Most difficulties arising in the development of speech understanding systems come from the great uncertainty inherent in the acoustic signal. The inaccuracy of acoustic-phonetic decoders makes it necessary to use higher level information: lexical, syntactic and semantic. We present here a possible solution for integrating some of this information in the particular case of an administrative database questioning system.First, we try to sketch how contextual knowledge can be extracted from local or global representations of an utterance, such as partial recognition or the dialogue history in the case of man-machine interaction.Then, we focus on the production of lexical hypotheses in top-down analysis, within the framework of a system architecture based on the cooperation of independent modules.We apply the Dempster-Shafer formalism in order to combine information in the lexicon, using a frequency distribution as the basis for evidence evaluation. This mathematical rule has suitable properties in the case of an oral dialogue system, as it preserves module autonomy and allows backtracking at any time during the recognition process.This approach seems promising in as much as it helps speech recognition converge faster to a satisfactory solution, thanks to linguistic information.