Ranking coherence in topic models using statistically validated networks

Probabilistic topic models have become one of the most widespread machine learning techniques in textual analysis. Topic discovering is an unsupervised process that does not guarantee interpretability its output. Hence, automatic evaluation coherence has attracted interest many researchers over last decade, and it open research area. This article offers a new quality method based on statistically validated networks (SVNs). The proposed probabilistic approach consists representing each as weighted network probable words. presence link between pair words assessed by validating their co-occurrence sentences against null hypothesis random co-occurrence. allows to distinguish high-quality low-quality topics, making use battery statistical tests. significant pairwise associations represented links SVN might reasonably be expected strictly related semantic topic. Therefore, more connected network, coherent question. We demonstrate effectiveness through analysis real text corpus, which shows measure correlated with human judgement than state-of-the-art measures.