关键词: 隐喻检测, 多重不一致性网络, 语言学方法, 大语言模型, LoRA（Low-Rank Adaptation）微调

Abstract: A metaphor detection model based on linguistic multi-incongruity was proposed to tackle the problem of metaphor occurrence caused by the incongruity between the target sentence meaning and the basic semantic meaning of the target word in a specific context where a target word has multiple semantic meanings (Polysemy), which has been ignored by the existing metaphor detection research. The model was structured with three main components. First, two separate encoders were employed in the feature encoding module to capture the target sentence meaning, the basic semantics of the target word, and its contextual semantics. Then, three linguistic methods - Selectional Preference Violation (SPV), Metaphor Identification Procedure (MIP), and Semantics Usage Comparison (SUC) - were integrated in the multi-incongruity modeling module to systematically model incongruity features. Finally, metaphor detection was performed through a dedicated recognition module. Furthermore, to validate Chinese metaphor detection performance, a Chinese word-level metaphor detection dataset named META-ZH was constructed through LoRA (Low-Rank Adaptation) fine-tuning of large language models combined with manual error correction in data annotation. The proposed model is demonstrated to achieve significant F1 value improvements of 0.8, 1.3, 1.5, and 2.3 percentage points over baseline systems such as QMM (Quantum-inspired Match for Metaphor) on the VUA-All, VUA-Verb, MOH-X, and META-ZH metaphor detection datasets respectively. Enhanced performance in metaphor detection is attributed to the comprehensive utilization of linguistic multi-incongruity.

Key words: metaphor detection, multi-incongruity network, linguistic methods, large language models, LoRA (Low-Rank Adaptation), fine-tuning