Natural Language Processing and XML Retrieval

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 165–166. Natural Language Processing and XML Retrieval Alan Woodley Faculty of Information Technology Queensland University of Technology a.woodley@qut.edu.au Marcus Hassler Department of Informatics Universitdat Klagenfurt Marcus.Hassler@uni-klu.ac.at Xavier Tannier Dept. Networks, Information, Multi-media Ecole des Mines Saint-Etienne tannier@emse.fr Shlomo Geva Faculty of Information Technology Queensland University of Technology s.geva@qut.edu.au Abstract XML information retrieval (XML-IR) sys- tems respond to user queries with results more specific than documents. XML-IR queries contain both content and struc- tural requirements traditionally expressed in a formal language. However, an intu- itive alternative is natural language queries (NLQs). Here, we discuss three ap- proaches for handling NLQs in an XML- IR system that are comparable to, and even outperform formal language queries. 1 Introduction Information retrieval (IR) systems respond to user queries with a ranked list of relevant documents, even though only parts of the documents are rel- evant. In contrast, XML-IR systems are able to exploit the separation of structure and content in XML documents by returning relevant portions of documents. To interact with XML-IR sys- tems users must specify both their content and structural requirements in structured queries. Cur- rently, formal languages are used to specify struc- tured queries, however, they have proven problem- atic since they are too difficult to use and are too tightly bound to the collection. A promising alternative to formal queries are natural language queries (NLQs). Here, we present justifications for NLQs in XML-IR, and describe three approaches that translate NLQs to an existing formal language (NEXI). When used in with an XML-IR system the approaches per- form strongly, at times outperforming a baseline consisting of manually constructed NEXI expres- sions. These results show that NLQs are poten- tially a viable alternative to XML-IR systems. 2 Motivation There are two major problems with formal query languages for XML-IR that could be rectified with NLQs. First, expressing a structural information need in a formal language is too difficult for many users. O’Keefe and Trotman (2004) investigated five structured query languages and concluded that all of them were too complicated to use. In prac- tice, 63% of the expert-built queries queries in the 2003 INEX campaign had major semantic or syn- tactic errors, requiring up to 12 rounds of correc- tions. In contrast, users should be able to express their need in NLQs intuitively. Second, formal query languages require an inti- mate knowledge of a document’s structure. So, in order to retrieve information from abstracts, sec- tions or bibliographic items, users need to know their corresponding tags. While this information is contained in the DTD/ Schema, it may not be publicly available, and is too much information to remember (INEX, for instance has 192 nodes). The problem extrapolates in a heterogenous col- lection since a single retrieval unit could be ex- pressed in multiple tags. In contrast, since struc- tures in NLQs are formulated at the conceptional level users do not have to know their actual tag names. 3 The Approaches Here, we …

Content truncated to 500 words