加入目标指导的强化对抗文本生成方法研究

Text generation via reinforced adversarial training with goal guidance

针对有监督的深度神经网络文本生成模型容易造成错误累积的问题,提出一种基于强化对抗思想训练的文本生成模型。通过将生成对抗网络鉴别器作为强化学习的奖励函数及时指导生成模型优化,尽量避免错误累积;通过在生成过程中加入目标指导特征帮助生成模型获取更多文本结构知识,提升文本生成模型真实性。在合成数据和真实数据集上的实验结果表明,该方法在文本生成任务中,较之前的文本生成模型在准确率和真实性上有了进一步的提高,验证了加入目标指导的强化对抗文本生成方法的有效性。

To address the problem that the supervised deep neural network text generation model was prone to the accumulation of errors,this paper proposed a text generation model based on reinforced adversarial training.By using the discriminative model to guide the training of the generative model as a reinforcement learning policy,it solved the accuracy of the traditional model.And by combining a goal direction model in the generation process,it helped the generation model to acquire more knowledge of sentence structure and improved the authenticity of the text generation.Experimental results on synthetic data and real data show that the generated model has a higher accuracy and authenticity than previous text generation models in the task of text generation,and verified the effectiveness of the reinforced adversarial text generation method with goal guidance.