Machine translation method and device based on term protection

The method addresses the challenge of translating rare words in machine translation by using word alignment and relative position markers to replace "unk" symbols, enhancing translation accuracy and quality.

CN114462427BActive Publication Date: 2025-03-18SICHUAN LAN-BRIDGE INFORMATION TECHNOLOGY CO LTD
2 Cites 0 Cited by

Patent Information

Application Number
CN202210094875.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-26
Publication Date
2025-03-18
Estimated Expiration
2042-01-26

AI Technical Summary

Technical Problem

Traditional machine translation methods using deep learning struggle to accurately translate rare words or characters, leading to information loss as they cannot identify which specific word in the original text corresponds to the "unk" symbol in the translated text, and performance deteriorates significantly when many rare words are present.

Method used

A method that collects vast parallel corpora, performs word alignment using an IBM Model, and introduces relative position markers for unknown words, allowing the translation model to learn and replace "unk" symbols with accurate translations by referencing a custom dictionary.

Benefits of technology

This approach effectively addresses the issue of information loss and improves translation quality by enabling accurate translation of rare words, even when numerous rare words are present in the text.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention provides a machine translation method and device based on term protection, which collects a large amount of parallel corpus, performs word segmentation on it, obtains a first set, and inputs it into a trained IBM Model model to output word alignment probability; according to the word alignment probability, replaces the words in the original text with special symbols, replaces the words in the translation with special symbols and marks the relative positions, and trains the machine translation model according to the replaced parallel corpus; processes the original text to be translated based on the trained machine translation model and a custom dictionary to obtain a translation result. The present invention marks rare words in the translation by introducing relative position marks, so that the translation model can learn to generate relative position information while generating special marks, so that it can know which original word corresponding to the translation is used for replacement, so as to achieve the purpose of correctly translating rare words.
Need to check novelty before this filing date? Find Prior Art