Detecting the strange: automatic recognition and evaluation of defamiliarization in modernist poetry

This study addresses the long-standing gap between literary theory and computational modeling by focusing on defamiliarization, a central technique in modernist poetics. While defamiliarization has been extensively theorized in literary studies, its computational treatment remains limited due to its semantic complexity and subjective nature. To operationalize this phenomenon, the research introduces MelPoet, a poetics-informed neural architecture adapted from MelBERT. Two tasks are formulated: defamiliarization identification as binary classification, and defamiliarization scoring as regression of estrangement intensity. A curated dataset of modernist poetry and control texts was annotated for both presence and degree of defamiliarization. Experimental results demonstrate that MelPoet substantially outperforms strong baselines in both classification accuracy and scoring correlation, confirming the efficacy of its theory-driven design. This work not only advances computational methods for modelling figurative language but also provides a systematic framework for integrating literary concepts into natural language processing, thereby opening new avenues for large-scale, data-driven analysis of poetic style.