Cross-language generative automatic summarization based on attention mechanism

Abstract

Generative automatic summarization is a basic problem in natural language processing. We propose a cross-language generative automatic summarization model. Unlike the traditional methods that have to go through machine translation, our model can directly generate a text summary of another language from a text body in one language. We use the RNNLM(Recurrent Neural Network based Language Model) structure to pre-train word vectors to obtain semantic information in different languages. We combined the Soft Attention mechanism in the Seq2Seq model, using Chinese, Korean and English to build a parallel corpus to train the model, thereby, cross-language automatic summarization can be achieved without the help of machine translation technology. Experiments show that the improvement of our proposed model on ROUGE-1, ROUGE-2 and ROUGE-L indicators reached 6%, 2.46%, and 5.13%, respectively. The experimental effect is friendly.