Cross-lingual Metaphor Recognition Algorithm Integrating GCN with Phenomenological Semantic Space
Year: 2026
Type: conference-paper
Abstract: Metaphor recognition remains a challenging task in natural language processing, particularly for cross-lingual scenarios where training data in target languages is scarce. Existing approaches rely heavily on distributional semantics learned from large-scale corpora, which capture language-specific patterns but fail to generalize across typologically diverse languages. In this paper, we propose a novel framework that integrates phenomenological semantic features—cognitively grounded dimensions such as concreteness, sensory modality, affective polarity, animacy, and dynamicity—with graph neural networks for cross-lingual metaphor recognition. Our approach constructs heterogeneous graphs that encode both syntactic dependencies and phenomenological similarities, enabling structured reasoning over metaphorical mappings. We hypothesize that phenomenological features represent universal cognitive primitives that remain invariant across languages, facilitating zero-shot and few-shot transfer. Extensive experiments on English and Chinese metaphor datasets demonstrate that our model achieves 90.9% of fully-supervised performance in zero-shot cross-lingual transfer, significantly outperforming multilingual BERT (83.2%) and XLM-R (87.5%). In few-shot scenarios, our approach requires only 100 Chinese training samples to match the performance of baseline models trained on 5,000 samples, demonstrating 50× sample efficiency. Ablation studies confirm that phenomenological features contribute 5.2% F1-score improvement, while the graph structure adds 4.7%, validating the synergy between cognitive semantics and structured neural architectures. Attention weight analysis reveals that our model correctly identifies source-target domain mappings by recognizing concrete-to-abstract transitions in phenomenological space. These findings suggest that incorporating cognitively motivated semantic representations offers a promising direction for building language-agnostic NLP systems. Our work bridges cognitive linguistics and deep learning, demonstrating that universal properties of human conceptualization can be operationalized to overcome data scarcity in multilingual metaphor processing.
Authors: HuiJie Zhou, Mohan Wu, Kunyuan Pan, Ye Tian
Institution: Guangdong University of Science and Technology
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204929653