Sentiment Classifier Scenarios - MangaAssist

This companion document grounds sentiment fine-tuning in MangaAssist. The goal is not generic positive/negative sentiment; the goal is to detect frustration, urgency, and escalation risk inside shopping and support conversations.

When This Topic Matters

Use a fine-tuned sentiment or frustration classifier when:

• users ask for a human agent,
• repeated support loops create frustration,
• a neutral support message hides urgency,
• routing should become more conservative.

Scenario 1 - Frustration Detection For Escalation

Example messages:

"I already asked this twice. Where is my order?"
"This refund answer is useless."
"Can I please talk to a real person?"

Labels:

Training setup:

Promotion gate:

Scenario 2 - Conversation-Level Sentiment

A single message may look calm, but the conversation pattern can show frustration.

Features:

• repeated question count,
• low-confidence answer count,
• user reformulation count,
• negative words,
• time since first issue.

Use a small sequence model or aggregate features around the message classifier.

Scenario 3 - Sentiment-Aware Routing

Routing policy:

if intent == escalation or sentiment == escalation_needed:
    offer human handoff
elif sentiment == frustrated and confidence_low:
    ask direct clarifying question
else:
    continue normal flow

Failure Modes

Production Log

{
  "event": "sentiment_route",
  "intent": "order_tracking",
  "sentiment": "frustrated",
  "frustration_score": 0.86,
  "turn_count": 5,
  "decision": "offer_human_handoff"
}

Final Decision

For MangaAssist, sentiment fine-tuning should optimize for operational safety. Missing frustrated users is more harmful than occasionally offering a human handoff too early.