Identifying discourse connectives in biomedical text.

Area of Science:

• Computational linguistics
• Bioinformatics
• Natural Language Processing

Background:

• Discourse connectives link sentences, indicating relationships crucial for text comprehension.
• Automatic recognition of these connectives has broad applications in Natural Language Processing (NLP).
• Identifying discourse connectives in specialized domains like biomedical literature presents unique challenges.

Purpose of the Study:

• To develop and evaluate supervised machine-learning classifiers for automatic discourse connective identification in biomedical articles.
• To assess the performance of a classifier trained on a general domain corpus (Penn Discourse Tree Bank - PDTB) when applied to biomedical text.
• To investigate the potential benefits of domain-specific adaptation for this task.

Main Methods:

• Development of supervised machine-learning classifiers using Conditional Random Fields (CRFs).
• Training a primary classifier on the open-domain Penn Discourse Tree Bank (PDTB) corpus (1 million tokens).
• Cross-validation of the PDTB-trained classifier on a custom-annotated biomedical corpus (approx. 100K tokens).

Main Results:

• The classifier trained on the general PDTB corpus achieved a 0.55 F1-score on biomedical text.
• Cross-validation within the biomedical corpus yielded a significantly higher F1-score of 0.69.
• This improvement was observed despite the smaller size of the biomedical training data.

Conclusions:

• A general-domain discourse connective classifier performs moderately on biomedical text.
• Domain-specific adaptation, even with limited data, substantially enhances identification performance.
• The findings suggest the presence of domain-specific linguistic features that warrant further investigation for improved NLP models.