Semantic resources are of great value for a lot advanced tasks in Natural Language Processing such as Information Extraction or text classification. However the maintenance of such resources presents great challenges. This thesis examines approaches to automatically extend an automotive domain taxonomy by learning synonyms and thus improving coverage. These domain-specific synonyms are learned directly from text in car repair reports. We evaluate two distributional semantic models as approach to model semantic similarity. Due to the difficulty of the task traditional methods have not yet examined how proper synonym candidates can be selected from text. Our proposed method produces satisfying results, but synonym discovery is still a difficult task and cannot be fully automated. The quality control of a human editor is still necessary.