Radar words are the primitives of radar phrase,and the extraction effect of radar words will directly affect the confidence of subsequent radar behaviour recognition.A solution is proposed for the radar word extraction problem in the case of unbalanced detection and acquisition data,i.e.an improved K-OPTICS radar word extraction algorithm based on K-means.By constructing virtual clustering centres and cluster merging,it achieves an extraction accuracy higher than 91.22% under various unbalanced sample simulation scenarios,and has better parameter insensitivity than the traditional algorithm does.