Adversarial sampling training method and device based on meta-learning

Abstract

The invention discloses an adversarial sampling training method and device based on meta-learning, and the method comprises the steps that: a K-dimensional probability vector P<T>(S) is outputted from a large task set T composed of K languages according to a strategy network, wherein P<T>(S),i is a sampling probability corresponding to a task set of the ith language, the first M languages with the maximum probability is selected according to the sampling probability, a task is sampled according to each language in the M languages with the maximum probability to form a training task set, and the training task set is divided into a support set and a query set; the support set carries out gradient descent on an initialization parameter theta of the speech recognition model to obtain an updated parameter;and the query set obtains a query loss vector according to an effect of querying the updated parameter, and the query loss vector is used for optimizing the initialization parameter theta to obtain an optimal model parameter. On the basis of a multi-language meta-learning speech recognition framework, the strategy network is introduced to form adversarial training, the problem of unbalanced low-resource language recognition is solved, and the training effect is improved.

Description

technical field [0001] The present invention relates to the technical field of speech recognition, in particular to a method and device for adversarial sampling training based on meta-learning. Background technique [0002] With the vigorous development of deep learning theory and related technologies, the field of speech recognition has made great progress. However, constructing an end-to-end deep speech recognition model often requires a large amount of annotated data, which is very difficult to obtain for many low-resource languages. In order to solve the above problems, there are many works using unsupervised pre-training and semi-supervised learning methods to use a large amount of unlabeled data to help low-resource target languages ​​improve the recognition effect, but these methods still require a large amount of unlabeled data in the target language. For some For small languages, unlabeled data is also very small. [0003] Therefore, transfer learning is introduce...

AI Technical Summary

Problems solved by technology

[0004] However, in the existing applications of low-resource speech recognition and low-resource code-switching speech recognition, the problem of unbalanced tasks in real scenarios is ignored. The existing technology uses the meta-information of each language equally, which leads to the effect of Loss

Images