Information Retrieval On Empty Fields

Proceedings of NAACL HLT 2007, pages 89–96, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Information Retrieval On Empty Fields Victor Lavrenko, Xing Yi and James Allan Center for Intelligent Information Retrieval Department of Computer Science University of Massachusetts Amherst, MA 01003-4610, USA {lavrenko,yixing,allan}@cs.umass.edu Abstract We explore the problem of retrieving semi-structured documents from a real- world collection using a structured query. We formally develop Structured Rele- vance Models (SRM), a retrieval model that is based on the idea that plausible values for a given field could be inferred from the context provided by the other fields in the record. We then carry out a set of experiments using a snapshot of the National Science Digital Library (NSDL) repository, and queries that only mention fields missing from the test data. For such queries, typical field matching would re- trieve no documents at all. In contrast, the SRM approach achieves a mean average precision of over twenty percent. 1 Introduction This study investigates information retrieval on semi-structured information, where documents con- sist of several textual fields that can be queried in- dependently. If documents contained subject and author fields, for example, we would expect to see queries looking for documents about theory of rela- tivity by the author Einstein. This setting suggests exploring the issue of inex- act match—is special theory of relativity relevant?— that has been explored elsewhere (Cohen, 2000). Our interest is in an extreme case of that problem, where the content of a field is not corrupted or in- correct, but is actually absent. We wish to find rele- vant information in response to a query such as the one above even if a relevant document is completely missing the subject and author fields. Our research is motivated by the challenges we encountered in working with the National Science Digital Library (NSDL) collection.1 Each item in the collection is a scientific resource, such as a re- search paper, an educational video, or perhaps an entire website. In addition to its main content, each resource is annotated with metadata, which provides information such as the author or creator of the re- source, its subject area, format (text/image/video) and intended audience – in all over 90 distinct fields (though some are very related). Making use of such extensive metadata in a digital library paves the way for constructing highly-focused models of the user’s information need. These models have the potential to dramatically improve the user experience in tar- geted applications, such as the NSDL portals. To illustrate this point, suppose that we are running an educational portal targeted at elementary school teachers, and some user requests teaching aids for an introductory class on gravity. An intelligent search system would be able to translate the request into a structured query that might look something like: subject=’gravity’ AND audience=’grades 1-4’ AND format=’image,video’ AND rights=’free-for- academic-use’. Such a query can be efficiently an- swered by a relational database system. Unfortunately, using a relational engine to query a semi-structured collection similar to NSDL …

Content truncated to 500 words