Domain Specific Document Representation for GPU Accelerated NLP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Information Retrieval (IR) systems incur high CPU cycles and are inefficient for GPU architectures due to processing tokens at the character level, leading to expensive data movement and suboptimal performance in finding patterns within documents.
Innovation Solution
A domain-specific representation of document text is converted into a token stream, then into an integer domain representation using a domain-specific ontology dictionary, which is stored on a GPU processing queue for accelerated natural language processing, enabling efficient GPU processing and reducing irrelevant token processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If tokens are processed at the character level in traditional IR systems, then pattern matching can be performed, but CPU cycles are consumed and data movement between CPU and GPU is expensive
Solution Approach 1:
The patent extracts only the relevant tokens from the character stream based on tokenization rules, removing irrelevant tokens before processing. This extraction principle reduces the amount of data that needs to be processed and transferred, thereby reducing CPU cycle consumption and data movement costs while maintaining pattern matching capability.
Solution Approach 2:
The patent converts tokens from their original character-level representation into an integer domain representation using a domain-specific ontology dictionary. This parameter change from character level to integer level enables more efficient processing and reduces the computational complexity of pattern matching operations.
2Productivity
If character-level token processing is used, then comprehensive text analysis is possible, but GPU processing efficiency is reduced due to conservative data movement requirements
Solution Approach 1:
The patent performs preliminary tokenization and filtering of the character stream into a token stream before transferring data to the GPU. By pre-processing the data on the CPU side to extract only relevant tokens and convert them to integer representation, the system reduces the complexity of data movement and prepares the data in a format optimized for GPU processing.
Solution Approach 2:
The patent introduces an intermediary integer domain representation layer between the character stream and GPU processing. This intermediary representation serves as a bridge that simplifies data movement and enables efficient GPU processing by converting variable-length character sequences into fixed-width integer values that are more suitable for parallel processing.
3Reliability
If all tokens in the token stream are processed, then complete document coverage is achieved, but irrelevant tokens increase processing overhead
Solution Approach 1:
The patent applies extraction by removing irrelevant tokens from the token stream based on domain-specific criteria while retaining relevant tokens. This selective extraction maintains complete coverage of meaningful document content while eliminating noise that would otherwise increase processing time and overhead.
Data Source
AI summary
Provided are techniques for a domain specific representation of document text for accelerated natural language processing. A document is selected from a set of documents to be analyzed. A character stream from the document is converted into a token stream based on tokenization rules. Irrelevant tokens are removed from the token stream. The tokens remaining in the token stream are converted into an integer domain representation based on a domain specific ontology dictionary. The integer domain representation are stored to a Graphics Processing Unit (GPU) processing queue of each of one or more GPUs. Then, a result set is received from the one or more GPUs.