Headline Generation with Recurrent Neural Network

Automatic headline generation is related to automatic text summarization and it is useful to solve information flood problems. This paper aims at generating a headline using a recurrent neural network which is based on a machine translation approach. Our headline generator consists of an encoder and a decoder and they are constructed with Long Short Term Memory, which is one of recurrent neural networks. The encoder constructs distributed representation from the first sentence in an article and the decoder generated headlines from the distributed representation. In our experiments, we confirmed that our proposed method could generate appropriate headlines but in some articles this method generates meaningless headlines. The results show that our proposed method is superior to another approach, statistical machine translation from the viewpoint of ROUGE, which is an evaluation score of automatic text summarization. Furthermore, we could find that using an input sentence in reverse order improves the quality of headline generation.