Structured Document Conversion Using Context-Aware LLM Fragments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques face challenges in converting unstructured document data into structured data accurately, often leading to information loss and unstable structuring due to summarization or fragmentation, particularly with large language models like ChatGPT.
Innovation Solution
A method involving dividing document data into fragments, creating conversion request sentences that include preceding and subsequent fragments, and inputting these sentences to a large language model to ensure structured conversion, guided by local and global context information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If document data is summarized to satisfy token length constraints, then the document can be processed by LLM, but information important for structuring is lost and conversion accuracy is impaired
Solution Approach 1:
The patent divides the document into multiple fragments and processes them sequentially through the LLM. Each fragment is converted individually with context from previous fragments, avoiding the need to summarize the entire document while maintaining structural information. This segmentation approach allows the full document to be processed without token constraints while preserving all structuring-relevant information.
2Quantity of substance
If document data is divided into fragments for LLM processing, then token constraints are satisfied, but each fragment may lack information involved in structuring and structuring becomes unstable
Solution Approach 1:
The patent performs preliminary processing by creating context information from previously converted fragments before processing each new fragment. This context includes structuring information from earlier parts of the document, which is provided to the LLM to ensure stable and consistent structuring decisions. The context is prepared in advance and included in each conversion request, preventing structuring instability.
3Loss of information
If the entire document is input to LLM at once, then complete context is available for structuring, but the document exceeds token length constraints and cannot be processed
Solution Approach 1:
The patent transforms the problem from a single-dimension approach (processing the entire document at once) to a multi-dimensional approach by adding the dimension of sequential processing with context accumulation. Documents are divided into fragments processed in sequence, with context from previous fragments carried forward. This dimensional change allows complete context to be available through accumulation rather than requiring simultaneous input of the entire document.
4Productivity
If document fragments are processed independently, then processing efficiency is improved, but conversion accuracy decreases due to lack of context
Solution Approach 1:
The patent ensures continuity of useful action by maintaining and transmitting context information across fragment processing boundaries. Each fragment processing operation continues the structuring task from where previous fragments left off, with context from converted fragments carried forward. This continuous context transmission maintains conversion accuracy while preserving the efficiency benefits of fragment-based parallel processing.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing program causes a computer to execute a process including: dividing document data intended to be converted, into a plurality of document fragments; when requesting a large language model to convert each of the divided document fragments into a document of which a document structure is specified, creating a conversion request sentence that includes the document fragments immediately preceding the document fragments requested to be converted, the document fragments requested to be converted, and the document fragments after conversion obtained by the large language model by converting the document fragments immediately preceding the document fragments requested to be converted; and inputting the created conversion request sentence to the large language model to execute the conversion into the document fragments of which the document structure is specified.