Summarizing Topics: From Word Lists to Phrases

• Activity Number: 433
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Lauren Hannah* and Hanna Wallach
• Companies: Columbia University and Microsoft Research
• Keywords: topic model ; Bayes factor ; interpretability

Abstract

We propose a two-stage approach to generate descriptive phrases from the output of a multinomial topic model. First, we propose a Bayesian way to statistically select phrases from a document corpus, using priors associated with Latent Dirichlet Allocation (LDA). Second, the selected phrases are combined with the topic dictionary to make a list of candidate phrases, which are ranked in terms of topic descriptiveness using a metric based on the weighted Kullback Leibler divergence between the topic probabilities implied by the phrase and topic probabilities implied by the underlying model. The results are assessed for interpretability on a set of diverse corpora using Mechanical Turk.