Natural language processing for biology

A large part of the information required for biology research can only be found in free-text form, as in MEDLINE abstracts, or in comment elds of relevant reports, as in GenBank feature table annotations. Such information is important for many t ypes of analysis, such as classiication of proteins into functional groups, extraction of protein-protein interaction facts, discovery of new functional relationships, maintaining information of material and methods, increasing the precision and relevance of hits returned by information retrieval systems, and so on. However, information in free-text form or in comment elds is very diicult for use by automated system. For example, annotation of biological function of diierent proteins is a time-consuming process currently performed by h uman experts because genome analysis tools encounter great diiculty in performing this task. The ability to extract information directly from MEDLINE abstracts and other sources can directly help in such a task. Five papers were accepted under peer-review in this session. Previous work in automated understanding of biomedical papers tended to concentrate on analytical tasks such as identifying protein names. We are delighted that all ve accepted papers considered substantially less constrained problems that involved nding relationships and contexts. The paper by Baclawski et. al. describes a diagrammatic knowledge representation technique called keynets. The rich o n tology of the Uniied Medical Language System was used to construct and index keynets. Fully using the domain-independent and domain-speciic knowledge, keynets parses texts and resolve references to construct new relationships between entities. The paper by Humphreys et. al. describes two information extraction applications in bioinformatics based on templates. The rst application is EMPathIE, which is able to extract details of enzyme and metabolic pathways from journal articles. The second application is PASTA, which is able to extract information on the roles of amino acids and active sites in protein molecules from journal articles. They clariied how important template matching is in this eld, and applied the technique to terminology recognition. The paper by Rinddesch el. al. describes EDGAR, a natural language processing system that extracts relationships between cancer-related drugs and genes from biomedical literature. EDGAR draws on a combination of technologies: a stochastic part of speech tagger, an underspeciied syntactic parser, a rule-based system, as well as semantics information from the Uniied Medical Language System. The metathesaurus and the lexicon in the knowledge base are used to identify the structure of noun phrases in MEDLINE …