The TextTilling algorithm computed the cosine between neighboring text spans (the normalized dot product of vectors of raw word counts), showing that sentences or paragraph in a subtopic have high cosine with each other, but not with sentences in a neighboring subtopic.