Research on Chinese multi-document hierarchical topic modeling automatic evaluation methods

Hierarchical Latent Dirichlet Allocation (hLDA) has achieved good results in the supervised and unsupervised multi-document hierarchical topic modeling. However, the result is diversified. The results maintain randomness even with the same parameters. Thus, this paper proposed automatic evaluation methods for unsupervised multi-document hLDA modeling results over previous studies. This paper used 10 topics of corpus of ACL2013 multilingual multi-document summarization and found 90 topics of news as experimental corpus, then compared the different modeling results. The results showed that automatic evaluation method can provide a good reference for the optimization of the modeling results.