LLM Text Rewriting Using Graph Prompts and Template Structure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text generation technologies struggle to incorporate the meaning and structure of multiple independent texts while maintaining the original text's meaning and structure, often resulting in complex, unnatural, or redundant outputs due to the complexity of graph structures and the need for extensive prompt engineering and fine-tuning.
Innovation Solution
A text generation device that utilizes a large-scale language model (LLM) trained with triplet and similar pair data to generate a rewritten text by extracting similar graph data from template and reference texts, using a hint sentence to maintain the structure of the template text and incorporate the meaning of the reference text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing text generation technologies are used to incorporate multiple independent texts, then the meaning and structure of the texts can be integrated, but the output becomes complex, unnatural, or redundant
Solution Approach 1:
The patent segments the text generation process into distinct components: graph data generation from input texts, prompt construction with segmentation instructions, and LLM-based text generation. By dividing the complex task of incorporating multiple texts into manageable segments with clear structural guidelines, the system maintains naturalness while achieving comprehensive text integration.
Solution Approach 2:
The patent introduces graph data as an intermediary representation that captures the semantic relationships between noun phrases from multiple input texts. This graph structure serves as a mediator between the raw input texts and the final generated text, enabling the LLM to understand and incorporate multiple sources naturally without direct concatenation or complex fusion operations.
2Productivity
If graph data and extensive prompt engineering are used to generate text from multiple sources, then text generation capability is improved, but the process complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-generating graph data from the input texts before the main text generation step. The graph data pre-processes and structures the semantic information from multiple sources, so that when the LLM generates the final text, it receives organized relationship information rather than raw unprocessed texts, reducing the complexity of prompt engineering needed.
Solution Approach 2:
The patent changes the parameter representation of input texts from raw text format to graph data format with structured noun phrases and relationships. This parameter transformation enables the system to handle multiple texts more efficiently by converting unstructured information into a standardized graphical representation that the LLM can process with simpler prompts.
3Loss of information
If multiple texts are integrated to generate comprehensive output, then information completeness is improved, but redundancy and unnatural content increase
Solution Approach 1:
The patent implements feedback mechanisms through the graph data structure that captures relationships between noun phrases from multiple input texts. The graph serves as a feedback loop that identifies overlapping information and relationships, allowing the system to integrate comprehensive information from all sources while avoiding redundant content in the final generated text by leveraging the structured relationship data.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer is caused to execute processing of acquiring a template text serving as a norm and a reference text, generating graph data including noun phrases included in the reference text and information about a relation between the noun phrases in the reference text, on the basis of the reference text, and inputting a prompt including the generated graph data and the template text to an LLM to generate a rewritten text satisfying a requirement defined in the template text.