Short Abstract

We present an adjective-centered lexical network for aiLanguageTutor, a Japanese learning platform. The network encodes 783 adjectives with synonymy, antonymy, and gradability relations, plus morpho-syntactic properties, supporting adaptive generation of teaching materials.

Long Abstract

The paper reports on the design of an adjective-centered lexical network being integrated into aiLanguageTutor, a Japanese language learning platform combining structured linguistic knowledge with adaptive generation of teaching materials. The network currently encompasses 783 adjectives (277 i-adjectives and 506 na-adjectives) drawn from the Japanese Language Education Vocabulary List (Sunakawa et al. 2012) connected through an estimated 2,500+ semantic and morpho-syntactic relations.

The network models adjectives as central nodes and encodes relations among adjectives only, capturing semantic similarity and opposition as well as systematic variation in degree and evaluation. The lexical graph, following previously developed methodology (Perak and Špica 2025), links adjectives via synonymy/near-synonymy (kirei–utsukushii), antonymy (atsui–samui), and gradability/intensity relations representing evaluative scaling and strength differences (kowai – osoroshii; oishii–zeppin da). Domain association is represented as labels on adjective nodes (e.g., travel, food, weather, service encounters), enabling topic-appropriate adjective selection without relying on noun–adjective pairings.

In parallel to semantic structure, the network encodes morpho-syntactic properties central to Japanese instruction and error diagnosis: adjective class (i-adjectives vs. na-adjectives: samui vs. shizuka da), conjugation patterns (samukatta/samukunai; shizuka datta/shizuka ja nai), and polarity as evaluative orientation with its interaction with negation.

For synonym relations, the network stores structured contrastive information for explainable paraphrase and choice guidance: shared core meaning, domain/register tendencies, context-driven contrasts, typical usage profiles, and explicit selection guidance including formality level and spoken/written preferences.

Within aiLanguageTutor, this adjective network supports level-appropriate sequencing from high-frequency cores to nuanced alternatives, contrastive practice and error diagnosis based on local neighborhoods (e.g., kirei vs. seiketsu; samui vs. suzushii), and controlled generation of dialogues and texts maintaining lexical coherence within chosen topics and proficiency levels.

Sunakawa, Yuriko, Lee, Jae-ho, and Takahara, Mari (2012) The Construction of a Database to Support the Compilation of Japanese Learners Dictionaries, Acta Linguistica Asiatica 2(2), pp.97-115

Perak, B., & Špica, D. (2025). Automating Lexical Graph Construction with Large Language Models: A Scalable Approach to Japanese Multi-Relation Lexical Networks. Knowledge, 5(4), 24. https://doi.org/10.3390/knowledge5040024

Keywords: lexical network, adjectives, synonymy, lexical relations