International audience Focusing on the human-computer interaction for natural language processing, we present a formal notion of interpretation, as a frame for the semantic analysis of texts. Moving away from classical automated processing, we define the role of the computer as a guide towards manual identification of semantic structures. We propose an interactive protocol in order to ``extract'' and explore a user's intuitions of the thematic structures of a given text. This protocol, along with the data structure which stores semantic information from the user, is managed by a software system, named PASTEL, which is described herein. Using an example, we show how this method can be useful in deeply exploring textual content.