gene pathway text mining and visualization - Artificial Intelligence ...

534 MEDICAL INFORMATICS3.2 GeneScene ParserThis case study discusses the Genescene parser, which extractsrelations between noun phrases and is tuned for biomedical text.3.2.1 Extracting Semantic ElementsThe parser begins by formatting, tokenizing, and tagging the PUBMEDabstract with part-of-speech and noun phrase tags. The abstracts are preparedby removing phrases referring to publisher and copyright information. Thenthe sentence splitter is run followed by the AZ Noun Phraser (Tolle andChen, 2000) to extract noun phrases. Verbs and adverbs are tagged with theirpart-of-speech (POS) based on a rule set and lexical lookup in the UMLSSpecialist Lexicon. Closed class words such as prepositions, negation,conjunctions, and punctuation are also tagged. Nouns and noun phrases arechecked for nominalizations. When a nominalization is discovered, e.g.,“activation,” then both the infinitive and the original nominalization areretained. Nominalizations can be replaced by the infinitive to facilitate textmining and visualization.3.2.2 Extracting Structural ElementsRelations have a syntactic basis: they are built around basic sentencestructures and prepositions. Prepositions were chosen because they form aclosed class and can help capture the structure of a sentence. The closedclasses’ membership does not change and allows us to build very specificbut semantically generic relation templates. In addition, prepositions oftenhead phrases (Pullum and Huddleston, 2002) and indicate different types ofrelations, such as time or spatial relations (Manning and Schütze, 2001).Although prepositional attachment ambiguity may become a problem, webelieve that researchers in biomedicine use a common writing style and sothe attachment structures will not vary much for a specific structure.This case study describes relations built around three prepositions, “by,”“of,” and “in,” which occur frequently in text and lead to interesting, diversebiomedical relations. “By” is often used to head complements in passivesentences, for example in “Mdm2 is not increased by the Ala20 mutation.”“Of” is one of the most highly grammaticised prepositions (Pullum andHuddleston, 2002) and is often used as a complement, such as for examplein “the inhibition of the activity of the tumor suppressor protein p53.” “In” isusually an indication of location. It forms interesting relations whencombined with verbs, for example in “Bcl-2 expression is inhibited inprecancerous B cells.”