Topic extraction over danmaku is an important task due to the prevalence of danmaku text on the video website. Directly applying traditional topic models on danmaku text can't work well. The underlying cause is that danmaku text is very short, unconventional and lacking explicit meaning. In this paper, an improved topic model that extends BTM is proposed to infer topics from danmaku text. The special steps in our method are that: (1) a pretraining model is trained on danmaku corpus to obtain word and sentence embeddings; (2) danmaku texts are clustered to generate distinct pseudo-danmaku texts; (3) biterms with the same or similar word pairs are removed from biterm set. Experimental results show that our method can improve the diversity among topics and find some special topic words.