DIINA-CodeSwitching-

Official implementation of DIINA: A neural architecture with Dynamic Inhibitory Regulation for stabilizing multilingual code-switching and reducing lexical interference. ‌ markdown

DIINA: Dynamic Inhibitory Interface for Neural Architectures

‌ Official implementation of DIINA: A neural architecture with Dynamic Inhibitory Regulation for stabilizing multilingual code-switching and reducing lexical interference. ‌

‌

🌟 Overview

DIINA is a specialized neural architecture designed to address the challenges of Code-Switching and Semantic Leakage in multilingual models. Unlike standard transformers, DIINA introduces a biological-inspired inhibitory mechanism that adaptively suppresses non-target language activation. ‌

Key Innovations:

• Dynamic Inhibitory Regulator (DIR): Estimates the degree of non-target activation.
• Inhibition-Augmented Attention: Modulates attention scores to ensure target-language stability.
• Treasure Corpus: A curated multilingual dataset (EN, FA, TR, AZ) for evaluating complex switching patterns. ‌

---

‌

📐 Mathematical Foundation (Appendix C)

‌ The core of DIINA lies in its modified attention mechanism: ‌

1. Inhibitory Signal Calculation

The DIR module computes the inhibitory strength ($w_{\mathrm{inh}}$) based on hidden states ($H$) and contextual cues ($C$): $$w_{\mathrm{inh}} = \sigma(W_{\mathrm{inh}}[H; C] + b_{\mathrm{inh}})$$ ‌

2. Augmented Attention

This signal is integrated into the attention mechanism as an inhibitory bias ($B_{\mathrm{inh}}$): $$\mathrm{Att}{\mathrm{inh}}(Q, K, V, w{\mathrm{inh}}) = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}} + B_{\mathrm{inh}}\right)V$$ ‌

‌

📊 Evaluation & Results

DIINA shows significant robustness in Intra-sentential Code-Switching, where lexical competition is strongest. ‌

Error Profile Mitigation:

• ✅ Reduces unintended lexical substitutions.
• ✅ Fixes delayed adaptation after switch points.
• ✅ Eliminates mixed-language phrases that reduce semantic precision. ‌

---

‌

🛠 Project Structure

• models/: PyTorch implementation of DIR and Attention layers.
• data/: Sample switching patterns from the Treasure Corpus.
• paper/: Supplementary materials and architectural details. ‌

---

‌

✍️ Citation

If you use DIINA in your research, please cite:bibtex @article{pegah2026diina, title={DIINA: Dynamic Inhibitory Interface for Neural Architectures}, author={Dr. Pegah}, year={2026} }