Task-oriented dialogue (TOD) is a system that helps users achieve their goals. While the task is reviewed and improved regularly, a formal system for industrial standards has not yet been established. Dialogue state tracking is a sub-task that involves predicting current dialogue slot values given the conversation and in some cases, the slots that are being required or informed. Based on a well-documented schema with instructions for possible slots and intents along with their descriptions, schema-guided TOD exploits a concrete set of guidelines to add extra context and perform general zero-shot ability on state tracking. Despite having contextual schema descriptions, language models hardly keep up with a full TOD dialogue flow. The TOD system as a whole lacks the mechanics to detect out-of-scope events, decide when to query the database, and is hardly extensible for further processing. To address these issues, we propose a full TOD system designed to overcome the listed weaknesses. Additionally, we experiment with dialogue state tracking, the system’s first stage, and measure out-of-scope detection effectiveness via user-undefined actions.
