Sequence Generative Adversarial Network for Chinese Social Media Text Summarization

Although the sequence-to-sequence models have achieved state-of-the-art performance in many summarization datasets, there are still some problems in the processing of Chinese social media text, such as short sentences, lack of coherence and accuracy. These issues are caused by two factors: the principle of the RNN-based sequence-to-sequence model is maximum likelihood estimation, which will lead to gradient vanishing or exploding when generating long summaries; the text in the Chinese social media is long and noisy, for which it is very difficult to generate high-quality summaries. To solve these issues, we apply a sequence generative adversarial network framework. The framework includes generator and discriminator, in which generator is used to generate summaries and discriminator is used to evaluate generated summaries. The softargmax layer is used as a connection layer to guarantee the co-training of generator and discriminator. Experiments are carried out on Large Scale Chinese Social Media Text Summarization Dataset. The length of the sentence, ROUGE score and artificial score of summary's quality are used to evaluate the generated summaries. The result shows that the sentences in the summaries generated by our model are longer and have higher accuracy.