Neural Keyphrase Generation via Reinforcement Learning with Adaptive\n Rewards
Abstract
Generating keyphrases that summarize the main points of a document is a\nfundamental task in natural language processing. Although existing generative\nmodels are capable of predicting multiple keyphrases for an input document as\nwell as determining the number of keyphrases to generate, they still suffer\nfrom the problem of generating too few keyphrases. To address this problem, we\npropose a reinforcement learning (RL) approach for keyphrase generation, with\nan adaptive reward function that encourages a model to generate both sufficient\nand accurate keyphrases. Furthermore, we introduce a new evaluation method that\nincorporates name variations of the ground-truth keyphrases using the Wikipedia\nknowledge base. Thus, our evaluation method can more robustly evaluate the\nquality of predicted keyphrases. Extensive experiments on five real-world\ndatasets of different scales demonstrate that our RL approach consistently and\nsignificantly improves the performance of the state-of-the-art generative\nmodels with both conventional and new evaluation methods.\n
Authors: Hou Pong Chan, Chen Wang, Lu Wang, Irwin King
Published in: arXiv (Cornell University) (2019)