Nowadays, internet public opinion has a rapid spread and great influence, and topic detection plays an irreplaceable role in the supervision of public opinion. Aiming at the problems of incomplete feature extraction and high feature dimension in traditional methods, this study proposes LDA&&Word2Vec text representation model based on time decay factor, which combines the hidden subject features by LDA model with the semantic features by Word2Vec model, and adds time decay factor, which can reduce the dimension and improve the integrity of text features. At the same time, this study proposes a Single-Pass-SOM clustering model, which solves the problem of setting initial neurons in SOM model, and improves the accuracy of topic clustering. Experimental results show that the text representation model and text clustering method proposed in this study have better topic detection effect than traditional methods.