An Attempt at Japanese Natural Language Lint Using a Local Large Language Model

— Implementation of a Prompt-Engineering-Independent Natural Language Lint System —

Abstract

This study proposes JapaneseLint, a natural language lint system designed to detect ambiguity in Japanese specifications and to automatically generate quantitatively constrained revision proposals. JapaneseLint adopts a lightweight fine-tuning architecture based on LoRA-based distillation, in which a task-specific reasoning style for specification review is embedded directly into a compact model using a general-purpose large language model (GPT-5.1) as a teacher. This design enables stable Lint-style inference without reliance on extensive prompt engineering, while minimizing context consumption at inference time.
For evaluation, twenty test specifications automatically generated by GPT-5.1 were used as inputs, and both inference time and textual output quality were assessed. The experimental results show that the average inference time was 1.803 s, with a maximum of 2.618 s, satisfying practical response-time requirements commonly reported in the HCI literature. Moreover, the generated diagnostic comments and revision proposals exhibited sufficient readability, objectivity, and quantitative rigor for practical use in real-world specification review tasks.
These results demonstrate that JapaneseLint can provide fast and practically usable support for ambiguity detection and quantitative revision of natural-language specifications under realistic computational constraints.