Neural Network Machine Translation Number Generalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network machine translation struggles with controlling content, leading to inconsistent number translations between the original text and the translated text, with difficulties in correcting these errors through post-processing.
Innovation Solution
A numerical generalization method that involves processing a training corpus using a hidden Markov model for word segmentation, alignment, and replacing number-containing words or phrases with generalization labels, allowing for accurate translation and reducing the word list size, thereby improving training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network machine translation is used, then translation quality is improved, but control over content and numbers is lost
Solution Approach 1:
The patent segments the translation process into two independent modules: a neural network translation module for overall translation quality and a post-processing control module for number accuracy. This segmentation allows each module to specialize in its function, with the neural network handling general translation and the post-processing module correcting numerical errors, thereby resolving the contradiction between translation quality and content control.
Solution Approach 2:
The patent introduces an intermediary post-processing module that acts as a mediator between the neural network translation output and the final translated text. This intermediary module specifically targets and corrects number translation errors without affecting the overall translation quality, enabling control over content while maintaining the benefits of neural network translation.
2Adaptability or versatility
If neural network model is used, then translation capability is enhanced, but numerical translation accuracy deteriorates
Solution Approach 1:
The patent divides the translation system into a neural network component for general translation capability and a separate post-processing component for numerical accuracy. This segmentation allows the neural network to maintain its versatility and translation capability while the post-processing module ensures precise number translation, resolving the contradiction between adaptability and precision.
Solution Approach 2:
The patent implements a feedback mechanism where the post-processing module receives the translation output from the neural network, identifies numerical errors, and corrects them. This feedback loop ensures that number translation accuracy is maintained while the neural network continues to provide enhanced translation capability for diverse content.
3Manufacturing precision
If post-processing is applied to correct errors, then number consistency is improved, but system complexity increases
Solution Approach 1:
The patent extracts the number correction function from the main translation system and places it in a separate, dedicated post-processing module. This extraction isolates the complexity of error correction from the neural network translation process, improving number consistency while keeping the overall system architecture manageable by separating concerns into distinct modules.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This method enhances the accuracy of translating numbers and reduces the word list size, improving the neural network model's training efficiency by applying generalization technology without altering the model structure, thus addressing the challenge of inconsistent number translations.
Implementation Method 1
applying a word segmentation algorithm based on a hidden Markov model on a corpus to perform a normal word segmentation
Implementation Method 2
translating a generalized original text by a neural network model trained by a generalized corpus
Data Source
AI summary
A numerical generalization method for machine translation and system, computer and computer program thereof includes a training stage, in the training stage, a training corpus is processed in a special manner, and a normal training is performed without changing a structure of a neural network model; and a translation stage, in the translation stage, a generalization label in a translation is replaced with a normal translation. In the present invention, only the pre-processing and post-processing are changed to make the generalization technology applicable, which expands the application of the generalization technology in a neural network machine translation, and better adapts to the new machine translation model structure.