Cross-language retrieval with the Twenty-One system.pdf

Cross Language Retrieval with the Twenty-One systemWessel KraaijTNO-TPDP.O. Box 155, 2600 AD DelftThe Netherlandskraaij@tpd.tno.nl Djoerd HiemstraUniversity of Twente/CTITP.O. Box 217, 7500 AE, EnschedeThe Netherlandshiemstra@cs.utwente.nlAbstractThe EU project Twenty-One will support cross lan-guage queries in a multilingual document base. Aprototype version of the Twenty-One system has beensubjected to the Cross Language track tests in orderto set baseline performances. The runs were based onquery translation using dictionaries and corpus baseddisambiguation methods.1 Introduction1.1 Twenty-One projectTwenty-One is a 2 MECU project with 11 partners1funded by the EU Telematics program, sector In-formation Engineering. The project subtitle is De-velopment of a Multimedia Information Transactionand Dissemination Tool. Twenty-One started early1996 and is currently in its building phase.The Twenty-One database consists of documents indierent languages, initially Dutch, English, Frenchand German but extensions to other European lan-guages are envisaged. The TREC Cross Language(CLIR) track task ts our needs to evaluate the sys-tem on the aspect of cross language retrieval perfor-mance.1.2 TREC6Although the development of the full scale Twenty-One system just started in the summer of 1997,Twenty-One accepted the challenge to participate inthe cross language track of TREC6.Whether we would complete the task was a com-plete question, because at that moment (May 1997),1Project partners are: Getronics software, TNO-TPD,DFKI, Rank Xerox Grenoble, University of Twente, Universityof Tubingen, MOOI foundation, Environ, Climate Alliance,VODO and Friends of the Earth the TNOmono-lingual vector space search engine wasstill under development and untested, The deliveryof a fast workstation was also delayed, and moreover,the consortium was still negotiating with two pub-lishers to acquire bilingual dictionaries. But nallyall hard-, soft- and lingware became available justin time to