A Deep News Headline Generation Model with REINFORCE Filter

Generating accurate and concise headlines based on news content can help people filter out interesting content and improve the quality of life. News headline generation, as the sub-application area of text summarization, has many methodological commonalities but with higher requirements for generated text quality, which is more challenging. In this paper, we propose a new model to generate news headline, named RADGen (REIN-FORCE Aided Deep Generator for news headline), which combines a Transformer-based generator with a sentence-selecting filter based on REINFORCE algorithm. We perform experiments and assessments on Chinese news headline dataset, which achieve 25.71, 8.26 and 23.58 for f-score of ROUGE-1, ROUGE-2 and ROUGE-L respectively, to demonstrate the effectiveness of the proposed model. In addition, ablation experiments show the positive roll of reinforcement learning filter in news headline generation.