Computational linguistics is a complex, interdisciplinary and very diversified field, which aims at developing IT programs for linguistic data processing. In the described project, we focus our attention on the understanding and communication side, under different aspects and application contexts. Based on the treated subject, we used different, specific approaches,  according to the related linguistic data and task: Logical formalisms based on the compositional nature of the language, to implement a module to understand and to detect emotions in the context of a man-machine dialogue system; Statistical analysis tools (Kohonen maps) to represent the linguistic knowledge of a child; Linguistic patterns, to detect semantic relations; Statistical quantitative approach based on term vectorization to measure sentence similarity, then used for a multi-document summarization task. The heterogeneity of approach, tasks and related corpora, naturally led us to deal with the problem of inter-annotator agreement since annotated corpora are material on which the systems are built and on which their evaluations are based. The eclecticism of tasks and approaches reflects the one typical of the language used as a communication tool. It is important to point out that statistical quantitative approaches give rather reliable results, but they cannot express all the communication possibilities given by natural language: the described experiments give some other points of view.