Porting a lexicalized-grammar parser to the biomedical domain.

Area of Science:

• Computational linguistics
• Bioinformatics
• Natural Language Processing

Background:

• Statistical parsers are crucial for biomedical text mining.
• Existing parsers, like those trained on Penn Treebank, are optimized for general text (e.g., newspaper text).
• Adapting these parsers to the specialized biomedical domain often requires extensive annotated data.

Purpose of the Study:

• Introduce a linguistically motivated statistical parser to the biomedical text mining community.
• Propose an efficient method for adapting a parser to the biomedical domain with minimal data annotation.
• Evaluate the parser's performance on biomedical text.

Main Methods:

• Utilized Combinatory Categorial Grammar (ccg), a lexicalized grammar formalism.
• Trained the parser at three levels of representation: part-of-speech (POS) tags, ccg lexical categories, and ccg derivations.
• Retrained the POS tagger on biomedical data and incorporated intermediate lexical category annotations.

Main Results:

• Retraining the POS tagger alone significantly improved parsing performance on biomedical text.
• Further improvements in parsing accuracy were achieved by using annotated data at the intermediate lexical category level.
• The adapted parser achieved accuracies comparable to those reported for newspaper text and surpassed previous benchmarks for the evaluated biomedical resource.

Conclusions:

• Adapting general-domain statistical parsers, particularly those based on lexicalized grammars, to the biomedical domain is more feasible than previously assumed.
• Limited data annotation, focusing on POS tags and lexical categories, can yield high parsing performance.
• This approach offers an efficient way to enhance biomedical text mining capabilities.