Title: Decoding Meaning: A Detailed Examination of Hybrid-MoE’s Impact on Idiomatic Comprehension in Language Models

Abstract:

In today’s era of multilingual education, the study of idioms offers a compelling entry point into the creativity, cultural heritage, historical backgrounds, and varied viewpoints embedded within different linguistic traditions. This study addresses the challenge of preserving figurative and cultural semantics in low-resource Southeast Asian languages, specifically Hindi, Bengali, and Thai. In these languages, culturally dense idioms create substantial hurdles for computational modeling and cross-linguistic transfer owing to their intricate metaphorical structures.

To overcome these complexities, we introduce Varnika, a rebuilt multimodal corpus containing 3,533 multilingual idioms. This dataset is enhanced with seven distinct idiomatic tones, supported by both textual and visual data. Furthermore, to facilitate a deeper inference of idiomatic understanding, we propose the Hybrid Mixture-of-Experts (HybridMoE) framework. This approach incorporates multiple expert opinions on idioms while reducing expert sparsity. It does so by combining outputs from both selected and unselected experts via controlled hybridization, and is further strengthened by Idiomatic Property Signals derived from masked multimodal embeddings.

For a comprehensive performance analysis, we developed the IDIO-TONE and Idiomatic Validation Score, a three-phase evaluation pipeline. This framework assesses (i) the fidelity of literal translations, (ii) the alignment between visual and semantic elements, and (iii) the preservation of idiomatic meaning. Our empirical results indicate that HybridMoE delivers a 5–6% improvement in performance across advanced vision-language models. These findings demonstrate enhanced representation of figurative language and culturally rooted meanings within multilingual, multimodal contexts.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC