Abstract EN:

This work deals with knowledge discovery in texts whose main applications are information retrieval using a search engine, communication filtering using an e-mail client and knowledge management using methods of knowledge engineering. We justify, define, illustrate and evaluate two new complementary extensions for text mining chains intended to save time and improve quality. These extensions apply to all the chain links and take into account the global chain and no more only one link in particular since the field is now sufficiently mature. On the one hand, one can improve the text mining chain while modeling locally and globally the information exchanges which take place. The objective is to optimize the dependencies between the various treatments in order to transmit the results between treatments among which those intractable but not those irretrievable. To reach it, we introduce recursive and mutually recursive dependent treatments. On the other hand, one can improve the text mining chain while taking into account that the majority of data are raw, without initial annotations. The objective is thus to make more controllable the training algorithms at the time of the training. To reach it, we introduce the user guided learning which consists in combining the stages of annotation and training by making coevolve the annotations, the learned model and knowledge from the user.