Comparison of Crosslingual Similarity Measures for Multilingual Documents Clustering
Résumé
This paper compares the performance of one thesaurus-based approach
against three lexicon-based techniques to the measurement of the crosslingual
similarity of domain-specific texts. These methods are applied to an unstructured
manually annotated corpus of texts in three languages: French, English
and German. We investigate the correlation between these measures and
human judgement as well as their ability to detect subtle (in the same topic) and
broader (in related topics) differences in comparability. Additional experiments
aim to determine the extent to which terminology helps improving measures of
similarity in a specialised domain. Results suggest that injecting domain-specific
knowledge, when available, is a good alternative to more shallow techniques.