Abstract
In the biomedical domain, queries often seek a comprehensive list of entities—such as genes, proteins, or diseases—covering various aspects of a topic. Traditional retrieval models, which assume document relevance is independent, often produce redundant results. This paper introduces RelNov, a relevance-novelty combined model based on an undirected graphical framework. RelNov incorporates two components: an aspect-term relevance model and an aspect-term novelty model. By leveraging Wikipedia for aspect detection, the model achieves significant performance gains, including a 16.4% improvement in Aspect MAP on the TREC 2007 Genomics track. Our results demonstrate that integrating aspects and document novelty effectively promotes ranking diversity and improves overall retrieval relevance.
Introduction
Biologists frequently search for information that requires a diverse set of answers. For instance, a query about proteins associated with a specific disease should ideally return a list of different relevant proteins rather than multiple documents discussing the same one. However, standard Information Retrieval (IR) systems typically rank documents based on individual relevance scores, ignoring the redundancy between results. This lack of "ranking diversity" is particularly problematic in genomics, where complex terminology and synonyms can mask repetitive information.
Methodology: The RelNov Model
To address the redundancy problem, we propose the RelNov model. Unlike traditional models, RelNov treats the ranking process as a search for documents that are both relevant to the query and novel relative to previously seen results.
Graphical Framework
The model is built on an undirected graphical framework (specifically a Markov Random Field). It defines potential functions for aspects, terms, topic relevance, and document novelty to calculate the probability of a document being both useful and unique.
Aspect Detection via Wikipedia
A key innovation is the use of Wikipedia as an external knowledge base to identify "aspects"—semantic concepts like specific biological entities. By "Wikifying" documents, the system maps various phrases (e.g., "SLE" and "Systemic lupus erythematosus") to a single canonical concept. This allows the model to recognize redundancy even when different terminology is used.
Two-Tiered Modeling
RelNov evaluates documents using both literal terms and underlying aspects. This dual approach ensures that the system captures both the specific nuances of the text and the high-level information "nuggets" provided by the document.
Experimental Evaluation
The model was tested using the TREC 2006 and 2007 Genomics track collections, which include over 160,000 full-text biomedical documents. Performance was measured using Aspect MAP (Mean Average Precision), a metric designed to reward the discovery of diverse information aspects.
Key Results
- Performance Gains: RelNov achieved a 16.4% improvement in Aspect MAP over the top-performing system in the TREC 2007 Genomics track and a 9.8% improvement in passage-level retrieval.
- Robustness: The model improved performance across various baselines, including standard Okapi BM25 and more advanced systems like NLMinter.
- Knowledge Base Efficacy: Wikipedia was found to be more effective for aspect detection than specialized biomedical ontologies like UMLS, largely due to its superior context for disambiguating acronyms.
Conclusion
The RelNov model demonstrates that explicitly modeling document novelty and semantic aspects significantly enhances the quality of genomics search results. By moving beyond keyword matching to concept-based diversity, the system provides biologists with more comprehensive and less redundant information. Future work will focus on improving the efficiency of the novelty calculations and exploring deeper semantic relationships between biological entities.
Summary updated