Comparing different knowledge sources for the automatic summarization of biomedical literature

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Automatic summarization of biomedical literature often requires external domain knowledge for semantic representation.
• The impact of different knowledge sources on summary quality remains an area for investigation.

Purpose of the Study:

• To investigate how the choice of knowledge source affects the quality of automatically generated biomedical summaries.
• To evaluate the performance of different ontology and vocabulary combinations within the Unified Medical Language System (UMLS).

Main Methods:

• Developed a method to represent documents as semantic graphs using domain concepts and relations.
• Constructed graphs using various combinations of UMLS ontologies (GO, SNOMED-CT, HUGO) and vocabularies.
• Employed different relationship types (co-occurrence, semantic) from UMLS Metathesaurus and Semantic Network.
• Inputted generated graphs into a summarization system to extract relevant sentences.

Main Results:

• The selection of knowledge sources significantly influences the quality of automatic biomedical summaries.
• Summaries of gene-related literature were notably improved when using Gene Ontology (GO), SNOMED-CT, and HUGO.
• This approach yielded superior results compared to utilizing all available UMLS vocabularies.

Conclusions:

• The choice of knowledge source is critical for effective biomedical text summarization.
• Ontology and vocabulary selection should align with the specific characteristics (coverage, specificity, relations) of the documents being summarized.
• Tailoring knowledge sources enhances the accuracy and relevance of automated summaries.