Dialogue and Discourse Structure
A call is not a bag of sentences: turns, dialogue acts, adjacency pairs and segments.
Why this matters
A customer saying “yes” means nothing on its own. It means *I'm salaried* after “aap salaried ho?”, and *send me the link* after “shall I send the link?”. Conversations have structure: who speaks when (), what each utterance does (), which utterances belong together (), and which phase the call is in ().
Every downstream stage benefits: evidence extraction looks for answers to qualification questions, intent classification weighs the closing more than the opening, and QA teams measure what agents do in each segment.
Core concept
Technical explanation + simple intuition + example
Dialogue acts
Label each utterance by its *job*: greeting, question, answer, request, objection, refusal, confirmation, clarification, closing — and “inform” for statements like a sales pitch. The same words can do different jobs: “haan” can be an answer, a confirmation, or just a back-channel meaning “I'm listening”.
Adjacency pairs and conversation segments
Some utterances come in pairs: question → answer, offer → accept/decline, greeting → greeting. The first part makes the second expected. A bare answer is only interpretable through its pair.
Calls also move through phases: Opening → Discovery → Qualification → Offer → Objection handling → Closing. A topic shift (“aur credit card ka bhi bata do”) can start a new mini-conversation about another product.
The 10 questions
Every important concept should answer all of these
1What is it?
The structural organization of conversation: turns, the function of each utterance (dialogue acts), links between utterances (adjacency pairs) and phases (segments/topics).
2Why does it exist?
Meaning in dialogue is distributed across turns; single utterances are often uninterpretable.
3What problem does it solve?
It lets systems locate evidence (answers to specific questions), weight later stances over earlier ones, and analyze agent behaviour per phase.
4How does it work internally?
Annotate or predict acts per turn using the turn text plus context; link first and second pair parts; group turns into segments by act patterns or topic shifts.
5What does it look like on real transcript data?
ASR merges or splits turns, back-channels (“haan”, “ji”) interleave with speech, Hinglish discourse markers (“aur”, “toh”, “matlab”) signal shifts, and diarization errors break pairs.
6What can go wrong?
Ambiguous acts (Inform vs Answer), back-channels mistaken for agreement, pairs broken by insertion sequences, and topic shifts that mix intents for multiple products in one call.
7How do I evaluate it?
Act-level accuracy/macro-F1 against annotated turns, pair-linking F1, and segment boundary F1 (with tolerance windows). Measure agreement between annotators first — acts are subjective.
8When should I use it?
Evidence extraction, agent QA (did the agent qualify, handle objections, close?), intent over time, and summarization structure.
9When should I NOT use it?
Don't build an elaborate act taxonomy if the downstream task only needs the final customer stance — an LLM given the whole call may suffice.
10How does it affect downstream systems?
Acts and pairs locate evidence; segments give context windows for long calls; act errors propagate into intent (e.g., reading a back-channel 'haan' as agreement).
Interactive lab: Conversation viewer & dialogue-act annotator
Label each utterance's dialogue act, compare with the reference, then see the conversation flow, adjacency pairs and segment boundaries built from your labels.
Try it yourself
Conceptual (A), applied (B) and engineering (C) exercises
Graded by a transparent keyword rubric (not an LLM) — it checks for the key ideas, then shows a model answer for self-assessment.
Real-world application
A QA team wanted to know whether agents *qualify* customers before pitching. Dialogue-act tagging plus segmentation let them measure, per agent, the fraction of calls where a Qualification segment (employment, income, tenure questions answered by the customer) precedes the Offer segment. Agents with low qualification rates had 2× more post-sale cancellations — a finding impossible to get from call-level intent labels alone.
Common mistakes
“Yes”, “haan”, “ok” change meaning with the preceding question.
Fix: Feed the previous turn(s) or the whole call; link adjacency pairs.
“haan… haan…” while the agent talks often means 'I'm listening'.
Fix: Distinguish back-channel acts from answers/confirmations.
Objections are often resolved later in the call.
Fix: Weight the final customer stance; model objection → resolution sequences.
Engineering decision
Do you need explicit dialogue structure, or can an LLM read the whole call?
- You measure agent behaviour per phase (QA, compliance)
- Calls are long and you need to select relevant segments
- Evidence must be tied to specific question→answer pairs
- The only output is a call-level label and calls are short
- You can't afford to annotate acts reliably
Quiz
Pass mark 60%. Missed questions go to your mistake log and spaced review.