Information Retrieval with Language Knowledge

The introduction of Swedish made it possible for Lexware to be tested for the first time in CLEF. Lexware is a natural language system applied in an information retrieval task and not an information retrieval systems using NLP techniques, therefore it is interesting to compare its results with other less odd IR systems. We experience that separate evaluation of document description and query building would provide yet better testing for our system. 1 A Natural Language System for Swedish Lexware is a natural language system applied in an information retrieval task and not an information retrieval systems using NLP techniques, like e.g. NLIR [4]. It can be considered odd also among natural language processing systems, if the latter are assumed to focus on syntactic analysis (c.f. [3]). Text-analysis is shallow and it is not demanding in terms of computing power and storage [1]. The strength of the system is its rich lexicon and the possibility to expand the lexicon with external items without negative impact on access time [2]. The vocabulary of about 80 000 lexical items is richly interconnected by relations of form and content: derivational origin, synonymy, components for complex items, hyponymy. Content words are categorized into about 100 content categories. There are also supplementary word lists which include about 50 000 non-appellatives like names of people, places, organizations, etc. plus basic glossaries of English, French, German, and Latin. 400 word formation rules cope with inflection, compounding and derivation, 500 general phrase rules plus 700 collocation patterns are used to disambiguate and to determine modifier–head roles. 2 Lexware in Another Information Retrieval Task Lexware has been extensively tested in another information retrieval task. The library of the Swedish parliament – Riksdagsbiblioteket, designed and conducted evaluation of software that could supplement or even substitute manual indexing of the documents of the parliament. The task is to select proper keywords among descriptors in a thesaurus specially created for this kind of documents.