NLM DIR Seminar Schedule

Abstract:

With the growing availability of full-text articles, integrating abstracts and full texts of documents into a unified representation becomes essential for performing comprehensive biomedical literature search. However, previous studies have shown that naïvely merging abstracts with full texts of articles does not consistently yield better performance.

In this work we study how to combine abstracts with available article full texts to improve the overall retrieval performance. For this purpose, we create an evaluation dataset and define a simulated environment that mimics the search environment of interest. The evaluation dataset consists of queries sampled from PubMed logs along with a subset of retrieved and clicked documents satisfying the requirement of the simulated search environment for each query. To remove PubMed users’ bias, we only consider documents for which none of the query tokens appear in the title. Another known source of biased clicks are clicks on the top rank. We removed these as they might simply represent a user’s urge to click on something indiscriminately. Summing up, to create our user click dataset, we collected only retrieved documents for which none of the query tokens appeared in the title and all of them appeared in the abstract, and we ignored clicks on the top tank. Using the dataset, we demonstrate that different sections of a full text document are of different value in deciding relevance and propose a method to combine information coming from various parts of a full text document by converting the information to log odds scores which can be treated uniformly.

Our experimental results suggest that although all query tokens appear in the abstract, incorporation of the body text of PMC improves the PubMed search over that of abstract only. Moreover, if one or more tokens are missing from the abstract our proposed technique can lead to better retrieval results by scoring the body text of the document. Considering that about 15% of PubMed queries return no results due to one or more query tokens not being present in the abstract or title, the proposed approach can provide useful retrieval results for a significant number of those failed queries by supplementing abstracts with full text information. Our approach may also be useful for queries which retrieve few documents; for these queries abstract-only retrieval can be augmented by full text search results.