BioVerbNet: a large semantic-syntactic classification of verbs in biomedicine

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Biomedical Informatics

Background:

• Verbal reasoning is a challenge for NLP systems, despite advances in representation learning.
• Structured verb knowledge resources improve NLP task performance but are costly to create for specialized domains like biomedicine.
• BioVerbNet addresses this by combining neural classification with expert annotation for biomedical verbs.

Purpose of the Study:

• To create BioVerbNet, a novel semantic-syntactic classification of biomedical verbs.
• To provide a resource that aids NLP systems in understanding verb meaning and behavior in biomedical texts.
• To demonstrate the utility of BioVerbNet in enhancing NLP model performance for biomedical applications.

Main Methods:

• Developed BioVerbNet by combining a neural classification method with expert annotation.
• Classified 693 biomedical verbs into 22 top-level and 117 fine-grained semantic-syntactic classes.
• Created knowledge-aware word embeddings using a retrofitting method with BioVerbNet classifications.

Main Results:

• BioVerbNet comprises 693 verbs with detailed semantic roles and VerbNet-style syntactic frames.
• Knowledge-aware embeddings derived from BioVerbNet significantly outperformed a non-specialized baseline.
• The resource demonstrated improved performance in both document- and sentence-level biomedical text classification tasks.

Conclusions:

• BioVerbNet is the first large-scale, annotated semantic-syntactic classification of biomedical verbs.
• The resource captures domain-specific verb properties and differences from general language.
• Leveraging BioVerbNet enhances NLP tasks in biomedicine, suggesting its potential for future applications.