Document Encoding System with Reasoning Retriever
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning models for natural language processing are computationally heavy and inefficient, requiring extensive resources and time to extract multiple pieces of information from documents, making them impractical for industry use cases.
Innovation Solution
A system comprising an encoder with a neural network and a retriever configured as a reasoning engine, which supports user-defined queries and allows multiple encoders to process documents in parallel, reducing computational resources and speeding up analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a sliding window approach is used to extract information from documents, then the model can analyse text in manageable segments, but the processing time increases significantly when multiple tasks need to be completed
Solution Approach 1:
The patent applies preliminary action by encoding the entire document once into a document embedding before any information extraction tasks are performed. This pre-computed embedding serves as a reusable representation that can be queried multiple times for different tasks, eliminating the need to re-process the document for each task and significantly reducing processing time.
Solution Approach 2:
The document embedding serves multiple functions: it can be used to answer various types of questions about the document, extract different types of information, and support multiple tasks simultaneously. This universal representation allows the system to handle diverse information extraction tasks without requiring separate processing pipelines for each task type.
2Productivity
If multiple encoders are used to process different tasks in parallel, then processing speed improves, but computational resources and memory requirements increase significantly
Solution Approach 1:
The patent extracts the computationally intensive document encoding operation as a separate, one-time preliminary step. By separating the document encoding from the task-specific querying operations, the system eliminates redundant computations while maintaining parallel processing capabilities for multiple tasks, thus reducing overall computational resource requirements.
Solution Approach 2:
Instead of creating multiple encoder instances or re-processing the document for each task, the system creates a single copy of the document embedding that can be reused across multiple tasks. This copying approach allows parallel task processing without duplicating the expensive encoding computations, optimizing the use of computational resources.
3Reliability
If the model re-runs completely from scratch for each window and task, then accurate predictions can be made, but the computational overhead becomes impractical for real-world use
Solution Approach 1:
The document embedding is computed as a preliminary action before any task-specific queries are processed. This pre-computed embedding captures the essential information from the document in a compressed form, allowing subsequent tasks to be performed by querying this embedding rather than re-processing the original document, thus maintaining accuracy while improving efficiency.
Solution Approach 2:
The patent merges the document understanding and information extraction processes into a unified framework where the document embedding serves as a shared representation for all tasks. This merging eliminates redundant processing steps and allows the system to leverage the same computed features across multiple tasks, improving both efficiency and consistency.
Data Source
AI summary
The invention pertains to a system for the extraction of information from documents, in particular natural language documents, the system comprising an encoder with a neural network; and a retriever that is configured as a reasoning engine. The system is configured such that it supports user-defined queries for at least two pieces of information; the encoder is applied to the documents, in particular the natural language documents, to generate document encodings; the user-defined queries, to generate encoded instructions; the document encodings are queried by the retriever in lookup steps based on the encoded instructions.