Abstract
Automatic news comment generation is beneficial for real applications but hasnot attracted enough attention from the research community. In this paper, wepropose a "read-attend-comment" procedure for news comment generation andformalize the procedure with a reading network and a generation network. Thereading network comprehends a news article and distills some important pointsfrom it, then the generation network creates a comment by attending to theextracted discrete points and the news title. We optimize the model in anend-to-end manner by maximizing a variational lower bound of the true objectiveusing the back-propagation algorithm. Experimental results on two publicdatasets indicate that our model can significantly outperform existing methodsin terms of both automatic evaluation and human judgment.