A New Evaluation Criterion with the Integration of Perplexity and Jensen-Shannon Divergence for Biterm Topic Model
Chengde Zhang, Jie Zheng, Zanbo Wang, Zhao Nan Jia, Fei Li · 2018
Perplexity, as a classical evaluation method for topic model, is good at predicting new documents. However, this method is not effective for Biterm Topic Model (BTM), especially short terms of web videos might leading to much more themes. In order to solve above problem, a new evaluation criterion is proposed with the integration of perplexity and Jensen-Shannon (JS) divergence. JS divergence focuses on differences and stability among themes for BTM model. Therefore, it can make up the deficiency of perplexity. The experiment demonstrates that our proposed method is more effective than several existing methods with a significant improvement.