Reinforced Keyphrase Generation with Multi-Dimensional Reward

Abstract

Keyphrase Generation(KG), aiming to generate a set of keyphrases from source documents to help people quickly understand sufficient information, is always a fundamental task in natural language processing. Traditional KG models tend to focus on the correctness of predictions but ignore the similarity between predictions and ground-truth keyphrases, restraining the model from learning deep semantic patterns. To address this problem, we propose a Multi-Dimensional Reward Reinforcement Learning model (MDRRL) for keyphrase generation. Specifically, MDRRL consists of two components: an Actor network that can generate keyphrases and interact with the environment and a Critic network that evaluates the behavior of the Actor network and provides corresponding reward. Additionally, we propose a Multi-Dimensional Reward (MDR) within the reinforcement learning framework, which accounts for both semantic similarity and quantity, to incentivize the model to generate more semantically appropriate and competent keyphrases. Experiments on five datasets show that our proposed Reinforcement Learning framework using Multi-Dimensional Reward outperforms the traditional keyphrase generation frameworks based on evaluation metrics.