Hybrid Deep Learning and Contextual Features for Figurative Language Detection in Product Reviews
International journal of intelligent engineering and systems · 2025
This study introduces a novel hybrid framework that uniquely integrates culturally tailored linguistic features with a multi-scale convolutional neural network (CNN) to address the underexplored challenge of figurative language detection in low-resource Indonesian e-commerce contexts.Unlike prior works focusing on English or social media, our method pioneers three innovations such as Culture-specific lexicons for Indonesian sarcasm markers (e.g., hyperbolic honorifics, temporal incongruities), iteratively refined via corpus analysis of 10,000 sarcastic phrases, Hierarchical feature fusion combining 25+ interpretable linguistic rules (e.g., oxymoron patterns, self-referential pronouns) with CNN-extracted semantic embeddings, enabling context-aware detection of implicit figurative cues, and Resource-efficient ensemble learning via AutoGluon, reducing computational costs by 63% compared to transformer baselines while maintaining interpretability.Evaluated on 147,636 annotated Indonesian reviews, the model achieves 89.2% accuracy and 81.5% F1-score for sarcasm, outperforming state-of-the-art methods (CNN: +20.3%, BERT: +4.1%) and demonstrating 41% reduction in manual moderation efforts in a real-world Tokopedia deployment.This work bridges critical gaps in low-resource NLP, offering a scalable blueprint for linguistically complex markets.