Discovering Correlated Entities from News Archives

Most textual documents contain references to real-word entities such as people, locations and organizations. The understanding of their correlations is behind many applications including social relationship construction platform and major search engines, etc. This paper aims to discover entity correlations from news archives by means of the proposed hierarchical Entity Topic Model (hETM). hETM is a semantic-based analysis model which follows the gist of probabilistic topic models and in which a directed acyclic graph (DAG) is leveraged to capture arbitrary topic correlations. Entity extraction is taken as a preprocessing step of our model and we then employ different generative processes for ordinary words and entities. The discovering of entity correlations is achieved via the analysis of the dependencies between entities and their associated topics as well as topic correlations. We evaluate the approach upon BBC news dataset and results demonstrate the higher quality of discovered entity correlations compared with existing methods.