Presentation Information
[3Yin-A-44]An LLM-Based Method for Automated Correction of Terminology Inconsistency in Design Documents Using a Term Dictionary
〇Kazu Nishikawa1, Yuta Koreeda1, Yasuhide Mori1, Shun Oida1, Daisuke Fukui1 (1. Hitachi)
Keywords:
AI Agent,Software Engineering,Document Proofreading
This paper presents an automated proofreading workflow for correcting terminology inconsistencies in Japanese
design documents by integrating project-specific term dictionaries with large language models (LLMs). The dictionary is built by extracting candidate terms and prompting an LLM to craft concise, context-aware definitions,
capturing proprietary abbreviations and jargon. The workflow then applies prompts tailored to six inconsistency
categories before rewriting affected passages. Evaluation on 30 financial-system design documents across three
GPT-4.1 model sizes shows that adding the dictionary to category-wise prompting raises precision from 0.325 to
0.432, highlighting the value of decomposition and explicit terminology grounding.
design documents by integrating project-specific term dictionaries with large language models (LLMs). The dictionary is built by extracting candidate terms and prompting an LLM to craft concise, context-aware definitions,
capturing proprietary abbreviations and jargon. The workflow then applies prompts tailored to six inconsistency
categories before rewriting affected passages. Evaluation on 30 financial-system design documents across three
GPT-4.1 model sizes shows that adding the dictionary to category-wise prompting raises precision from 0.325 to
0.432, highlighting the value of decomposition and explicit terminology grounding.