In the context of globalization, companies need to capitalize on their knowledge. The knowledge of a company is present in two forms tacit and explicit. Explicit knowledge represents all formalized information i.e all documents (pdf, words ...). Tacit knowledge is present in documents and mind of employees, this kind of knowledge is not formalized, it needs a reasoning process to discover it. The approach proposed focus on extracting tacit knowledge from textual documents. In this paper, we propose hierarchical word clustering as an improvement of word clusters generated in previous work, we also proposed an approach to extract relevant bigrams and trigrams. We use Reuters-21578 corpus to validate our approach. Our global work aims to ease the automatic building of ontologies.