Abstract:
As a hot topic in the domain of text information mining, text clustering has been widely used in information retrieval, search engines, etc. However, short text has high dimensionality and sparsity of features, which make the clustering on short text face many challenges. Although a lot of efforts have been made to overcome the sparsity and high dimensionality of short text, there still exist many problems on how to further improve the quality of short text clustering. To this end, by means of biterm topic model (BTM) and paragraph vector (PV), this poper proposes a short text clustering method from the perspective of statistics and neural network algorithm, which includes two main steps: (1) Utilize the obtained word-document-topic probability distribution via BTM, in combination with the local outlier factor and JS divergence, to achieve semantic splitting of the words in the entire set of texts; (2) Input the above texts into the vectorization model PV-DBOW (Distributed Bag of Words version of Paragraph Vector) to obtain the paragraph vectors, which will further be combined with their corresponding document-topic probability distribution to form the eigenvectors of text feature. The proposed method can enrich the information carried by text features and ensures a low feature dimension. Finally, it has been shown via the experimental results that the obtained eigenvectors have stronger discriminative ability for short texts such that the clustering effect of short texts can be effectively improved, meanwhile, the negative effects of sparseness in short texts may be avoided.