Translation Model Training Data Refinement via Confidence Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional translation machine-learning models face challenges in generating high-quality translations due to the use of low-quality training examples, which are often generated by other translation models and may contain incorrect correspondences between source and target languages.
Innovation Solution
The method involves using two additional translation machine-learning models to refine the training data. The first model translates directly from the source to the target language, while the second model translates through an intermediate language. By comparing the confidence scores of these models with the base confidence score of the existing target sentence, the method replaces low-quality target sentences with more accurate variants, thereby improving the quality of the training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If training data is generated using existing translation models, then the training process can be automated and scaled, but the translation accuracy deteriorates due to propagation of errors from low-quality source translations
Solution Approach 1:
The patent introduces an intermediary verification step using a separate validation model to assess the quality of training examples before they are used for training. This mediator model evaluates whether source-target sentence pairs are genuine translations or false parallels, thereby preventing error propagation while maintaining automated training scalability
Solution Approach 2:
The system implements feedback mechanisms where translation models evaluate their own training data quality. The validation model provides feedback on training example quality, and this information is used to filter or reweight training samples, creating a closed-loop system that continuously improves translation accuracy while maintaining automated training
2Quantity of substance
If low-quality translation examples are used for training, then the training data quantity can be increased, but the model learns incorrect correspondences and translation quality deteriorates
Solution Approach 1:
The patent changes the quality parameter of training data by introducing confidence scoring and filtering mechanisms. Instead of uniformly using all available training examples, the system evaluates each example's quality metric and selectively uses only high-confidence training pairs, thereby maintaining large data volume while ensuring high translation quality
Solution Approach 2:
The system performs preliminary validation and filtering of training data before the actual training process. By pre-assessing the quality of source-target sentence pairs using validation models and confidence scoring, the system ensures that only reliable training examples are used, preventing error propagation from the outset
3Manufacturing precision
If multiple translation models are used for validation, then the translation accuracy improves through cross-verification, but the system complexity increases
Solution Approach 1:
The patent segments the translation validation task into distinct functional components: a validation model for quality assessment, confidence scoring mechanisms, and filtering modules. This segmentation allows each component to specialize in a specific aspect of quality control, improving overall accuracy while managing system complexity through modular design
Solution Approach 2:
The validation model serves multiple functions: it validates training data quality, generates confidence scores, and identifies false parallels. This multi-functionality reduces the need for separate specialized models for each task, thereby improving translation accuracy through comprehensive validation while limiting the increase in system complexity
Data Source
AI summary
A method and a server for generating a training set of data for training a translation machine-learning model are provided. The method comprises: receiving: (i) a source language corpus of texts in the source language; and (ii) a corresponding target language corpus of texts in the target language; generating a first variant of translation of a given source language phrase from the source language corpus into the target language; and a first confidence score thereof; generating a second variant of translation of the given source language phrase into the target language; and a second confidence score thereof; and in response to one of the first and second confidence scores being greater than a base confidence score associated with a respective target language phrase from the corresponding target language corpus, replacing the respective target language phrase with a respective one of the first and second variants of translation.