Unsupervised grammar induction of clinical report sublanguage

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Medical Informatics

Background:

• Clinical reports utilize specialized sublanguages, varying across genres and institutions.
• Supervised training of clinical text parsers is challenging due to the need for extensive, domain-specific annotations.
• Adapting parsers to diverse clinical language conventions requires significant manual effort.

Purpose of the Study:

• To develop an unsupervised method for automatically inducing grammars and parsers for clinical sublanguages.
• To overcome the limitations of supervised methods requiring annotated clinical text.
• To enable robust parsing of clinical sentences without manual annotation.

Main Methods:

• An unsupervised grammar induction method is presented, utilizing semantic classes and part-of-speech tags.
• Grammar induction is achieved by minimizing the combined encoding cost of the grammar and sentence derivations.
• Expectation-maximization algorithm is used to learn grammar production probabilities from unannotated corpora.

Main Results:

• The induced grammar successfully parses novel clinical sentences.
• An F-measure of 60.5% for parse-bracketing was achieved on unannotated discharge summary sentences (max length 10).
• The method allows for inducing a spectrum of grammars, from specific to general, with optimal performance at an intermediate setting.

Conclusions:

• Unsupervised grammar induction is effective for parsing clinical sublanguages.
• The developed method reduces the need for manual annotation in clinical natural language processing.
• This approach offers a flexible and efficient way to build parsers for diverse clinical text genres.