Semantic Topic Modelling

Abstract

Topic models came to improve the way search, browse and summarization of large sets of texts is performed. These models are used for uncovering the main theme of the documents in a corpus, where topics are probability distributions over a collection of words that is representative of a document. The most widely used topic model is called Latent Dirichlet Allocation (LDA) and it enables for documents to be characterized by more than one topic. This allows for a more accurate representation of what happens with real documents, where a text may have more than one underlying theme. However, this popular model is still far from producing excellent topics, given that it does not account for the semantic relations between words. It may thus result in redundant topics that contain di erent words, but with the same meaning. This thesis o ers a way to improve the LDA algorithm and, hence, solve the problem of not considering the semantics of words. The model proposed here uses the LDA algorithm as a starting point, however some changes are made, since it is our interest to introduce semantic relations in this model. A main component of the proposed model is the use of a lexical database for English, WordNet, which enables the integration of semantics by accessing its content.