Slot filling
Slot filling is the process of collecting the specific pieces of information required to complete a conversational task. Each “slot” represents a field—such as a service type, date, address, or customer name—and the agent continues the conversation until required fields are known, validated, or deliberately left unresolved.
Consider an appointment request. The workflow might require the caller's name, the reason for the visit, a preferred date, and a time. Some values may arrive before the agent asks for them: “Book a furnace inspection Friday morning” supplies the service and a time range in one turn. A capable dialogue manager records those values, asks only for what is missing, and updates a slot when the caller corrects it.
Entity extraction and slot filling are related but different. Entity extraction identifies a phrase and its type. Slot filling assigns the resulting value to a field in the active task and tracks whether that field is usable. A detected date may still be incomplete if the year is ambiguous, and an address may need validation before it can support a service call.
Good slot design distinguishes required, optional, derived, and conditional information. An email address may be optional for a basic inquiry but required for a follow-up. A transfer reason may become necessary only when a caller asks for a specialist. Treating every possible field as mandatory creates long, rigid calls; collecting too little can make the resulting action unreliable.
Voice workflows also need explicit rules for confirmation and correction. High-impact values such as dates, phone numbers, and account identifiers often deserve confirmation. Lower-risk values can sometimes be accepted without repetition. If a caller says “Actually, make that Monday,” the newest clear value should replace the earlier one while the rest of the task state remains intact.
When evaluating slot filling, test incomplete answers, multiple values in one response, mid-call corrections, unavailable choices, silence, and requests that fall outside the workflow. The goal is not to fill every field at any cost. It is to collect enough reliable information to complete the task, ask a useful follow-up, or hand the caller to a person with the context already gathered.