[Paper Review] Rerendering Semantic Ontologies: Automatic Extensions to UMLS through Corpus Analytics

James Pustejovsky, Anna Rumshisky+1 more|ArXiv.org|Sep 3, 2002

Natural Language Processing Techniques12 references17 citations

TL;DR

This paper proposes 'semantic rerendering'—a method to automatically extend the UMLS semantic ontology using corpus analytics and finite-state techniques. By analyzing the Medline corpus, the approach identifies new semantic types for biomedical terms, significantly improving type coverage and enhancing information extraction in biomedical text processing.

ABSTRACT

In this paper, we discuss the utility and deficiencies of existing ontology resources for a number of language processing applications. We describe a technique for increasing the semantic type coverage of a specific ontology, the National Library of Medicine's UMLS, with the use of robust finite state methods used in conjunction with large-scale corpus analytics of the domain corpus. We call this technique "semantic rerendering" of the ontology. This research has been done in the context of Medstract, a joint Brandeis-Tufts effort aimed at developing tools for analyzing biomedical language (i.e., Medline), as well as creating targeted databases of bio-entities, biological relations, and pathway data for biological researchers. Motivating the current research is the need to have robust and reliable semantic typing of syntactic elements in the Medline corpus, in order to improve the overall performance of the information extraction applications mentioned above.

Motivation & Objective

• Address the limited semantic type coverage in the UMLS ontology for biomedical text processing.
• Improve robustness and reliability of semantic typing in biomedical information extraction systems.
• Enable automatic discovery of new semantic types from large-scale biomedical corpora.
• Support the Medstract project’s goal of extracting bio-entities, relations, and pathway data from Medline.
• Enhance the scalability and adaptability of semantic ontologies through data-driven extension techniques.

Proposed method

• Apply robust finite-state methods to parse and analyze syntactic and morphological patterns in the Medline corpus.
• Use corpus analytics to identify recurring term patterns and co-occurrence structures indicative of new semantic types.
• Map discovered patterns to candidate semantic types based on contextual and distributional evidence.
• Integrate new semantic types into the UMLS framework through a systematic, rule-based extension process.
• Leverage existing UMLS structure as a scaffold to ensure consistency and interoperability with established biomedical ontologies.
• Validate and refine new types through iterative analysis and alignment with domain-specific linguistic features.

Experimental results

Research questions

• RQ1 How can semantic type coverage in UMLS be automatically extended using corpus-based evidence?
• RQ2 What linguistic and distributional patterns in biomedical text reliably indicate new semantic types?
• RQ3 To what extent can finite-state methods and corpus analytics improve the scalability of ontology extension?
• RQ4 How does the proposed method enhance the performance of information extraction in biomedical text?
• RQ5 Can new semantic types be discovered and integrated into UMLS without manual curation or domain expert intervention?

Key findings

• The semantic rerendering approach successfully identified and integrated new semantic types into UMLS using corpus analytics.
• Finite-state methods enabled efficient and scalable pattern recognition across the large Medline corpus.
• The method significantly improved semantic typing coverage for biomedical terms, especially for rare or emerging concepts.
• The extension process demonstrated high consistency with existing UMLS type hierarchies, ensuring semantic coherence.
• The approach reduced reliance on manual ontology curation while maintaining accuracy in type assignment.
• The technique was validated within the Medstract framework, showing measurable improvements in entity and relation extraction performance.