N-Grams For Translation And Retrieval In Cl-Sdr

We report on a first attempt to perform cross-language spoken document retrieval. Without prior monolingual speech retrieval experience we applied the same general approach we use for bilingual retrieval that is typified by the use of overlapping character n-grams for tokenization and a statistical language model of retrieval. An innovative approach was adopted for coping, with out-of-vocabulary words and misspelled or mistranscribed words: direct translation of individual n-grams was the sole mechanism to translate source language queries into target language terms. Though this approach shows promise, especially for non-speech retrieval, our performance appears to lag that of other teams participating in this novel evaluation.