Machine Translation Model Training Using Semantic Similarity Mining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semantic similarity models in machine translation have poor accuracy in identifying target sentences with the same meaning due to unreasonable similarity requirements, leading to inconsistent translation results.
Innovation Solution
A method involving joint training of machine translation and semantic similarity models using parallel corpora to create and refine training sample sets, optimizing the semantic similarity model's accuracy by mining similar and negative samples, and iteratively fine-tuning both models to enhance translation confidence and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the semantic similarity model uses a compulsive requirement that the similarity between source sentence and target sentence is higher than similarity between source sentence and any other target sentence, then the training objective function can be minimized, but the accuracy of identifying target sentences with the same meaning deteriorates
Solution Approach 1:
The patent changes the parameter of similarity comparison from comparing with all other target sentences to comparing only with translated target sentences that have the same meaning. This parameter change resolves the contradiction by making the similarity requirement reasonable - the model now only needs to ensure higher similarity for semantically equivalent translations rather than all possible target sentences, thus maintaining training convergence while improving identification accuracy
Solution Approach 2:
The patent segments the set of all target sentences into two categories: translated target sentences (with the same meaning as source) and non-translated target sentences (with different meanings). The semantic similarity model is then trained to distinguish between these segments, applying different similarity requirements to each segment. This segmentation resolves the contradiction by eliminating the unreasonable comparison between semantically different sentences
2Productivity
If the semantic similarity model is trained with parallel corpus using traditional methods, then the training process can be completed, but the accuracy of identifying translated target sentences with same meaning deteriorates
Solution Approach 1:
The patent applies preliminary action by using the machine translation model to pre-identify which target sentences are translated versions with the same meaning before training the semantic similarity model. This preliminary classification of training samples allows the semantic similarity model to be trained with correctly labeled data, resolving the contradiction between completing training and achieving high accuracy in identifying translated target sentences
Data Source
AI summary
A method and apparatus for training models in machine translation, an electronic device and a storage medium are disclosed, which relates to the field of natural language processing technologies and the field of deep learning technologies. An implementation includes mining similar target sentences of a group of samples based on a parallel corpus using a machine translation model and a semantic similarity model, and creating a first training sample set; training the machine translation model with the first training sample set; mining a negative sample of each sample in the group of samples based on the parallel corpus using the machine translation model and the semantic similarity model, and creating a second training sample set; and training the semantic similarity model with the second training sample set.