Emotional Human Machine Conversation Generation Based on SeqGAN

摘要

In recent years, artificial intelligence has made a significant breakthrough and progress in the field of humanmachine conversation. However, how to generate high-quality, emotional and subhuman conversation still a troublesome work. The key factor of man-machine dialogue is whether the chatbot can give a good response in content and emotional level. How to ensure that the robot understands the user's emotions, and consider the user's emotions then give a satisfactory response. In this paper, we add the emotional tags to the post and response from the dataset respectively. The emotional tags, as the emotional tags of post and response, represent the emotions expressed by this sentence. The purpose of our emotional tags is to make the chatbot understood the emotion of the input sequence more directly so that it has a recognition of the emotional dimension. In this paper, we apply the mechanism of GAN network on our conversation model. For the generator: We make full use of Encoder-Decoder structure form a seq2seq model, which is used to generate a sentence's response. For the discriminator: distinguish between the human-generated dialogues and the machine-generated ones.The outputs from the discriminator are used as rewards for the generative model, pushing the system to generate dialogues that mostly resemble human dialogues. We cast our task as an RL(Reinforcement Learning) problem, using a policy gradient method to reward more subhuman conversational sequences, and in addition we have added an emotion tags to represent the response we want to get, which we will use as a rewarding part of it, so that the emotions of real responses can be closer to the emotions we specify. Our experiment shows that through the introduction of emotional intelligence, our model can generate responses appropriate not only in content but also in emotion, which can be used to control and adjust users emotion. Compared with our previous work, we get a better performance on the same data set, and we get less ''safe'' response than before, but there will be a certain degree of existence.