Learning to rank diversified results for biomedical information retrieval from multiple features

Area of Science:

• Biomedical Informatics
• Information Retrieval

Background:

• Traditional information retrieval (IR) often overlooks document relationships, leading to redundant results.
• Promoting diversity in IR is crucial for satisfying varied user intents, especially in specialized domains like biology.
• Biologists require diverse search results that capture different facets of their queries.

Purpose of the Study:

• To develop and evaluate a novel learning-to-rank (LTR) framework for diversity-enhanced biomedical information retrieval.
• To improve the novelty and reduce redundancy of search results in the biomedical domain.

Main Methods:

• A combined LTR framework integrating a general ranking model (gLTR) and a diversity-biased model was developed.
• Diversity-indicating features were extracted based on passage topics (using Wikipedia) and the gLTR model's output.
• Final rankings were achieved by combining the general and diversity-biased models.

Main Results:

• The general LTR (gLTR) model demonstrated significant improvements in Aspect Mean Average Precision (Aspect MAP) over baseline methods (BM25, DirKL).
• The combined LTR method further outperformed gLTR, showing notable percentage improvements in Aspect MAP on benchmark collections.
• The proposed LTR approach effectively enhanced diversity in biomedical search results.

Conclusions:

• The developed learning-to-rank method is an effective strategy for improving biomedical information retrieval.
• Incorporating diversity-biased features significantly enhances the diversity of ranking results in biomedical IR.