from the biomedical literature different types of knowledge such as, protein–protein interactions (Leroy et al., 2003) (He and DiMarco, 2005) (Fundel et al., 2007) (kiong Ng and Wong, 1999) (Yu et al., 2005) (Bui et al., 2011), new hypotheses(Swanson, 1986) (Hristovski et al., 2005) (Hristovski et al., 2006), relations between drugs, genes, and cells (Rindflesch et al., 2000) (Friedman et al., 2001) (Tanabe et al., 1999), relations between genes and diseases (Rindflesch et al., 2003) (Coulet et al., 2010), protein structure (Humphreys et al., 2000) (Gaizauskas et al., 2003), and protein function (Andrade and Valencia, 1998) (Valencia, 2005). Fundamental to each of these applications is the Named Entity Recognition and Classification (NERC) task. Research over the past years has shown that recognizing the names of biomedical objects is not a simple task. Factors that preclude a straightforward procedure include the existence of an ever-increasing large (millions) set of entity names, a penchant for the use of abbreviations, the use of synonyms, and the fact that some biological entities have complex names consisting of many words, like “increased erythrocyte adenosine deaminase activity”, and lacking agreement on the name boundaries, even among biologists (Leser and Hakenberg, 2005). Named Entity Recognition and Classification in the biomedical domain has been extensively studied. As a consequence, many methods have been proposed. Generic methods, like MetaMap (Aronson, 2001) and mgrep (Dai et al., 2008) recognize and classify many kinds of entities in text. Some methods, however, are specialized to recognize particular types of entities like gene or protein names (Krauthammer et al., 2000) (Gaizauskas et al., 2003), diseases and drugs (Rindflesch et al., 2000) (Xu et al., 2008) (Segura-Bedmar et al., 2008), mutations (Horn et …

Content truncated to 500 words