=Paper=
{{Paper
|id=Vol-1170/CLEF2004wn-other-Thurmair2004
|storemode=property
|title=Issues When Building CLIR Applications: Outline of Invited Talk at CLEF 2004 Workshop
|pdfUrl=https://ceur-ws.org/Vol-1170/CLEF2004wn-other-Thurmair2004.pdf
|volume=Vol-1170
|dblpUrl=https://dblp.org/rec/conf/clef/Thurmair04
}}
==Issues When Building CLIR Applications: Outline of Invited Talk at CLEF 2004 Workshop==
<pdf width="1500px">https://ceur-ws.org/Vol-1170/CLEF2004wn-other-Thurmair2004.pdf</pdf>
<pre>
                     Issues When Building CLIR Applications

                 Outline of Invited Talk at CLEF 2004 Workshop

                                    G. Thurmair, linguatec

The talk will discuss some design issues which influence the quality of crosslingual retrieval.
1. On the Analysis Side, index creation is a key point. The index should be based on base
    forms, not on text forms or stems; therefore state-of-the-art stemmers are not the best tools
    to use. Also, there should be one index per language, to avoid clashes of terms occurring
    in several languages; this implies to run a language-detection component before indexing.
2. On the Query side, the quality of search crucially depends on the resource used for
    translation. While several techniques have been proposed, preference should be given to a
    multilingual conceptual network, to be used for a term-to-term translation. The closer this
    network matches the domain to be searched the better the search quality will be.
3. Once the search result is available, the questions of ranking of multilingual documents, as
    well as retranslation into the query language need to be addressed. Technology
    components to support these tasks must be provided.
4. The CLIR application requires a new kind of linguistic resource which merges
    requirements of terminology, formal NLP type dictionaries, conceptual hierarchies, and
    multilinguality like in Machine Translation. Such a multifunctional resource must be
    designed and maintained, and the paper ends with a list of requirements for such a
    resource.
As a result, a good CLIR system shows significant differences as compared to a conventional
IR implementation.

</pre>