LLM Text Rewriting Using Graph Prompts and Similar Triplets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text generation technologies struggle to incorporate the meaning and structure of an independent text while maintaining the original text's meaning and structure, often leading to complex, unnatural, or redundant outputs, and require significant time and effort in prompt engineering and fine tuning.
Innovation Solution
A text generation device utilizing a large-scale language model (LLM) trained with triplet and similar pair data to generate a rewritten text by extracting similar triplets from template and reference texts, incorporating a hint sentence generation process to maintain the original text's structure and meaning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing text generation technologies are used to incorporate meaning and structure from independent texts, then the output may reflect reference text meaning, but the generation becomes complex, unnatural, or redundant and requires significant time and effort in prompt engineering and fine tuning
Solution Approach 1:
The patent segments the text generation process into distinct stages: extracting triplets from the reference text, filtering similar triplets, generating rewritten sentences only for necessary portions, and assembling the final output. This segmentation reduces the complexity of handling entire texts by processing them in manageable units (triplets and sentences), thereby improving reliability while reducing the burden of prompt engineering and fine-tuning.
Solution Approach 2:
The patent extracts meaningful units (triples) from the reference text and separates them from the template text structure. By extracting only the essential semantic information (subject-predicate-object relationships) and filtering for similarity, the system can incorporate reference meaning without requiring complex processing of the entire reference text, thus reducing generation complexity while maintaining accuracy.
2Stability of the object's composition
If existing text generation technologies are used to maintain original text structure and meaning, then the output structure is preserved, but the generation becomes complex and requires extensive prompt engineering and fine tuning
Solution Approach 1:
The patent performs preliminary extraction and filtering of triplets from the reference text before the actual text generation process. By pre-processing the reference text to identify and filter similar triplets in advance, the system prepares reusable semantic units that can be directly inserted into the template structure, thereby preserving original structure while significantly reducing the time needed for prompt engineering and fine-tuning.
Solution Approach 2:
The patent creates rewritten sentences by copying and adapting triplets from the reference text into the template text structure. Instead of generating entirely new content or requiring complex transformations, the system copies semantic units (triples) and positions them within the template framework, maintaining structural integrity while incorporating reference meaning with minimal prompt engineering and fine-tuning requirements.
3Manufacturing precision
If existing text generation technologies are used to generate rewritten text, then the output may satisfy requirements, but redundancy and complexity in the output increase
Solution Approach 1:
The patent extracts only the necessary semantic information (triples) from the reference text and filters out redundant or unrelated content. By extracting and selecting only the most relevant triplets that satisfy the generation requirements, the system produces concise rewritten text without redundancy, achieving high precision while minimizing unnecessary content in the output.
Solution Approach 2:
The patent applies partial action by generating rewritten sentences only for portions of the template text that require modification based on the reference text. Instead of rewriting the entire template, the system identifies specific sections needing updates and applies changes only there, thereby satisfying requirements while avoiding redundant rewriting of unchanged portions and reducing overall text complexity.
Data Source
AI summary
A non-transitory computer-readable recording medium stores therein a text generation program that causes a computer to execute a process including acquiring a first text serving as a norm and a second text related to a case example, first generating graph data of the second text including noun phrases included in the second text and information about a relation between the noun phrases in the second text, based on the second text, and first inputting a prompt including the graph data of the second text generated, and the first text, to a large-scale language model to generate a third text satisfying a requirement defined in the first text.