1	Biomedical corpora for natural language processing: What’s a nice Librarian like you doing in a project like this? K. Bretonnel Cohen Lynne Fox (lynne.fox@uchsc.edu) Philip V. Ogren Lawrence Hunter University of Colorado Health Sciences Center
2	Today’s talk The Center for Computational Pharmacology and their work Define key terms Illustrate corpus creation using our project as an example Summarize findings of research on other biomedical corpora Describe my role (10% appointment) Encourage you to get involved at your institution
3
4	Center for Computational Pharmacology http://compbio.uchsc.edu/ In the School of Medicine, Department of Pharmacology Purpose: Creating novel algorithms and knowledge-based tools for the analysis and interpretation of high-throughput molecular biology data Test and prove theories about biomedical linguistics using computer programs. (Can we create a natural language processing program that recognizes named entities [locations, proteins, etc.], or identifies a physiological process, or finds the co-occurrence of two proteins?)
5	Why biologists (and librarians) care about NLP High-throughput data is generated (dna microarray chips) Literature searches needs to be more automated (A+B, A+C, A+D . . . ) Annotation of data improves access via computer algorithms (More refined searching of relationships) Knowledgebase creation provides access to new knowledge and discoveries
6	Definitions
7	Corpus (plural: corpora) A body of data in digital form that is used to build and test natural language processing systems, that may include: Textual data (naturally occurring in any source – transcripts of spoken communication, textbooks, articles, etc.) Annotation (“mark-up” can be automated or manual) The ability to compare the results from computer programs to “right” and “wrong” answers in the corpus
8	Types of annotations in corpora: Sentence segmentation: <sentence>Here, we show that Bifocal (Bif), a putative cytoskeletal regulator, is a component of the Msn pathway for regulating R cell growth targeting. </sentence> <sentence>bif displays strong genetic interaction with msn. </sentence> Tokenization: Divides the text into smallest units (usually words), removing punctuation. Challenge: What should be done with punctuation that has linguistic meaning? Negative charge (Cl-) Absence of symptom (-fever) Knocked-out gene (Ski-/-) Gene name (IL-2 – mediated) Plus, “syntactic” uses (insulin-dependent) Part-of-speech tagging: Met28/NOUN binds/VERB to/TO DNA/NOUN./.
9
10	Our Corpus Experts manually annotate Gene References Into Function sentences from Entrez Gene NLP researchers use the corpus to compare the results of tasks executed by their computer programs to right and wrong answers and “train” the programs to improve their results The goal is to automatically process the texts from NCBI GeneRIFs (National Center for Biotechnology Information Entrez Gene References Into Function), into a formal knowledge-base of the current understanding of protein transportation.
11	GRIFs A sentence about a specific gene and the function of the gene, usually taken from a sentence in a PubMed abstract APOE does not play an important role in susceptibility to Parkinson's disease, but may play a role in dementia associated with Parkinson's disease (PMID: 11863377, 2002)
12
13	How can we learn from extant corpora? Our research What we wanted to know: What corpora exist? What design features did they include? How did that impact use? How did we find out? Citation index search Manually examined papers Contacted corpus designers
14	Which corpora did we study?
15	Our Findings Format of corpora determined use (portability of markup method) Manual curation matters (the gold standard and training set) Size doesn’t matter (within limits) Named entity markup matters Annotation style matters (linguistic and structural)
16	My Role I’m familiar with user needs and I know why it’s tough to find information Mediate in the process of creating data and using data Consult on PubMed searching, MEDLINE structure, and controlled vocabularies (ontologies) Draw attention to ways commercial databases mine data Assist with Gene Reference into Function (GRIF) corpus annotation project Idea generation Copyright “buzz-kill”
17	What’s Next For Me? Continue GRIF annotation project, corpus creation & design research Grant application to bring my participation to 8 hours/week, 20% time Consult on projects, search strategy, MEDLINE KEEP LEARNING!
18	What can you do? Network, network, network Connect with folks on your campus or at a university in your area Look for departments of linguistics, biocomputing, bioinformatics Look for units “hidden” within research divisions Electronic medical records managers Attend AMIA, Association for Computational Linguistics, Intelligent Systems for Molecular Biology, Pacific Symposium for Biocomputing Suggest grant ideas, especially for clinical data mining Ask for funding for participation
19	Further Reading Fuller SS, Revere D, Bugni PF, Martin GM. A knowledgebase system to enhance scientific discovery: Telemakus. Biomedical Digital Libraries. 21;1(1):2, September 2004. Rzhetsky A, Iossifov I, Koike T, et al. GeneWays: a system for extracting, analyzing, visualizing, and integrating molecular pathway data. Journal Of Biomedical Informatics. 37 (1): 43-53, February 2004. Srinivasan P, Hristovski D. Distilling conceptual connections from MeSH co-occurrences. Medinfo. 11(Pt 2):808-12, 2004. Watson LA, Hersh WR Information retrieval: A health and biomedical perspective. Springer; 2nd edition, 2003, ISBN: 0387955224.
20	Summary Librarians and computational linguists have common goals, can benefit from mutual cooperation, and librarians have unique skills that can contribute to the study of natural language processing in biomedicine. Go forth and find NLPeople!
21	Acknowledgements NIH grant R01-LM008111 to Larry Hunter Lynette Hirschman, Alex Morgan, Kristofer Franzén (Comments on study) Christian Blaschke, Mark Craven, Lorraine Tanabe, Kristofer Franzén (Corpus designers)