Delta Language Model for Machine Translation Quality Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine translation systems face challenges in filtering out test data from training data to avoid poor feature weights and incorrect quality estimation, especially when updating or expanding test data affects other language pairs, leading to quality loss and resource inefficiency.
Innovation Solution
Generating a delta language model based on test data to evaluate and store difference values, allowing for efficient evaluation of new data without regenerating the full language model, and enabling flexible addition of new test sentences without affecting other language pairs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If test data is removed from training data to avoid poor feature weights and incorrect quality estimation, then translation quality is improved, but filtering complexity increases and data loss occurs
Solution Approach 1:
The patent segments the language model into two distinct components: a base language model trained on training data without test data, and a delta language model trained specifically on test data. This segmentation allows the base model to provide reliable quality estimation while the delta model handles test data-specific patterns, eliminating the need for complex filtering of test data from the training set.
Solution Approach 2:
The patent extracts test data from the training data set to create a separate delta language model. By taking out test data and training a dedicated delta model on it, the system can evaluate translation quality using the base model (which wasn't exposed to test data) while still capturing test data characteristics in the delta model for final translation enhancement.
2Measurement precision
If test data is filtered at the single-sentence level to handle frequent web sentences, then translation accuracy is improved, but data loss increases and processing time increases
Solution Approach 1:
The patent performs preliminary action by pre-training the base language model on training data excluding test data, and pre-training the delta language model on test data. This preliminary separation and training eliminates the need for real-time filtering during translation operations, significantly reducing processing time while maintaining high translation accuracy.
3Reliability
If updating test data for a particular language pair is performed, then translation quality for that pair is improved, but other language pairs are affected
Solution Approach 1:
The patent segments the language model training into language-pair-specific components. Each language pair has its own base language model trained on its specific training data and its own delta language model trained on its specific test data. This segmentation ensures that updating test data for one language pair only affects that specific language pair's delta model, leaving other language pairs completely independent and unaffected.
4Reliability
If the full language model is regenerated when adding new test sentences, then translation quality is maintained, but resource efficiency decreases
Solution Approach 1:
The patent extracts the delta language model as a separate, updatable component that handles test data. When new test sentences need to be added, only the delta language model for the affected language pair needs to be regenerated, not the entire full language model. This extraction approach maintains translation quality while dramatically improving resource efficiency by limiting regeneration scope to only the necessary delta model.
Data Source
AI summary
Among other disclosed subject matter, a computer-implemented method for generating a language model includes generating a language model based on a data set, the data set including at least training data and test data. The test data is to be used in testing the language model. The method includes identifying the test data in the data set. The method includes generating a delta language model based on at least the test data, the delta language model configured for evaluating the test data.