Method for accelerating running speed of translation software in small intelligent mobile equipment

A mobile device and running speed technology, which is applied in natural language translation, neural learning methods, biological neural network models, etc., can solve the problems of large-scale models and inability to deploy, so as to improve running speed, reduce hardware cost, and solve Easy to freeze effect

CN112287697APending Publication Date: 2021-01-29沈阳雅译网络技术有限公司
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Current Assignee / Owner
Publication Date
2021-01-29

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention discloses a method for accelerating the running speed of translation software in small intelligent mobile equipment. The method comprises the following steps of: 1) training two neural machine translation models based on a self-attention mechanism, including a large model and a small model; 2) constructing a parameter generator of which the input is the parameter of the large model and the output is the parameter of the small model, and transmitting knowledge in the parameter of the large model to the small model through a training parameter generator to complete a knowledge migration process; 3) decoding the small model to obtain a decoding result of the machine translation model, and proving the practical value of the small model; and 4) applying the small model to the small intelligent mobile device to realize the application of rapid decoding. According to the method, the advantage that knowledge of a large model is transferred to a small model through knowledge migration is fully utilized, the method can be achieved on the basis of the model with the best current performance, and on the premise that the performance of the model is hardly reduced, model compression is achieved, and the real-time corresponding speed is increased.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to a neural machine translation model compression and acceleration technology, in particular to a method for accelerating the running speed of translation software in a small intelligent mobile device. Background technique

[0002] Small smart mobile devices have become an indispensable source of knowledge for people. Due to people's demand for various types of applications, various APP software is frequently downloaded and stored by users of small smart mobile devices, resulting in the problem of small smart mobile devices lagging. Simply increasing the capacity of small smart mobile devices will lead to high hardware costs and increase the economic burden of users. Therefore, from another perspective, the space occupied by APP software on small smart mobile devices should be reduced as much as possible.

[0003] Machine translation has a wide range of application requirements at this stage, and intelligent translation software i...

Examples

Embodiment Construction

[0031] The present invention will be further elaborated below in conjunction with the accompanying drawings of the description.

[0032] From the perspective of model compression, the present invention transfers knowledge from a large model to a small model in a neural machine translation system based on a self-attention mechanism, aiming to achieve model compression and improve real-time response speed at the cost of small performance loss.

[0033] A kind of small model generation method based on machine translation task of the present invention comprises the following steps:

[0034] 1) Train two neural machine translation models based on the self-attention mechanism, including a large model and a small model;

[0035] 2) Build a parameter generator whose input is the parameter of the large model and output is the parameter of the small model, and transfer the knowledge in the parameters of the large model to the small model by training the parameter generator to complete t...