As Natural Language Processing (NLP) models grow more complex, explaining their predictions becomes increasingly challenging—particularly for tasks involving structured outputs or multilingual input. This paper introduces adaptations of three explainability methods; SHAP, LIME and Integrated Gradients; to enhance interpretability for Sentiment Analysis and Named Entity Recognition on Arabic–English code-switched data. To the best of our knowledge, this is the first work to systematically adapt these techniques for structured prediction in this language setting. We propose class-specific filtering for SHAP, a language-aware perturbation strategy for LIME, and a refined token-level IG approach (X-DIGG) for more faithful attributions. To improve accessibility, we present Explainify, an interactive tool that combines token-level visualizations with natural language summaries. A human-grounded user study and quantitative analysis demonstrate that our adaptations significantly improve interpretability across tasks.