Translation Model Training With Error-Driven Dataset Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language translation models face challenges in achieving high translation performance due to translation errors such as grammatical errors, punctuation errors, and inconsistencies in terminology, which affect the accuracy and quality of translated text.
Innovation Solution
A method for training a natural language translation model involves processing electronic parallel documents to obtain aligned sentences, identifying and correcting translation errors, and iteratively refining the training set to improve translation performance, using a neural-machine translation model specifically adapted for legal contexts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a natural language translation model is trained using conventional methods with static training sets, then the training process is simple and fast, but translation accuracy and performance are insufficient due to grammatical errors, punctuation errors, and terminology inconsistencies
Solution Approach 1:
The patent implements a feedback mechanism where translation errors (grammatical errors, punctuation errors, terminology inconsistencies) are detected in the model's output and fed back into the training process. The error detection module identifies specific error types, and this feedback is used to iteratively refine the training set and adjust model parameters, continuously improving translation accuracy through multiple training cycles.
Solution Approach 2:
The patent performs preliminary actions by pre-processing parallel sentences to identify and correct potential errors before they affect model training. The system pre-processes training data to fix grammatical errors, punctuation issues, and terminology inconsistencies in advance, and maintains a dynamic error database that is prepared beforehand to guide the iterative training process.
2Reliability
If the training set is refined multiple times to correct translation errors, then translation performance improves, but training time and computational resources increase
Solution Approach 1:
The patent applies partial action by focusing error correction efforts on the most critical error types in each iteration rather than attempting to correct all errors simultaneously. The error detection module prioritizes grammatical errors, punctuation errors, and terminology inconsistencies based on their impact on translation quality, allowing the model to achieve significant improvement without requiring exhaustive correction in every training cycle.
Solution Approach 2:
The patent implements periodic action through iterative training cycles where the model is trained, evaluated, and then re-trained with corrected errors at regular intervals. The system performs multiple rounds of training with progressively refined training sets, allowing the model to improve translation quality over time while managing computational resources through structured periodic updates rather than continuous retraining.
3Measurement precision
If error detection and correction is performed on all parallel sentences, then translation accuracy improves, but processing complexity and computational load increase
Solution Approach 1:
The patent applies local quality by focusing error detection and correction efforts on specific critical areas rather than uniformly processing all parallel sentences. The error detection module targets grammatical errors, punctuation errors, and terminology inconsistencies in high-priority regions of the training data, while applying less intensive processing to other areas, thereby improving translation accuracy without proportionally increasing overall processing complexity.
Solution Approach 2:
The patent segments the error detection and correction process into distinct modules that handle different error types independently. The system divides parallel sentences into manageable units and processes them through specialized error detection components that identify grammatical errors, punctuation errors, and terminology issues separately, making the complex processing task more manageable and efficient.
Data Source
AI summary
A computer-implemented method for training a natural language translation model. The computer-implemented method includes: processing one or more sets of electronic parallel documents to obtain a plurality of aligned parallel sentences; creating a first training set including a subset of the plurality of aligned parallel sentences; and training the natural language translation model in a first stage using the first training set. The computer-implemented method further includes: modifying the first training set based on translation errors detected after the first stage of training; creating a second training set based on the modified first training set and at least some of the plurality of aligned parallel sentences not in the first training set; and training the natural language translation model in a second stage using the second training set so as to improve translation performance of the natural language translation model.