Up to now, the core technology of the commercial bilingual concordancer TransSearch mainly relies on sentence level-alignment. The main objective of the TS3 project conducted at the RALI is to automatically identify the translations of a user query in pairs of sentences retrieved from a parallel corpus. In my talk, I will discuss and evaluate the embedding of statistical word-level alignment. Since word alignment is a notoriously difficult problem, I will also discuss two novel issues that are essential to the success of TransSearch: the identification of erroneous alignments and the grouping of translation variants.