Guidelines for the Cross-Lingual Intent Model

Utterances

Intents are built from conversational phrases called utterances. Utterances, also referred to as intent data, represent variations of customer input for an intent.

• Write high-quality utterances that vary in word length and structure. Make sure you use multiple words in utterances. Utterance quality is more important than quantity for the cross-lingual intent model.

  Sample utterances for the intent “Check the status of an order”:

  • “What is the status of my order?”
  • “I’m checking up on my order, can you look up the status”
  • “I would like to know the status of an order I placed”
  • “Status of my order please”
• Write utterances that clearly illustrate the customer’s goal or intention. Use actionable verbs, such as “look up,” “order,” and “update.” Include a noun that’s acted on, such as “order,” “password,” or “transactions.” For example, the utterance “I want to update my password.”
• Make utterances short and to the point. We recommend keeping them under 150 characters.
• We recommend including 1 to 20 utterances for each intent for your primary language. For secondary languages, add at least 1 utterance and iterate based on performance.
• Include a similar number of utterances in each intent.
• If you’re moving your bot from the original model to the cross-lingual model, review and revise your utterances so that they follow the utterance best practices.

Intents

Intents power your bot’s dialogs. If the model matches a customer’s input to an intent, the bot moves to the dialog associated with that intent.

• Create distinct and focused intents. The model performs better when each intent-enabled dialog has different goals and utterances.

  Sample intents:

  • Distinct: “Order Status” and “Cancel Order”
  • Too similar: “Order Status” and “Order Related”
  • Too broad: “Manage Order”
• When you can’t break down an intent any further, include utterances that vary in word length and structure.

  Sample utterances for the intent “End Chat”:

  • “I’m done for today, thanks”
  • “End chat”
  • “I have no further questions”
• Test your intents and iterate with bot training.

Troubleshooting

• If the customer input doesn’t invoke the Confused dialog when expected, determine which intents are invoked incorrectly. Then revise those intents and their utterances to follow best practices.
• If the customer input matches to the wrong intent, revise the correct and wrong intent to follow best practices. Remove low-quality utterances, including one-word utterances and similar utterances, and then retrain the model.