DirBN

The demo code of PFA+DirBN in the paper of "Dirichlet belief networks for topic structure learning", NeurIPS 2018 Arxiv.

Key features:

1. DirBN discovers topic hierarchies on topic-word distributions.
2. DirBN flexibly combines with many other topic models.
3. DirBN enjoys better perplexity and topic coherence, especially for short texts.

Run PFA+DirBN

0. The code is a mixture of Matlab and C++. The code has been tested in MacOS and Linux (Ubuntu). To run it on Windows, you need to re-compile all the .c files with MEX and a C++ complier.

1. Requirements: Matlab 2016b (or later).

2. We have offered the TMN dataset used in the paper, which is stored in MAT format, with the following contents:

• x: a V by N count (sparse) matrix for N documents with V words in the vocabulary
• voc: the words in the vocabulary
• train_idx: the indexes of documents for training
• test_idx: the indexes of documents for testing

Please prepare your own documents in the above format. If you want to use this dataset, please cite the original papers, which are cited in our paper.

3. Run PFA_DirBN_demo.m

Use DirBN with other models

DirBN is a hierarchical construction on top of topic-word distributions and leaves the construction on doc-word distributions untouched. init_DirBN.m, sample_DirBN.m, sample_DirBN_beta.m, and sample_DirBN_counts.m can be viewed as an independent package of DirBN.

To combine DirBN with other topic models than PFA, simply call init_DirBN.m before the inference begins and call sample_DirBN.m in each iteration after the topic assignments are sampled.

Notes

1. CRT_sum_mex_matrix_v1.c, CRT_sum_mex_v1.c, Mult_Sparse.c, Multrnd_Matrix_mex_fast_v1.c, PartitionX_v1.m, and Sample_rk.m are borrowed from GBN of Mingyuan Zhou. If you want to use the above code please cite the related papers. collapsed_gibbs_topic_assignment_mex.c is modified from the code of GBN.

2. If you find any bugs, please contact me by email (ethanhezhao@gmail.com).