<p>Tackling Cross-Domain Intent Variability With HyTra-ID: A Robust Transformer-Based Framework for Real-World Dialogue Systems</p>

Intent detection is the basis for real-world dialogue systems such as virtual assistants, conversational agents, and automated customer support, enabling these applications to understand the user’s intent. While deploying a single unified classification model across multiple domains is difficult because of the variety of linguistic structures and annotation systems in the numerous datasets (ATIS, SNIPS, and TOPv2) available, the resulting domain variability can lead to a lack of generalizability and stability in the unified intent classification model. To solve this problem, we present Heterogeneity Aware Transformer for Intent Detection (HyTra-ID), a robust transformer-based framework that addresses the inconsistencies that arise from varying domains through a unified preprocessing pipeline, domain-aware dataset embedding modules, and an enhanced classification architecture specifically designed for large-scale multi-domain intent detection. We learn a single-fused representation that combines semantic sentence features with dataset-specific embedding cues needed for HyTra-ID to adapt to various sources of conversation. Our extensive experimental and ablation study results in a combined benchmark of 102 unified intent classes. HyTra-ID shows robust performance across the combined ATIS, SNIPS, and TOPv2 benchmark datasets and in producing accurate predictions of 96.34% for overall accuracy on HyTra-ID unified test set and outperforming all simplified ablation variants tested. HyTra-ID provides a scalable, domain-robust framework for real-world dialogue systems with consistent intent understanding across multiple domains.