Research

This work addresses the critical challenge of deploying accurate multi-turn intent classification in production conversational AI systems, where scalability and latency constraints often conflict with model performance. We introduce two novel approaches: Symbol Tuning, which simplifies intent labels to reduce task complexity and improve classification accuracy, and C-LARA (Consistency-aware, Linguistics Adaptive Retrieval Augmentation), a framework that leverages LLMs for synthetic data generation and pseudo-labeling to train lightweight, deployable models. The methods achieve a 5.09% improvement in classification accuracy while reducing annotation costs by 40%, enabling scalable deployment in low-resource multilingual industrial systems. This research demonstrates practical solutions for bridging the gap between research-grade models and production requirements in conversational AI.

Training effective multi-turn intent classification models requires large-scale, domain-specific, multilingual dialogue datasets—a significant bottleneck in conversational AI development. We introduce Chain-of-Intent, a novel framework that integrates Hidden Markov Models (HMMs) with Large Language Models (LLMs) to generate intent-driven, context-aware dialogues through self-play. The framework first extracts domain-specific intent transition patterns from real-world e-commerce chat logs, then uses LLMs to produce natural utterances aligned with predicted intents. We also present MINT-CL, a multi-task contrastive learning approach that improves classification performance while reducing dependence on large-scale annotated datasets. Additionally, we release MINT-E, a comprehensive multilingual intent-aware multi-turn dialogue corpus derived from e-commerce interactions. Experiments demonstrate superior performance compared to existing baselines across multilingual dialogue generation and classification tasks.

Multi-turn intent classification presents unique challenges due to the complexity and evolving nature of conversational contexts. LARA introduces a Linguistic-Adaptive Retrieval-Augmentation framework that enhances classification accuracy across six languages while accommodating numerous intents in chatbot interactions. The architecture combines a fine-tuned smaller model with a retrieval-augmented mechanism integrated within LLMs, enabling dynamic utilization of past dialogues and relevant intents for improved context understanding. Our adaptive retrieval techniques bolster cross-lingual capabilities without extensive retraining, making the approach practical for multilingual production deployments. Comprehensive experiments demonstrate state-of-the-art performance, achieving 3.67% average accuracy improvement over existing single-turn intent classifiers. This work addresses real-world requirements for scalable, multilingual conversational AI systems in industrial applications.

This paper proposes a hybrid quantum-classical convolutional neural network (QCCNN) that combines quantum computing with classical deep learning to enhance feature extraction capabilities. Inspired by classical CNNs but adapted for quantum computing, QCCNN leverages quantum circuits to generate complex distributions that are computationally expensive for classical computers to produce. The architecture is designed to be compatible with current noisy intermediate-scale quantum (NISQ) devices, respecting practical constraints on qubit count and circuit depth while retaining essential CNN properties like nonlinearity and scalability. We present a framework for automatically computing gradients of hybrid quantum-classical loss functions, applicable to other hybrid algorithms. Experiments on a Tetris classification dataset demonstrate that QCCNN achieves learning accuracy surpassing classical CNN, validating the potential of quantum-enhanced neural networks for pattern recognition tasks.