Multilingual Short Text Classification Based on LDA and BiLSTM-CNN Neural Network

Abstract

The rapid development of the Internet makes information resources have the characteristics of language diversity, and the differences between different languages can cause difficulties in information exchange. Therefore, the integration of multilingual data resources has important application value and practical significance. In this paper, we propose a novel model based on LDA and BiLSTM-CNN to solve the multilingual short text classification problem. In our method, we make full use of topic vectors and word vectors to extract text information from two aspects, in each language, to solve the sparse problem of short text features. Then, we use Long Short Term Memory (LSTM) neural network to capture the semantic features of documents and use Convolutional Neural Network (CNN) to extract local features. Finally, we cascade the features of each language and use SoftMax function to predict document categories. Experiments on parallel datasets of Chinese, English and Korean scientific and technological literature show that our proposed the short text classification model based on LDA and BiLSTM-CNN can effectively extract text information and improve the accuracy of multilingual text classification.