Are Semantically Coherent Topic Models Useful for Ad Hoc Information Retrieval?

International audienceThe current topic modeling approaches forInformation Retrieval do not allow to explicitlymodel query-oriented latent topics.More, the semantic coherence of thetopics has never been considered in thisfield. We propose a model-based feedbackapproach that learns Latent Dirichlet Allocationtopic models on the top-rankedpseudo-relevant feedback, and we measurethe semantic coherence of those topics.We perform a first experimental evaluationusing two major TREC test collections.Results show that retrieval performancestend to be better when using topicswith higher semantic coherence