Intelligent translation model training optimization method based on transfer learning

By dynamically calculating activation weights and selectively activating adapters, the intelligent translation model is optimized, solving the problems of misjudgment and adapter selection bias in cross-domain text translation, improving the accuracy and stability of translation, and performing particularly well in professional applications.

CN122087463BActive Publication Date: 2026-07-10EC INNOVATIONS (SHENYANG) INC
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN Β· China
Patent Type
Patents(China)
Current Assignee / Owner
EC INNOVATIONS (SHENYANG) INC
Filing Date
2026-04-27
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing intelligent translation models suffer from domain misjudgment, adapter selection bias, and catastrophic forgetting when processing cross-domain text, resulting in inaccurate and inconsistent translation results, especially in professional applications with high precision requirements.

Method used

By dynamically calculating activation weights, the target adapter in the multilingual backbone model is selectively activated, and the translation model is optimized to adapt to the domain characteristics of the target text. By combining the multilingual backbone model, the domain classifier, and the vertical domain adapter, accurate translation of the target text is achieved.

Benefits of technology

It improves the accuracy and consistency of translation, enhances the model's translation performance in specialized domain texts, solves the problems of domain misjudgment and adapter selection bias, and improves the stability and applicability of the translation model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122087463B_ABST
    Figure CN122087463B_ABST
Patent Text Reader

Abstract

The application relates to the field of natural language processing, in particular to an intelligent translation model training optimization method based on transfer learning, which comprises the following steps: obtaining a pre-trained translation model; obtaining target text to be translated and inputting the target text into the translation model; analyzing the domain features of the target text through a domain classifier; calculating the activation weight according to the domain features; activating a target adapter in multiple adapters according to the activation weight to adjust the multilingual backbone model; and using the translation model including the adjusted multilingual backbone model to translate the target text to obtain a target translation text. In this way, the pre-trained translation model is obtained, the domain features of the target text are analyzed to calculate the activation weight, the target adapter is activated to adjust the model and translation is performed, dynamic model adjustment is realized, the problems of domain misjudgment and adapter selection deviation in the prior art are solved, and the accuracy of translation is improved.
Need to check novelty before this filing date? Find Prior Art