Compact Neural Translation Model Construction via Imitation Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural translation models require high memory and computation resources, making them unsuitable for terminal environments with limited resources, and they often rely on network connections, which may not be stable for translation and interpretation services.
Innovation Solution
A method and apparatus for constructing a compact translation model by imitation learning a parameter distribution and tree search structure from a pre-built reference model, involving pruning and quantizing parameters, and re-learning to optimize the model for reduced resource usage and network independence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural translation model is used to achieve high translation accuracy, then translation quality is improved, but memory capacity and computation performance requirements increase significantly
Solution Approach 1:
The patent applies parameter changes by quantizing model parameters from high precision (32-bit floating point) to low precision (8-bit integer), and by pruning unnecessary connections in the neural network. This reduces the memory capacity and computation performance requirements while maintaining acceptable translation accuracy, enabling deployment on terminal devices with limited resources.
Solution Approach 2:
The patent extracts and removes unnecessary or less important parameters and connections from the neural translation model through pruning techniques. By eliminating redundant components, the model size and computational complexity are reduced, allowing the model to run on devices with constrained memory and processing power while preserving essential translation functionality.
2Measurement precision
If a large-sized neural model is deployed to provide high-quality translation, then translation performance is improved, but network connection becomes mandatory due to resource constraints
Solution Approach 1:
By quantizing model parameters to low precision and optimizing the model structure, the patent enables the translation model to be deployed locally on terminal devices without requiring network connection for inference. This provides adaptability to work in offline environments while maintaining translation quality, resolving the contradiction between model size and network dependency.
3Device complexity
If a compact translation model is constructed for terminal deployment, then resource usage is reduced, but translation accuracy and speed may deteriorate
Solution Approach 1:
The patent applies partial action by selectively pruning only the least important parameters and connections while retaining the essential ones needed for accurate translation. This selective approach ensures that the compact model maintains sufficient translation accuracy and speed by preserving the most critical model components while removing only the redundant parts.
Solution Approach 2:
The patent employs feedback mechanisms during the pruning and quantization process, where the model performance is continuously evaluated and the pruning/quantization levels are adjusted accordingly. This iterative optimization ensures that the compact model achieves the best possible translation accuracy and speed given the resource constraints, preventing excessive degradation of performance.
Data Source
AI summary
Provided are a method and apparatus for constructing a compact translation model that may be installed on a terminal on the basis of a pre-built reference model, in which a pre-built reference model is miniaturized through a parameter imitation learning and is efficiently compressed through a tree search structure imitation learning without degrading the translation performance. The compact translation model provides translation accuracy and speed in a terminal environment that is limited in network, memory, and computation performance.