Electronic Device Sentence Correction for Translation Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine translation models often produce low-quality translations due to the lack of familiarity between input sentences and the learning corpus, particularly when input sentences contain honorific expressions, grammatical features, or spacing differences not present in the training data, making it impractical to construct a comprehensive corpus for all human language expressions.
Innovation Solution
An electronic device corrects input sentences to make them more familiar to the learning corpus by using language models trained on parallel and mono corpora, determining the need for correction based on perplexity values, and then inputting the corrected sentences into a translation model to improve translation quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a translation model is trained on a limited learning corpus, then the model can be trained efficiently with available data, but the translation quality deteriorates when input sentences contain expressions not present in the corpus
Solution Approach 1:
The system performs preliminary correction of the input sentence before translation by using a language model to generate a corrected version that is more likely to appear in the training corpus. This preliminary action ensures that the translation model receives familiar input patterns, thereby improving translation quality without requiring an expanded corpus.
Solution Approach 2:
A correction model serves as an intermediary between the input sentence and the translation model. This intermediary component transforms unfamiliar or poorly formatted input sentences into corrected versions that better match the training corpus patterns, enabling the translation model to produce higher quality translations without changing the translation model itself.
2Adaptability or versatility
If the learning corpus is expanded to include all possible human language expressions, then the translation model can handle all expressions, but it becomes practically impossible to construct such a comprehensive corpus
Solution Approach 1:
Instead of creating a comprehensive corpus covering all possible expressions, the system copies or replicates existing corpus patterns through the correction model. The correction model learns from the existing corpus and generates corrected sentences that follow the same patterns, effectively extending the corpus coverage without actually expanding the training data.
Solution Approach 2:
The system changes the parameters of the input sentence (through correction) to match the distribution and patterns of the training corpus. By transforming unfamiliar expressions into familiar patterns, the system achieves broader effective coverage without requiring an expanded corpus.
3Productivity
If correction is always applied to input sentences, then translation quality may improve for unfamiliar expressions, but unnecessary corrections may degrade translation quality for already optimal sentences
Solution Approach 1:
The system applies correction selectively rather than universally. The correction model is trained to identify when correction is actually needed based on the input sentence characteristics, applying correction only in cases where it will improve translation quality while leaving already-optimal sentences unchanged.
Solution Approach 2:
The correction model uses feedback from the language model and perplexity calculations to determine whether correction is necessary. By monitoring how well the input sentence matches the training corpus patterns, the system can decide whether correction will help or harm translation quality, thereby maintaining consistency.
Data Source
AI summary
An electronic device is provided. The electronic device includes a memory and a processor. The processor is configured to, based on acquiring a first sentence in a first language, determine whether to correct the first sentence to another sentence in the first language by using a second language model trained based on a learning corpus, and based on determining to correct the first sentence to another sentence in the first language, input the first sentence into a conversion model trained to acquire another sentence having a similarity greater than or equal to a threshold value to an input sentence and acquire a second sentence in the first language which is a corrected form of the first sentence, and based on acquiring the second sentence, input the second sentence into a translation model trained based on the learning corpus and acquire a third sentence in a second language.