The ontology learning from text cycle consists of the consecutive phases of term, synonym, concept, taxonomy and relation extraction. The paper touches the problems of a low efficiency in the current term extraction methods which are handled by a combination of statistic (frequency-based) and linguistic approaches. We present a novel method to extract terms that uses only shallow linguistic information. It is proposed to explore a different set of linguistic layers and support a classic POS n-gram model with additional context information basedon proximity window features. The method is evaluated on twosubstantially different corpora to produce better results than the classic measures, including standard n-gram models and frequency-based approaches.