Part-of-Speech Tagger for Biomedical Text

TODO

• convert the upenn biocorpus data into the "pos" format. (done)
  • Spaces within a token were replaced with underscores.
  • training data (the first 90% of each (cyp450 and oncology) )
  • test data (the rest)
• train the tagger (done)
  • use PTB, GENIA and the above upennbio corpus as the training data.
• evaluate the tagger (done)
  • tagging accuracy: 96.94% on PTB, 98.26% on GENIA, 97.78% on UpennBio.
  • compare with the GENIA tagger and the TnT tagger:
    • three articles from the latest issue of Nucleic Acid Research (2005 March)
    • tagging results
• build the server

Acronym Generator

The acronym generator takes a term (typically multi-word term) and generates possible acronyms for the term.
ex.) human growth hormone -> HGF

TODO

• collect papers on acronym detection (done)
• make the training data (done)
  • extract definition-acronym pairs from medline abstracts
  • annotate each pair with positional information
  • 1,901 pairs
• build the generator (done)
  • MEMM-based algorithm
• evaluate (done)
  • 5-fold cross-validation
  • Coverage: 55.2% (Top1), 75.4% (Top 5), 82.2% (Top 10)
• write a paper (done) "A Machine Learning Approach to Acronym Generation"(pdf)

Chunker for Biomedical Text

TODO

• convert the GENIA and upenn-biocorpus data into the "chunk" format.
• train the chuker