Neural Headline Generation with Minimum Risk Training

Abstract

Automatic headline generation is an important research area within text summarization and sentence compression. Recently, neural headline generation models have been proposed to take advantage of well-trained neural networks in learning sentence representations and mapping sequence to sequence. Nevertheless, traditional neural network encoder utilizes maximum likelihood estimation for parameter optimization, which essentially constraints the expected training objective within word level instead of sentence level. Moreover, the performance of model prediction significantly relies on training data distribution. To overcome these drawbacks, we employ minimum risk training strategy in this paper, which directly optimizes model parameters with respect to evaluation metrics and statistically leads to significant improvements for headline generation. Experiment results show that our approach outperforms state-of-the-art systems on both English and Chinese headline generation tasks.