Chatbot Miscommunication Repair Responses

Description

Design escalating responses for input the bot does not understand: notify, ask the user to rephrase, then reveal what the bot can help with, and measure breakdowns in logs.

Every conversational agent misunderstands some input, and how it recovers shapes whether users persist. The escalating repair pattern has three tiers. The first response notifies the user that the bot did not understand. Paired with it is a repair request asking the user to rephrase. If the failure repeats, the bot tells the user what kinds of information or actions it can provide, a capability hint that earlier chatbot studies found effective. Liebrecht, Kamoen and Aerts used this set in voting-advice chatbots and also randomised connective sentences (acknowledging, restating, confirming) so conversation felt varied. Coding chat logs then showed far more miscommunication with voice input than with text. The practice is only half design: repair responses need log-based measurement of how often each tier fires and whether users recover.

Attributes

Loss type
Active
Data
Mixed
Effort
Days

Steps

  1. List the fallback situations: no intent matched, low confidence, out-of-scope request, and repeated failure.
  2. Write a short error notification for the first failure, avoiding blame.
  3. Pair it with a repair request asking the user to rephrase, ideally with an example phrasing.
  4. On a second consecutive failure, list the types of information or actions the bot supports, as buttons where possible.
  5. On a third failure, offer handover to a human or another channel.
  6. Write several variants of each response and of connective sentences, and rotate them so the bot does not repeat itself.
  7. Instrument each tier and review logs for how often it fires and whether users recover, per modality.

Sources

  • Chatbot Research and Design: CONVERSATIONS 2022 (Følstad et al., 2023)

Also known as

  • Chatbot error recovery strategy
  • Escalating repair responses
  • Chatbot miscommunication repair responses with capability hints

Categories

  • Design
  • Design & Product Practice