AI uses for records
Call intent change dataset: from a billing question to a cancellation
By SourceX Editorial · Updated
10 sections
Short answer
Represent a call that moves from billing to cancellation as intent segments over timestamped, speaker-attributed turns, joined by explicit transition events. Each transition records the anchor turn where the new intent first appears, its trigger, who raised it, its type and what happened to the earlier intent. Take outcomes such as an actual cancellation from account records, kept apart from what was said.
Call to transcript
Call intent change dataset
Key takeaways
- A single intent label per call erases the opening issue, the turning point and any attempt to resolve it.
- Label intent segments over speaker turns and join them with typed transitions anchored to a specific turn and time.
- Tell escalation, addition, disclosure, conditional, agent-led and return transitions apart, and never count a repair as a switch.
- Take outcomes from billing or account records with an observation window, not from the call or the agent's note.
- Validate every record so anchors, triggers, speakers and outcome citations agree before it enters a split.
How should a call dataset represent a conversation that changes from billing to cancellation?
Represent it as a sequence of intent segments laid over timestamped, speaker-attributed turns, joined by explicit transition events. Each transition records the turn and time where the new intent first appears, what triggered it, who raised it and what happened to the earlier intent. Outcomes, such as whether the account was actually canceled, come from the billing or account system and sit in separate fields from what anyone said on the call.
A single call-level label of cancellation hides most of what a model needs to learn from this call: that it opened as a billing question, which agent response or account fact turned it, whether the billing issue was resolved before the caller moved on, and whether the stated intent to cancel became an actual cancellation. A single label of billing is worse, because it teaches the model that the call stayed routine.
The schema below keeps all of that. It has five record types, a worked record for a call that moves from a disputed charge to a cancellation, a set of validator checks and one record that fails them.
Why one intent label per call loses the transition
Most call and chat datasets were designed around one goal per conversation, and contact center dispositions usually follow the same habit. The Action-Based Conversations Dataset (ABCD), for example, holds more than 10,000 human-to-human customer service dialogues with 55 user intents, and each conversation's scenario carries one flow and subflow as its ground-truth intent. That design suits intent classification and policy compliance, but it has no place to record a second goal arriving mid-call.
Task-oriented dialogue research has better structures to borrow. The Schema-Guided Dialogue dataset annotates every user turn with frames, one per service, and each frame carries a dialogue state whose active_intent names the intent being fulfilled, or NONE. In its multi-domain dialogues some turns carry more than one frame. MultiWOZ reports 7,032 multi-domain dialogues that involve between two and five domains. Both show that turn-level state, not a conversation-level label, is how a change of goal becomes visible.
Production voice bots treat switching as a measured event too. Amazon's Lex V2 analytics define an intent as Switched when the bot recognizes a different intent and moves to it before the original is classified as a success or a failure. A dataset built from human calls should be at least that explicit.
Which kinds of intent change should the dataset tell apart?
Not every change of topic is the same event, and a model that conflates them will mishandle the call. Label the type of each transition, because the right agent response differs for each one.
At a glance
Which kinds of intent change should the dataset tell apart
- Escalation: the original issue drives the new intent, as when a disputed charge leads to a request to cancel. The earlier intent is usually unresolved or resolved badly.
- Addition: the caller adds an unrelated task, such as updating an address after the billing question is settled. Both intents can succeed.
- Disclosure: the caller had the second intent from the start and opened with something easier, for example asking about a charge before saying they want to leave. Mark the earliest turn with evidence of it, which may precede the explicit request.
- Conditional: the caller states a new intent as a condition, such as canceling unless a fee is removed. Record the condition and whether it was met.
- Agent-led: the agent proposes the new intent, such as a plan change or a retention offer. Record that the agent, not the caller, raised it.
- Repair: the apparent change came from a misunderstanding that was corrected. This is not a real transition, and the segment should be relabeled rather than counted as a switch.
- Return: the caller goes back to an earlier intent after a detour. Lex documents the same pattern for bots, where an application can switch intents and later restore a previous intent's saved dialog state.
The multi-intent transition schema
The schema has five record types: call, turn, intent segment, transition and outcome. Turns hold what was said and when; segments and transitions hold the interpretation; outcomes hold what the business systems show afterward. Keeping interpretation and evidence in separate records lets reviewers correct a label without touching the transcript.
The table lists the fields that carry the transition. Call-level fields such as channel layout, audio rights status and recording disclosure belong on the call record and are checked before any audio is used.
| Record and field | Definition | Validation rule |
|---|---|---|
| turn.turn_id, speaker, start_ms, end_ms | One speaker turn with its role (caller, agent, system) and audio offsets | Turns ordered by start_ms; end_ms after start_ms; overlaps allowed only when flagged |
| turn.text and asr_source | Transcript of the turn and whether it is ASR output or a human-corrected transcript | Every turn has text or an explicit inaudible marker |
| segment.intent and status | Intent from a fixed taxonomy and how it ended: resolved, abandoned, superseded or transferred | Each segment ends with exactly one status |
| segment.first_turn and last_turn | The turns the segment covers | Segments cover every caller turn after greeting; overlaps only for parallel intents |
| transition.from, to and type | The two segments joined and the transition type (escalation, addition, disclosure, conditional, agent-led, return) | A repair is never stored as a transition |
| transition.anchor_turn and anchor_ms | The first turn and time where evidence of the new intent appears | anchor_turn falls inside the new segment and belongs to the party named in raised_by |
| transition.trigger_turns and raised_by | The turns that caused the change and who raised it | Trigger turns precede or equal the anchor turn |
| outcome.system, event, effective_at, maturity | A business-system event for each segment, such as a fee reversal or account cancellation, with its date and whether the observation window has closed | Each outcome cites a system record id; immature outcomes are never used as final labels |
Timing the transition point in the audio
Anchor the transition to a speaker turn with millisecond offsets, and score predicted transition times with a tolerance rather than an exact match. Turn boundaries in human transcripts are approximate, and diarization research already handles this: pyannote.metrics notes that it is common to exclude a 500 ms collar around each speaker turn boundary, 250 ms on each side, because manual annotations are not precise to the sample, and its metrics take a collar parameter that defaults to 0. State the tolerance you use alongside the data.
Speaker attribution matters as much as time. A transition raised by the agent and one raised by the caller are different labels, so the turn record must name the speaker reliably. Dual-channel recordings make that easier; for mixed audio, keep the diarization source and confidence. For delivery, WebVTT, the W3C caption format, has cue voice spans that let each cue say who is talking, which keeps speaker and time together in one file.
Repair turns need explicit handling. When the agent hears cancel my order and the caller meant cancel the autopay, the transcript may show a switch that never happened. Mark the misunderstanding and the correcting turn, keep the original segment, and do not create a transition.
Checking the audio against the transcript, and the transcript against the account
Check both links separately. The recording supports the transcript when the audio at the anchor turn contains the words the label relies on, spoken by the labeled party; a reviewer should listen to every anchor turn rather than trust ASR output there. The transcript supports the outcome only when a system record shows the event: a stated intent to cancel is a conversational signal, while a cancellation is an account status change with an effective date.
Keep those as different fields with different maturity rules. A retention offer accepted on the call can be reversed the next week, and a cancellation can be reactivated. Set an observation window per outcome type before labeling, record whether it has closed, and leave outcomes as pending rather than guessing from the call's tone. Window lengths are a dataset decision; document them and apply them to every call.
Worked record (illustrative): a disputed charge that becomes a cancellation
The record below is invented to show the schema; it does not come from any real call or company. A subscriber calls about a late fee on the latest invoice. The agent explains the fee and declines to remove it because the payment arrived after the grace period. The caller says they will cancel unless it is removed, the agent offers a one-time courtesy credit, the caller declines and asks to cancel, and the agent processes the cancellation.
| Turn and time | Speaker and words (abridged) | Segment | Transition and outcome fields |
|---|---|---|---|
| t3 at 00:41 | Caller: why is there a late fee on this bill | S1 billing_dispute begins | S1 first_turn t3 |
| t6 at 01:52 | Agent: the payment posted after the grace period, so the fee stands | S1 continues | trigger candidate for the next transition |
| t7 at 02:05 | Caller: then I want to cancel unless you take it off | S2 cancel_service begins | transition S1 to S2, type conditional, anchor t7, trigger t6, raised_by caller |
| t9 at 02:40 | Agent: I can offer a one-time courtesy credit on the next invoice | S3 retention_offer begins | transition S2 to S3, type agent-led, anchor t9, raised_by agent |
| t10 at 02:58 | Caller: no, just cancel it | S2 resumes | transition S3 to S2, type return, anchor t10; S3 status abandoned |
| after call | Billing system and account system records | Outcomes | S1 superseded with fee unchanged; S2 cancellation effective next cycle; maturity closed after an illustrative 30-day window with no reactivation |
Validator checks, and a record that fails them
Run these checks on every record before it enters a training or evaluation split. They are cheap, they catch most labeling drift, and the failures point reviewers to the exact field to fix.
- Every caller turn after the greeting belongs to at least one segment, and every segment has exactly one end status.
- Each transition's anchor turn lies inside the new segment, and its trigger turns do not come after the anchor.
- raised_by matches the speaker of the anchor turn, unless the type is agent-led and the anchor is an agent turn.
- No transition has type repair; repairs are stored as turn annotations instead.
- Every outcome cites a system record id and an effective date, and records whose window is still open are marked pending.
- Anchor turns were reviewed against the audio, and the review result is stored per transition.
- The call record shows audio rights and the recording disclosure as reviewed before audio is used.
What the failing example gets wrong
Take the same call labeled the quick way: one call-level intent of cancel_service, a transition from billing to cancellation anchored at t10, and an outcome of canceled taken from the agent's after-call note. It fails four checks. Turns t3 to t6 belong to no segment, so the billing dispute disappears. The anchor is late: the first evidence of cancellation is t7, where the caller set a condition, and anchoring at t10 erases the conditional and the retention offer. The retention segment is missing, so there is no record that the agent tried to keep the customer. And the outcome cites a note rather than an account record, so a cancellation that was later reversed would still read as canceled.
A model trained on the failing version would learn to treat late-fee calls as cancellations and would never see the window in which an agent's response could have changed the result. That window is usually what a buyer building a call agent most needs, which is why the anchor and trigger fields are worth the labeling cost.
Finding calls whose outcomes can be checked
Transition data is only as good as the link between the call and the account systems that recorded what happened next. Businesses with subscription billing, service plans or memberships often hold both: call recordings or transcripts, plus billing and account histories with cancellation, credit and reactivation events.
Call-to-account data of this kind comes through SourceX by request, not from a back catalog. A voice team names its intent taxonomy, whether it needs audio or transcripts, which outcome systems must be linked and its observation window, along with volume, language, format, timing and licensing terms; SourceX then looks for U.S. businesses whose recordings and billing histories fit and handles qualification, rights review, contracting, preparation and delivery. Recordings move only after the supplier has approved the scope and signed.
Frequently asked questions
What is a call intent change dataset?
It is a set of call recordings or transcripts labeled with every intent the caller pursued, the point where each change happened and the business outcome of each intent. Unlike a single-intent dataset, it records transitions such as a billing question turning into a cancellation request.
Should a call that ends in cancellation be labeled as a cancellation call?
Only at the call-summary level, and only if the account system confirms it. The dataset should still keep the billing segment that started the call, the turn where cancellation first came up and any retention attempt, because those are what an agent model needs to learn from.
How do you label a caller who threatens to cancel but does not?
Record a conditional transition with the condition, such as removing a fee, and whether it was met. Link the outcome from the account system. If the account stayed active, the cancel segment ends as abandoned and the outcome shows no cancellation.
How precise does the transition timestamp need to be?
Anchor it to the turn where evidence of the new intent first appears and store that turn's start time. When scoring a model's predicted transition, use a stated tolerance, because human turn boundaries are approximate.
Can transcripts alone support outcome labels?
No. A transcript shows what was said, such as a request to cancel or an accepted offer. Whether the account was canceled, credited or reactivated has to come from the billing or account system, checked after an observation window.
Sources
- ABCD contains over 10,000 human-to-human customer service dialogues with 55 user intents, and each scenario's flow and subflow are the ground-truth intent. Source
- Schema-Guided Dialogue groups turn annotations into per-service frames whose state includes active_intent, or NONE, and multi-domain turns may carry several frames. Source
- MultiWOZ contains 7,032 multi-domain dialogues involving two to five domains. Source
- Amazon Lex V2 analytics classify an intent as Switched when the bot recognizes a different intent before the original succeeds or fails. Source
- Amazon Lex V2 lets an application switch intents with PutSession and resume a previous intent from saved state. Source
- pyannote.metrics documents the common 500 ms collar around speaker turn boundaries and a collar parameter that defaults to 0. Source
- The WebVTT specification defines cue voice spans that identify who is talking in each cue. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.