The growing use of ontologies in a variety of application areas has stimulated the development of approaches proposing different degrees of automation of the ontology construction process. However, despite the real interest of these approaches, sometimes their results are of low quality. The aim of the work presented in this thesis is to contribute to the improvement of the quality of ontologies constructed automatically from texts. Our main contributions are : (1) a method for the comparison of the approaches, (2) a typology of problems that affect the quality of ontologies, and (3) a first reflection on automating the detection of quality problems. Our method for the comparison of approaches consists of three complementary steps : (1) on the basis of their degree of automation and completeness, (2) on the basis of their technical and functional characteristics, and (3) experimentally by comparing their results with a manually constructed ontology. The proposed typology organizes the quality problems according to two dimensions : errors versus unsuitable situations and logical aspects versus social aspects. Our typology contains 24 classes of problems that cover and complement the problems described in the literature. Concerning the automatic detection we have inventoried some of the existing methods for each problem in our typology and we have highlighted the problems for which the automatic detection remains an open issue. We have also proposed a heuristic for the detection of a quality problem that appears frequently in our experimentations (polysemic labels).
