Compressing Conditional Random Field Models via Hashing and Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Conditional Random Field (CRF) models used in natural language understanding applications, such as named entity recognition, are large in size due to extensive dictionaries and parameter values, leading to significant storage and transmission bandwidth requirements, and existing compression techniques do not provide usable models without decompression.
Innovation Solution
The approach involves reducing the size of CRF models by eliminating or reducing input dictionaries through hashing functions, applying regularization to eliminate insignificant parameters, and quantizing parameter values to lower precision, resulting in a substantially smaller model without compromising accuracy or speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional CRF models use extensive dictionaries and parameter values, then model accuracy is maintained, but storage size and transmission bandwidth requirements increase significantly
Solution Approach 1:
The patent extracts and removes extensive dictionaries from the CRF model structure, retaining only essential parameter values. This extraction principle directly reduces storage size while maintaining the core functionality needed for model accuracy.
Solution Approach 2:
The patent uses compact representations and encoding schemes that create efficient copies or references to parameter data, reducing the actual storage footprint while preserving the information necessary for accurate model operation.
2Reliability
If conventional CRF models use extensive dictionaries and parameter values, then model accuracy is maintained, but transmission bandwidth requirements increase significantly
Solution Approach 1:
By extracting and removing redundant dictionary data from the model structure, the patent significantly reduces the total data volume that must be transmitted, thereby reducing bandwidth consumption while preserving essential model parameters for accuracy.
3Quantity of substance
If compression techniques are applied to CRF models, then storage size is reduced, but the model becomes unusable without decompression
Solution Approach 1:
The patent adopts a disposable approach to compression by using lossless compression algorithms that can be quickly applied and decompressed when needed, accepting the temporary overhead of decompression in exchange for significantly reduced storage size and the ability to maintain model usability.
Data Source
AI summary
Features are disclosed for generating models, such as conditional random field (“CRF”) models, that consume less storage space and/or transmission bandwidth than conventional models. In some embodiments, the generated CRF models are composed of fewer or alternate components in comparison with conventional CRF models. For example, a system generating such CRF models may forgo the use of large dictionaries or other cross-reference lists that map information extracted from input (e.g., “features”) to model parameters; reduce in weight (or exclude altogether) certain model parameters that may not have a significant effect on model accuracy; and/or reduce the numerical precision of model parameters.