An experimental comparison of explicit semantic analysis implementations for cross-language retrieval

Explicit Semantic Analysis (ESA) has been recently proposed as an approach to computing semantic relatedness between words (and indirectly also between texts) and has thus a natural application in information retrieval, showing the potential to alleviate the vocabulary mismatch problem inherent in standard Bag-of-Word models. The ESA model has been also recently extended to cross-lingual retrieval settings, which can be considered as an extreme case of the vocabulary mismatch problem. The ESA approach actually represents a class of approaches and allows for various instantiations. As our first contribution, we generalize ESA in order to clearly show the degrees of freedom it provides. Second, we propose some variants of ESA along different dimensions, testing their impact on performance on a cross-lingual mate retrieval task on two datasets (JRC-ACQUIS and Multext). Our results are interesting as a systematic investigation has been missing so far and the variations between different basic design choices are significant. We also show that the settings adopted in the original ESA implementation are reasonably good, which to our knowledge has not been demonstrated so far, but can still be significantly improved by tuning the right parameters (yielding a relative improvement on a cross-lingual mate retrieval task of between 62% (Multext) and 237% (JRC-ACQUIS) with respect to the original ESA model).