Most of the knowledge achieved from research activities are available as computer-like unstructured data written in natural-language papers. Automatically retrieving and representing knowledge from natural-language papers as input for computer processing is complex and challenging. In this paper, we propose a novel syntactic-relationship approach based on natural language processing, efficiently applying clustering algorithms to generate knowledge taxonomies about specific domain texts automatically. The approach considers a cloud computing case study through the collection and analysis of a set of recent publications. To assess our proposal, we conducted a quantitative comparison between different clustering-intrinsic metrics. Results showed higher popularity and coverage of the present proposal than the state-of-the-art, especially when using hierarchical clustering. The differential of our proposal lies in building a well-informative representation of knowledge with only three-quarters of the original textual data, and without any ground truth labeling.