Dynamic Multi-Attention Fusion for Joint Intent Detection and Slot Filling in Code-Mixed Language Understanding

Code-mixed languages, characterized by the seamless integration of multiple languages, pose a significant challenge to natural language understanding (NLU) systems. In this work, we propose a novel approach to address the complexities of code-mixed text by leveraging a BiLSTM-BERT architecture enhanced with a multi-attention fusion mechanism. This approach is specifically designed for joint intent classification and slot filling (JIC-SF) tasks in code-mixed language scenarios. Our proposed model incorporates self-attention and cross-attention mechanisms to dynamically prioritize the most relevant parts of the input text for effective JIC-SF. The effectiveness of the proposed model was evaluated on three benchmark datasets: ATIS, SNIPs, and Hind-English Code-Mixed (HiEn-CMD). The results demonstrate its superiority over state-of-the-art models, achieving an intent accuracy of 97.87% and a slot F1-score of 95.97% on the ATIS dataset, an intent accuracy of 98.86% and a slot F1-score of 96.25% on the SNIPs dataset, and an intent accuracy of 84.24% with a slot F1-score of 82.68% on the HiEn-CMD dataset. This BiLSTM-BERT multi-attention model offers a robust solution for improving the accuracy and efficiency of NLU systems. Its applications span a variety of domains, including virtual assistants, chatbots, and customer service platforms, providing enhanced capabilities for processing code-mixed text.