Extracting semantic prototypes and factual information from a large scale corpus using variable size window topic modelling

摘要

In this paper a model of textual events composed of a mixture of semantic stereotypes and factual information is proposed. A method is introduced that enables distinguishing automatically semantic prototypes of a general nature describing general categories of events from factual elements specific to a given event. Next, this paper presents the results of an experiment of unsupervised topic extraction performed on documents from a large-scale corpus with an additional temporal structure. This experiment was realized as a comparison of the nature of information provided by Latent Dirichlet Allocation and Vector Space modelling based on Log-Entropy weights. The impact of using different time windows of the corpus on the results of topic modelling is presented. Finally, a discussion is suggested on the issue if unsupervised topic modelling may reflect deeper semantic information, such as elements describing a given event or its causes and results, and discern it from pure factual data.