Weighted-LDA-TVM: Using a Weighted Topic Vector Model for Measuring Short Text Similarity

Abstract

Topic modeling is the core task of the similarity measurement of short texts and is widely used in the fields of information retrieval and sentiment analysis. Though latent dirichlet allocation provides an approach to model texts by mining the underlying semantic themes of texts. It often leads to a low accuracy of text similarity calculation because of the feature sparseness and poor topic focus of short texts. This paper proposes a similarity measurement method of short texts based on a new topic model, namely Weighted-LDA-TVM. Latent dirichlet allocation is adopted to capture the latent topics of short texts. The topic weights are learned by using particle swarm optimization. Finally, a text vector can be constructed based on the word embeddings of weighted topics for measuring the similarity of short texts. A group of text similarity measurement experiments were performed on biomedical literature abstracts about antidepressant drugs. The experimental results prove that the proposed model has the better distinguish ability and semantic representation ability for the similarity measurement of short texts.