Geographical Latent Variable Models for Microblog Retrieval

Although topic models designed for textual collections annotated with geographical meta-data have been previously shown to be effective at capturing vocabulary preferences of people living in different geographical regions, little is known about their utility for information retrieval in general or microblog retrieval in particular. In this work, we propose simple and scalable geographical latent variable generative models and a method to improve the accuracy of retrieval from collections of geo-tagged documents through document expansion that is based on the topics identified by the proposed models. In particular, we experimentally compare the retrieval effectiveness of four geographical latent variable models: two geographical variants of post-hoc LDA, latent variable model without hidden topics and a topic model that can separate background from geographically-specific topics. The experiments conducted on TREC microblog datasets demonstrate significant improvement in search accuracy of the proposed method over both the traditional probabilistic retrieval model and retrieval models utilizing geographical post-hoc variants of LDA.