Advancing Task-Oriented Dialog Systems

Task-oriented dialog (TOD) systems enable conversational interfaces for complex tasks like flight booking and restaurant reservations. However, deploying TOD systems at scale faces three critical barriers: scalability, generalization, and evaluation. Scalability is primarily restricted by the human-annotation bottleneck, as current systems depend on vast quantities of manually labeled data for every new domain, making deployment prohibitively expensive. Generalization remains a persistent challenge, as systems optimized for known domains often suffer significant performance degradation when encountering new, unseen ones. Existing evaluation metrics measure response quality and fluency, but fail to measure functional task success. As TOD systems are deployed across diverse real-world domains powering millions of daily interactions, overcoming these three barriers is critical to advancing practical and scalable dialog systems. This dissertation addresses these barriers through three progressive contributions. First, we propose SCot, a framework for annotation-free slot filling that uses co-training on unannotated data to achieve performance comparable to supervised methods. Building on this foundation, we introduce ZS-ToD, an end-to-end TOD system designed for zero-shot generalization, enabling adaptation to new domains without retraining. Finally, we present ZeroToD, which synthesizes all three dimensions. ZeroToD removes per-turn annotation requirements for scalability, enhances generalization through schema augmentation, and evaluates task completion through structured API calls, demonstrating that fine-tuned open-source systems can exceed proprietary approaches on unseen domains.