We describe our information retrieval system which allows a convivial access to databases. It is based on a multi-expert architecture using a database management system and a blackboard to control the progressive analysis of the user's sentence. We have integrated a numerical method based on fuzzy rules to deal with uncertainty. This method is used to optimize the analysis process and the cooperation between the different experts. On the other hand, the goal of our recent research is to ease or eliminate the knowledge-acquisition bottleneck for expert system creation and to make a connectionist model behave as much as possible like an expert system. We describe our experience using neural networks to represent the knowledge bases of the different experts (i.e., lexical entries expert, homographs expert, template expert, grammatical word expert and words expert).