Effective extraction of domain-specific terms and named entities is a key challenge in text mining. This paper investigates the use of the k-means clustering algorithm for unsupervised extraction of unigrams and named entities from text data. The approach groups terms based on their vector representations, enabling the identification of semantically similar words without labeled data. Experiments conducted on the ACTER (Annotated Corpora for Term Extraction Research) corpus evaluate the method using precision, recall, and F1-score. Results show average scores of 25.79% precision, 40.05% recall and 30.47% F1-score, with optimal performance achieved using 40 to 60 clusters. Future work will explore algorithm optimization and comparisons with alternative extraction techniques.