Translation Model Training via Adversarial Discriminator Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network machine translation systems face challenges in balancing multiple tasks, leading to low translation accuracy and issues like missed-translation and repeated-translation phenomena due to the complexity of the decoder's tasks.
Innovation Solution
A translation model is trained using a generative adversarial network (GAN) with a discriminator model to improve translation accuracy by calculating the similarity between the target and reference sentences, and using this similarity as a weight coefficient for the translation model, thereby addressing the decoder's task imbalance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the decoder performs multiple tasks (recording translated content, content to be translated, and fluency information), then the translation system can handle complex translation requirements, but the decoder cannot balance all tasks leading to low translation accuracy and missed-translation or repeated-translation phenomena
Solution Approach 1:
The patent introduces an attention mechanism that segments the decoder's focus across different source sentence positions. The attention mechanism divides the translation task into position-specific sub-tasks, allowing the decoder to selectively attend to relevant parts of the source sentence for each target word generation, thereby resolving the conflict between handling multiple translation requirements and maintaining accuracy.
2Loss of information
If the decoder records all necessary information (current translated content, content to be translated, fluency information), then the translation process can be comprehensive, but the complexity of tasks increases making it impossible to balance all tasks
Solution Approach 1:
The patent introduces an attention mechanism as an intermediary between the source sentence and the decoder output. This intermediary selectively weights and transmits relevant information from different source positions to the decoder at each generation step, reducing the burden on the decoder to manually track and balance multiple information streams while preserving comprehensive information retention.
3Ease of manufacture
If the translation model uses a standard training approach without adversarial training, then the training process is simpler, but the translation accuracy remains low and missed-translation phenomena occur
Solution Approach 1:
The patent implements adversarial training where a discriminator network provides feedback to the translation model about the quality and authenticity of generated translations. This feedback mechanism guides the translation model to improve its output by learning from the discriminator's evaluations, thereby increasing translation accuracy while maintaining a relatively straightforward training framework through the use of standard GAN training procedures.
Data Source
AI summary
A translation model based training method is provided for a computer device. The method includes inputting a source sentence to a translation model, to obtain a target sentence outputted by the translation model; determining a fidelity of the target sentence to the source sentence; using the target sentence and a reference sentence as input of a discriminator model, using the fidelity as output of the discriminator model, and training the discriminator model on a performance of calculating a similarity between the target sentence and the reference sentence; outputting the similarity by using the discriminator model; and using the source sentence as input of the translation model, using the target sentence as output of the translation model, and using the similarity as a weight coefficient, and training the translation model on a performance of outputting the corresponding target sentence according to the input source sentence.