Speakers
Description
This study presents a graph-based workflow for extending the lexicographic description of Japanese adjectives with large language model (LLM) relation candidates and expert review. JMdict provides the principal dictionary evidence. Lexical-item nodes carry written form, reading, grammatical category, English glosses, learner information, and provenance; relations record a proposed type together with semantic, register, domain, scalar, provenance, and review information. Written form plus hiragana reading serves as an operational lookup key that distinguishes homographs such as 辛い (karai ‘spicy’) and 辛い (tsurai ‘painful’). The pilot stored 1,960 proposal observations from gpt-4.1-mini and 2,890 from gpt-5.4-mini across 400 source records representing 305 distinct items; the latter total includes a 101-observation preliminary run. An exploratory run added 76 observations. Three linguists supplied 420 judgments on 214 proposals: 314 judgments combined acceptance with an overall score of at least four, and 180 proposals received at least one such judgment. The results describe the reviewed subset and establish the feasibility of preserving dictionary evidence, model observations, target-resolution outcomes, and expert judgments as distinct components of a lexicographically interpretable resource.