Topic Model with Fully-Connected Layers for Short-Text Classification

Abstract

Listen

Translate article

The Internet news always contain a few words, which affects the performance of text analysis and classification. To solve the problem, based on the co-occurrence of words, Biterm Topic Model (BTM) builds the word biterms in corpus to extract the topic features for short-text classification. However, BTM ignores the relationship of topics. To overcome the limitation, we propose a model which integrates fully-connected layers of convolutional neural networks (CNN) into the framework of BTM. The model reflects the co-occurrence of words by biterms and builds the relationship of topics in the fully-connected layers by connecting the topic features with the cells of neighbor fully-connected layer. Thus, the model makes full use of topic features and the relationship of topics and enhances the representations of short-text documents. In the experiments of short-text classification, the performances of the model outperform the other baseline models on two benchmark datasets.