Project ANR Emotirob aims at detecting emotions from an original point of view: realizing an emotional companion robot for weakened children. In our approach, linguistic detection and prosodie are combined. Our experiments show that human beings can estimate the emotional value of an utterance from its propositional content in a reliable way. So we have implemented a first model of linguistic detection, based on the principle that emotions can be compound: lexical words have an emotional value while predicates can modify emotional values of their arguments. This paper presents a short description of the logical understanding system, the outputs of which are used for the final emotional value calculus. Then, the creation of a lexical emotional reference standard is presented with an ontology of emotional predicate classes for children, aged between 5 and 7.