Using concept relations to improve ranking in information retrieval

Abstract

Despite improved search engine technology, most searches return numerous documents not directly related to the query. This problem is mitigated if relevant documents appear high on a ranked list of search results. We propose that some queries and the underlying information needs can be modeled as relationships between concepts (relations), and we match relations in queries to relations in documents to try to improve ranking of search results. We investigate four techniques to identify two relationships important in medicine, causes and treats, to improve the ranking of medical text documents relevant to clinical questions about causation and treatment. Preliminary results suggest that identifying relation instances can improve the ranking of search results.

Figure 1

Proposed query types. Example queries are shown with relational representations. The letters X and Y represent variables, meaning the slot can be filled by any concept or relationship.

Figure 2

Calculation of similarity score for document d relative to query q based on terms t in q. coord: fraction of terms in q that appear in d tf (term frequency): factor for frequency of t in d idf (inverse term freq.): factor for number of documents containing t (increases when t is in fewer documents) lengthNorm: normalizes for the length of the document

Figure 3

An example of a regular expression and two phrases that it matched. The word “metronidazole” from the query relation treats(“metronidazole”, *) was used to instantiate the variable $intervention.

Figure 4

Query histograms showing the effects of using expansion, regular expressions, and proximity matching on each query. Bars show change in average precision compared to the baseline run. Queries 1–6 have full treats relations, queries 7–12 have full causes relations, queries 13–18 have partial treats relations and queries 19–24 have partial causes relations.