Domain adaption of parsing for operative notes

Area of Science:

• Computational linguistics
• Natural Language Processing (NLP)
• Clinical informatics

Background:

• Syntactic parsing of clinical text is crucial for clinical NLP applications.
• Existing general English parsers often require adaptation for specialized clinical text.
• Operative reports present unique linguistic challenges for standard NLP tools.

Purpose of the Study:

• To adapt a general English syntactic parser for clinical operative reports.
• To enhance parser performance through lexicon augmentation and statistical adjustments.
• To modify grammar rules based on the linguistic characteristics of operative reports.

Main Methods:

• Expanded the Stanford unlexicalized probabilistic context-free grammar (PCFG) parser lexicon with the SPECIALIST lexicon.
• Incorporated statistics from operative notes tagged by GENIA and MedPost POS taggers.
• Modified parser grammar rules and verb entries based on manual review of clinical text.

Main Results:

• The adapted parser achieved an improved F-score of 89.90%, a 2.26% increase from the baseline.
• Lexicon augmentation and corpus statistics yielded the most significant performance gains.
• The approach demonstrated generalizability, improving performance on the GENIA corpus by 3.81%.

Conclusions:

• Adapting unlexicalized PCFG parsers using clinical text statistics and modified lexicons/grammars enhances performance on specialized clinical text.
• This method offers a viable strategy for improving NLP tools in the medical domain.
• The findings support the customization of NLP models for specific clinical document types.