Word Sense Disambiguation in Biomedical Applications: A Machine Learning Approach

Word Sense Disambiguation in Biomedical Applications: A Machine Learning Approach

Torsten Schiemann (Humboldt-Universität zu Berlin, Germany), Ulf Leser (Humboldt-Universität zu Berlin, Germany) and Jörg Hakenberg (Arizona State University, USA)
DOI: 10.4018/978-1-60566-274-9.ch008
OnDemand PDF Download:
$37.50

Abstract

Ambiguity is a common phenomenon in text, especially in the biomedical domain. For instance, it is frequently the case that a gene, a protein encoded by the gene, and a disease associated with the protein share the same name. Resolving this problem, that is, assigning to an ambiguous word in a given context its correct meaning is called word sense disambiguation (WSD). It is a pre-requisite for associating entities in text to external identifiers and thus to put the results from text mining into a larger knowledge framework. In this chapter, we introduce the WSD problem and sketch general approaches for solving it. The authors then describe in detail the results of a study in WSD using classification. For each sense of an ambiguous term, they collected a large number of exemplary texts automatically and used them to train an SVM-based classifier. This method reaches a median success rate of 97%. The authors also provide an analysis of potential sources and methods to obtain training examples, which proved to be the most difficult part of this study.
Chapter Preview
Top

Introduction

Ambiguity, i.e., words with multiple possible meanings, is a common phenomenon in natural languages (Manning & Schütze 1999). Which of the different meanings of a word is actually meant in a concrete text depends on the context the word appears in and cannot be judged based only on the appearance of the word itself. For instance, the terms ‘sin’ and ‘soul’ both are common English words – but they are also names of proteins. If a person only sees one of these two words on a piece of paper, he cannot decide which of the two meanings (or senses) the paper tries to convey. However, given a phrase such as “Salvation from sins”, humans immediately recognize the correct sense of the ambiguous word.

From a linguistics point of view, the term ambiguity in itself has different senses. The most common form is homonymity, that is, words that have multiple, possibly unrelated meanings. ‘Sin’ and ‘soul’ both are homonyms. However, there are also more complex forms of ambiguity, such as polysemy, which describes cases where a word has different yet closely related senses. Examples in the life sciences are identical names for a gene, the protein it encodes, and the mRNA in which it is transcribed.

Word sense disambiguation (WSD) is the problem of assigning to an ambiguous term in a given context its correct sense (Ide & Veronis 1998). Although humans usually have no problem in disambiguating terms, the problem is challenging for computers due to the necessity to capture the context of a word in a text, which, in general, not only encompasses the preceding and following words, but also background knowledge on the different senses of the word that might apply in the topic of the text. Obviously, WSD is the more difficult the more related the different senses are. It should be rather simple to distinguish the senses mtg gene and particular tube announcement of the term ‘mind the gap’, but much more difficult to tell the senses gene and drug from the term ‘oxytocin’.

However, solving the WSD-problem is a pre-requisite for high-quality named entity recognition and, in particular, entity normalization (Leser & Hakenberg 2005, Cohen 2005). As an example, consider the AliBaba system described in (Plake et al. 2005). AliBaba extracts occurrences of cells, diseases, drugs, genes/proteins, organisms, and tissues from PubMed abstracts using class-specific dictionaries. In this context, the WSD problem appears with terms that are contained in multiple dictionaries, and whose occurrences therefore indicate entities from different classes. Note that such terms are more frequent than one might expect. In a small study, we found 175 terms indicating both a species and a protein, 67 terms indicating a drug and a protein, and 123 terms indicating a cell and a tissue. Furthermore, names of biological entities often are homonym with common English word, like the above mentioned ‘sin’ and ‘soul’, but also terms such as ‘black’, ‘hedgehog’, or ‘mind the gap’, which are all names of genes in Drosophila (see http://flybase.bio.indiana.edu/). Some terms even appear in more than two classes, such as ‘dare’, which may indicate a protein, an organism, or the English verb. (Figure 1)

Figure 1.

Screenshot of the Ali Baba main window after searching PubMed for the term ‘soul’. Colored boxes represent biological entities. One can see that various Meanings of ‘soul’ are intermixed in the display – mentioning of the immortal soul by the Greek poet Homer, results from a large international study called ‘Heart and Soul’, and facts describing the protein ‘soul’.

Complete Chapter List

Search this Book:
Reset