Text Translation Model Balancing Speed and Quality via Dual Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text translation models face challenges in balancing translation speed and effectiveness, as autoregressive decoding provides good translation quality but is slow, while non-autoregressive decoding improves speed but reduces translation quality due to lack of dependency-based parallel decoding.
Innovation Solution
A training method that combines autoregressive and non-autoregressive decoding with length prediction to enhance feature extraction and translation effectiveness, by performing feature extraction on source text data, obtaining target feature vectors, and training a model using both decoding types and length predictions to achieve a balanced translation model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If autoregressive decoding is used for text translation, then translation quality is improved, but translation speed deteriorates
Solution Approach 1:
The patent merges autoregressive decoding and non-autoregressive decoding into a unified model framework. The model can dynamically switch between or combine both decoding strategies, allowing it to leverage the quality benefits of autoregressive decoding while incorporating the speed advantages of non-autoregressive decoding, thus resolving the contradiction between translation quality and translation speed
Solution Approach 2:
The patent introduces dynamic length prediction and adaptive decoding mechanisms that allow the model to adjust its decoding strategy based on the input sequence characteristics. This dynamic approach enables the model to optimize the balance between translation quality and speed for different translation scenarios, rather than being constrained to a fixed decoding mode
2Productivity
If non-autoregressive decoding is used for text translation, then translation speed is improved, but translation quality deteriorates
Solution Approach 1:
The patent introduces length prediction modules as intermediary components that guide the non-autoregressive decoding process. These modules predict the length of the target sequence in advance, allowing the model to allocate computational resources more effectively and improve translation quality while maintaining the speed benefits of parallel decoding
Solution Approach 2:
The patent performs length prediction as a preliminary action before the main translation decoding process. By predicting the target sequence length in advance, the model can better plan its decoding strategy and improve translation quality without sacrificing the speed advantages of non-autoregressive parallel decoding
Data Source
AI summary
A training method, a text translation method, an electronic device, and a storage medium, which relate to a field of artificial intelligence, in particular to fields of natural language processing and deep learning technologies. A specific implementation solution includes: performing a feature extraction on source sample text data to obtain a sample feature vector sequence; obtaining a target sample feature vector according to the sample feature vector sequence; performing an autoregressive decoding and a non-autoregressive decoding on the sample feature vector sequence, respectively; performing a length prediction on the target sample feature vector; training a predetermined model by using translation sample data, the autoregressive text translation result, the non-autoregressive text translation result, a true length value of the source sample text, the first predicted length value, a true length value of the translation sample text, and the second predicted length value to obtain the text translation model.