Journal

Volume 27, Issue 3 (October 31, 2026)

3 articles

• Response Patterns to Prosodically Disambiguated Questions Containing Wh-Words in Korean
  by Jeonghwa Cho, Jina Song and Harim Kwon
  J. CS. 2026, 27(3), 311-334;
  Abstract Korean wh-words are ambiguous between an existential quantifier reading and a wh-interrogative reading. Thus, questions containing wh-words can be interpreted as either a yes/no-question or a wh-question depending on their prosody. However, previous studies have primarily relied on forced-choice tas... [Read more].
  Abstract Korean wh-words are ambiguous between an existential quantifier reading and a wh-interrogative reading. Thus, questions containing wh-words can be interpreted as either a yes/no-question or a wh-question depending on their prosody. However, previous studies have primarily relied on forced-choice tasks, and thus little is known about whether listeners consider multiple response types appropriate for each question type. The current study aimed to address this gap by allowing participants to select more than one response. Thirty-four Korean native speakers listened to questions containing wh-words that are spoken with prosody typical for either yes/no-questions or wh-questions and selected appropriate answers between yes/no-answer type and wh-answer type. Importantly, they were allowed to select both answers when applicable. Results show a clear asymmetry when the questions were spoken with yes/no-question prosody and wh-question prosody: participants most frequently selected the wh-answer for wh-question prosody, but most frequently selected both answers for yes/no-question prosody. This indicates that for yes/no-questions, Korean speakers allow wh-answers in addition to yes/no-answers. These findings suggest that conventional forced-choice paradigms may underestimate the range of responses that Korean speakers consider for prosodically disambiguated questions. The asymmetry can be interpreted in terms of the entailment relation between the two question types and the communicative goals of the interlocutors. [Collapse]
• Similar Ratings, Different Mechanisms: Korean Inanimate Honorification as a Benchmark for Pragmatic Competence in Large Language Models
  by Youngin Lee, Haayan Jang, Seohyun Jung and Sung-Eun Lee
  J. CS. 2026, 27(3), 335-378;
  Abstract Large language models (LLMs) have exhibited remarkable capabilities across a wide range of language tasks. However, existing evaluation benchmarks remain predominantly English-centric and rarely assess pragmatic factors in a systematic manner. This study investigates LLMs’ judgments of inanimate hon... [Read more].
  Abstract Large language models (LLMs) have exhibited remarkable capabilities across a wide range of language tasks. However, existing evaluation benchmarks remain predominantly English-centric and rarely assess pragmatic factors in a systematic manner. This study investigates LLMs’ judgments of inanimate honorification, a pervasive phenomenon in contemporary Korean. We introduce a novel benchmark that systematically manipulates linguistic, possession, and spatial factors, and compare model-generated acceptability ratings with human judgments. Human results revealed that the use of the honorific suffix -si- with an inanimate subject depends on two factors: whether the referent of the subject is possessed by or otherwise associated with the honored referent, and its physical proximity to the honored referent. These findings support accounts that view honorification as a pragmatic phenomenon rather than strict grammatical agreement. While several models showed substantial alignment with human judgments, none fully matched human rating patterns. Moreover, even when overall rating patterns appeared similar, models varied considerably in their sensitivity to possession and spatial factors, suggesting differences in the mechanisms underlying their evaluations. These findings demonstrate that similarity in acceptability ratings does not necessarily reflect human-like pragmatic competence. By incorporating socially and culturally grounded variables, our benchmark provides a mechanism-level evaluation of pragmatic processing and expands LLM assessment beyond predominantly English-centric benchmarks. [Collapse]
• Cross-Validated Computational Cognitive Modeling of Behavioral Visual-Field Effects in Korean Eojeol Recognition with Implications for Human-Informed AI
  by Sangyub Kim
  J. CS. 2026, 27(3), 379-412;
  Abstract The primary aim of the present study was to construct and cross-validate a functional cognitive-processing architecture from human trial-level data on Korean Eojeol recognition. The architecture separated shared lexical difficulty (D), a behavioral visual-field asymmetry adjustment (A), and a bilate... [Read more].
  Abstract The primary aim of the present study was to construct and cross-validate a functional cognitive-processing architecture from human trial-level data on Korean Eojeol recognition. The architecture separated shared lexical difficulty (D), a behavioral visual-field asymmetry adjustment (A), and a bilateral-presentation adjustment (B), and tested whether A and B improved prediction for unseen participants and items. We conducted a secondary analysis of 10,800 word trials from 36 participants. In participant-plus-item cross-validation, adding A improved reaction time (RT) prediction (R²=.015 to .034; mean absolute error=131.36 to 129.70 ms) and accuracy prediction (log loss=.461 to .429). Adding B further improved accuracy (log loss=.423) but not RT (R²=.034 to .034), and item modulation improved neither outcome. Complementary item-level analyses used repeated nested cross-validation to compare psycholinguistic, character n-gram language form, structural Hangul visual form, and combined representations. Psycholinguistic variables gave the clearest predictions for several outcomes. Under the present framework and dataset, adding the specific language-form and visual-form representations did not consistently improve out-of-sample prediction over the training-mean baseline. Split-half reliability was poor for item-level right visual field advantage (RVFA) and bilateral redundancy gain (BRG), with median Spearman-Brown estimates of −.021 for RT RVFA, .168 for accuracy RVFA, .019 for RT BRG, and −.068 for accuracy BRG. These estimates limited item-specific inference while trial-level modeling tested population-level components. The findings provide partial out-of-sample support for a behavioral architecture in which A contributed to response speed and accuracy, whereas B contributed only to accuracy. The architecture provides a testable behavioral account and a potential human-data constraint for future AI models. [Collapse]