2019/09/03 by Florian Böhm, Yang Gao, Böhm, Florian +9 · 3 citations
Computer Science · #Advanced Text Analysis Techniques #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1909.01214
openalex publication_date 2019/09/03 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Reinforcement Learning (RL) based document summarisation systems yield\nstate-of-the-art performance in terms of ROUGE scores, because they directly\nuse ROUGE as the rewards during training. However, summaries with high ROUGE\nscores often receive low human judgement. To find a better reward function that\ncan guide RL to generate human-appealing summaries, we learn a reward function\nfrom human ratings on 2,500 summaries. Our reward function only takes the\ndocument and system summary as input. Hence, once trained, it can be used to\ntrain RL-based summarisation systems without using any reference summaries. We\nshow that our learned rewards have significantly higher correlation with human\nratings than previous approaches. Human evaluation experiments show that,\ncompared to the state-of-the-art supervised-learning systems and\nROUGE-as-rewards RL summarisation systems, the RL systems using our learned\nrewards during training generate summarieswith higher human ratings. The\nlearned reward function and our source code are available at\nhttps://github.com/yg211/summary-reward-no-reference.\n