Paraphrase Generation with Deep Reinforcement Learning

Automatic generation of paraphrases for a given sentence is an important yet challenging task in natural language processing (NLP), and plays a key role in a number of applications such as question answering, information retrieval and dialogue. In this paper we present a deep reinforcement learning approach to paraphrase generation. Specifically, we propose a new model for the task, which consists of a generator and a teacher. The generator, built on the sequence-tosequence learning framework, can generate paraphrases given a sentence. The teacher, modeled as a deep neural network, can decide whether the sentences are paraphrases of each other. After construction of the generator and teacher, the generator is further fine-tuned by reinforcement learning in which the reward is given by a teacher. Empirical study shows that the teacher can provide precise supervision to the generator, and guide the generator to produce more accurate paraphrases. Experimental results demonstrate the proposed model outperforms the state-of-theart methods in paraphrase generation in both automatic evaluation and human evaluation.