Evaluating Symmetry and Asymmetry in Large Language Models’ Focus-Span Identification: Evidence from Chinese shi…de Cleft Constructions

Large language models (LLMs) have achieved strong performance in many linguistic tasks, but their ability to process discourse-level information structure remains insufficiently understood. In particular, current models may identify locally coherent spans while failing to determine the minimal constituent that carries informational prominence in context. Chinese “shi…de” cleft constructions provide a theoretically important testing ground for this problem because they combine a stable formal pattern with context-dependent focus interpretation, exhaustivity, and discourse-sensitive boundary variation. This study investigates whether current LLMs can identify the minimal focus domain in Chinese “shi…de” clefts and whether their performance goes beyond simple surface-form heuristics. Based on 105 human-validated gold-standard samples, we compared three API-accessible models, ChatGPT 5.4, Claude Opus 4.6, and DeepSeek-V4-Pro, with two rule-based baselines. Baseline 1, which extracted the full span between “shi” and “de”, achieved only 2.86% accuracy, while Baseline 2, a stronger minimal-cue heuristic, reached 46.67%. Under the main prompt condition, DeepSeek-V4-Pro achieved the highest accuracy (65.71%), followed by Claude Opus 4.6 (60.00%), whereas ChatGPT 5.4 (41.90%) did not outperform Baseline 2. A prompt-level QUD ablation showed no stable or statistically significant improvement, indicating that explicit discourse-question guidance alone is insufficient for minimal focus-boundary identification. Performance across focus types further showed that topical focus was relatively easier than informational and contrastive focus, suggesting the importance of topic continuity. Overall, the findings reveal both symmetry and asymmetry in LLM focus processing: models share certain task-level constraints, but differ in cue weighting and boundary-compression strategies. The study argues that Chinese “shi…de” focus identification is better modeled as a multi-cue focus-span ranking problem rather than as direct QUD-answer matching. Future research should extend the dataset and further test whether explicit multi-cue ranking methods can improve focus-boundary identification across models and languages.