The natural language descriptions of the capabilities of manufacturing companies can be found in multiple locations including company websites, legacy system databases, and ad hoc documents and spreadsheets. To unlock the value of unstructured capability data and learn from it, there is a need for developing advanced quantitative methods supported by machine learning and natural language processing techniques. This research proposes a hybrid unsupervised learning methodology using K-means clustering and topic modeling techniques in order to build clusters of suppliers based on their capabilities, automatically infer topics from the created clusters, and discover nontrivial patterns in manufacturing capability corpora. The capability data is extracted either directly from the website of manufacturing firms or from their profiles in e-sourcing portals and directories. Feature extraction and dimensionality reduction process in this work are supported by N-gram extraction and latent semantic analysis (LSA) methods. The proposed clustering method is validated experimentally based on a dataset composed of 150 capability descriptions collected from web-based sourcing directories such as the Thomas Net directory for manufacturing companies. The results of the experiment show that the proposed method creates supplier cluster with high accuracy. Two example applications of the proposed framework, related to supplier similarity measurement and automated thesaurus creation, are introduced in this paper.