SynergyScore: Leveraging Pretrained Language Models for Effective Evaluation of Code-Switched Sentences

Code-switching, a common phenomenon in multilingual communication, presents significant challenges for linguistic analysis, particularly in evaluating semantic and contextual nuances such as meaning, fluency, and factuality. Existing metrics, including the Code-Mixing Index (CMI), Multilingual Index (M-index), and Language Entropy (LE), primarily quantify surface-level language switching characteristics but fail to address deeper linguistic dimensions. This paper introduces SynergyScore, a novel composite metric designed to reflect human-like evaluations of code-switched text by integrating traditional measures with insights from pretrained language models (PLMs). SynergyScore provides a holistic evaluation framework by combining structural analysis with semantic and contextual assessments. Comparative experiments demonstrate its superiority over traditional metrics, achieving stronger alignment with human judgments across dimensions such as fluency, coherence, and meaning. This study highlights SynergyScore’s potential to enhance code-switching evaluation and sets a foundation for further advancements in multilingual natural language processing.