Hierarchical Topic Model Inference by Community Discovery on Word Co-occurrence Networks
Résumé
The most popular topic modelling algorithm, Latent Dirichlet Allocation, produces a simple set of topics. However, topics naturally exist in a hierarchy with larger, more general super-topics and smaller, more specific sub-topics. We develop a novel topic modelling algorithm, Community Topic, that mines communities from word co-occurrence networks to produce topics. The fractal structure of networks provides a natural topic hierarchy where sub-topics can be found by iteratively mining the sub-graph formed by a single topic. Similarly, super-topics can by found by mining the network of topic hyper-nodes. We compare the topic hierarchies discovered by Community Topic to those produced by two probabilistic graphical topic models and find that Community Topic uncovers a topic hierarchy with a more coherent structure and a tighter relationship between parent and child topics. Community Topic is able to find this hierarchy more quickly and allows for on-demand sub- and super-topic discovery, facilitating corpus exploration by researchers.