Attractive or Faithful? Popularity-Reinforced Learning for Inspired Headline Generation

  • 2020-02-06 04:37:44
  • Yun-Zhu Song, Hong-Han Shuai, Sung-Lin Yeh, Yi-Lun Wu, Lun-Wei Ku, Wen-Chih Peng
  • 8

Abstract

With the rapid proliferation of online media sources and published news,headlines have become increasingly important for attracting readers to newsarticles, since users may be overwhelmed with the massive information. In thispaper, we generate inspired headlines that preserve the nature of news articlesand catch the eye of the reader simultaneously. The task of inspired headlinegeneration can be viewed as a specific form of Headline Generation (HG) task,with the emphasis on creating an attractive headline from a given news article.To generate inspired headlines, we propose a novel framework calledPOpularity-Reinforced Learning for inspired Headline Generation (PORL-HG).PORL-HG exploits the extractive-abstractive architecture with 1) Popular TopicAttention (PTA) for guiding the extractor to select the attractive sentencefrom the article and 2) a popularity predictor for guiding the abstractor torewrite the attractive sentence. Moreover, since the sentence selection of theextractor is not differentiable, techniques of reinforcement learning (RL) areutilized to bridge the gap with rewards obtained from a popularity scorepredictor. Through quantitative and qualitative experiments, we show that theproposed PORL-HG significantly outperforms the state-of-the-art headlinegeneration models in terms of attractiveness evaluated by both human (71.03%)and the predictor (at least 27.60%), while the faithfulness of PORL-HG is alsocomparable to the state-of-the-art generation model.

 

Quick Read (beta)

loading the full paper ...