Training method of dialogue generation model and dialogue generation method and device

Abstract

The invention discloses a training method of a dialogue generation model and a dialogue generation method and device, and relates to the field of artificial intelligence. The method comprises the following steps: encoding a context sample to obtain a first hidden layer variable, and identifying the first hidden layer variable to obtain a prior hidden layer variable; encoding the reply sample to obtain a second hidden layer variable; encoding the reply similar sample to obtain a third hidden layer variable; according to the Gaussian mixture distribution of the first hidden layer variable, the second hidden layer variable and the third hidden layer variable, identifying to obtain a posterior hidden layer variable; and matching the prior hidden layer variable and the posterior hidden layer variable, and performing adversarial training on the dialogue generation model. According to the method, similar samples in a sample set are adopted, posteriori distribution of a dialogue generation model is fitted through Gaussian mixture distribution of the similar samples, the purpose of fitting more complex semantics is achieved, and complex semantics and high variability of dialogues can be captured.

Description

technical field [0001] The present application relates to the field of artificial intelligence, in particular to a training method for a dialogue generation model, a dialogue generation method and a device. Background technique [0002] Dialogue systems in the open domain are widely used in industry and academia, which can generate diverse and relevant responses. [0003] Dialog generation models based on Variational Auto-Encoders (VAEs) generate responses with diversity and relevance given the context of different topics. Taking the Wasserstein Auto-Encoder (Dialog Wasserstein Auto-Encoder, DialogWAE) as an example, the training process of the dialog generation model is explained. First, DialogWAE obtains a mixed Gaussian distribution through prior network (Prior Network) learning context, based on the mixed Gaussian Distribution sampling to obtain a random variable; then, the above random variable is transformed into a priori hidden layer variable through the generator; s...

AI Technical Summary

Problems solved by technology

[0005] The embodiment of the present application provides a dialogue generation model training method, dialogue generation method and device, which can solve the problem that the posterior distribution fitted by a simple Gaussian distribution may not be able to capture the complex semantics and high variability required for answer generation question

Images