| EIT
worked on monolingual experiments in German and English
as baselines for the cross-language experiments. Two sets
of relevance assessments were produced by separate groups
of medical experts (at ZInfo and at CMU). In order to check
the influence of the relevance assessments on the retrieval
results the monolingual German evaluation runs were repeated
for ZInfo and CMU relevance assessments and checked against
each other. For most of the cross-language evaluation runs
German queries were used to retrieve English documents.
The results of these experimenst were compared to the monolingual
runs but also between different approaches. All CLIR experiments
were performed with the ZInfo relevance assessments. Experiments
included: CLIR via Vocabulary Overlap; CLIR via Machine
Translation of the Queries; CLIR via Semantic Codes; CLIR
via a Bilingual Similarity Thesaurus; CLIR with English
queries and German documents; Different weighting schemes.
XRCE investigated whether an optimal combination of different
resources for query translation could lead to improved results
in the context of cross-language information retrieval.
Data for estimating mixture weights was automatically derived
by building pseudo-translation pairs in which the source
words are extracted from the set of queries used in the
training phase, and their corresponding target words are
taken to be those occurring in all the documents judged
to be relevant to the query. Experiments in this context
showed that the combination thus obtained was always better
than a simple merge of the resources, but we nevertheless
failed to report significant improvements on the retrieval
results.
|
