Best Practices to Create Zero Shot Intents
Updated
Before You Begin
Learn how to create generative AI-powered zero-shot intents.
Overview
Zero-shot intents utilize Sprinklr AI+ to predict the relevant intent based solely on the Name and Description provided. Therefore, it is crucial that both the Name and Description are clear and highly relevant.
Best Practices for Zero Shot Intent Names
Keep intent names short, ideally consisting of 4-5 words, to clearly convey the user's intention. A descriptive intent name helps Sprinklr AI+ better understand user queries.
Use a verb followed by a noun to indicate the action or task the user wants to perform. For example, "Apply for a car loan" or "Check flight status" are clear and informative intent names.
Avoid generic or placeholder names like Intent 1, FAQ_3, or Bot_2 as they provide little context and can be confusing.
While abbreviations can be used, ensure they are widely understood or provide the full form for clarity. Avoid overly technical or industry-specific abbreviations that may not be familiar to all users.
Instead of using just a noun or a product/brand name, opt for descriptive phrases that provide more context about the user's intention. For example, instead of "Car Loan," use "Apply for a car loan" to specify the user's action.
Best Practices for Zero Shot Intent Descriptions
When crafting descriptions for intents, it's crucial to bear in mind that the prompt serves solely to classify the intent and lacks any context regarding the bot or its capabilities. Therefore, it's recommended to avoid phrases like: "Route the customer to an agent if the customer asks for one." Instead, phrase it as: "Analyze the customer query to determine if they seek interaction with an agent or human."
Here are some best practices to enhance your Zero Shot Intent Model performance:
1. Be Clear and Concise:
- Clearly and briefly explain the intent without unnecessary words. Avoid ambiguity.
- Example: "Order Status" intent could be described as "Inquiries and updates about the status of an existing order."
2. Provide Specific Examples:
- Include one or two specific examples to demonstrate typical messages falling under this intent.
- Example: For "Technical Support," add "Example: 'My internet is not working,' or 'How do I reset my password?'"
3. Use Unique Descriptors:
- Use distinct terminology in each description to minimize overlap and confusion between intents.
- Example: For "Billing Issue," use terms like "invoice," "charge," "payment," whereas for "Account Issue," use "login," "username," "account settings."
4. Ensure Contextual Relevance:
- Capture the context in which the intent would be used.
- Example: For "Product Inquiry," describe it as "Questions about product details, features, and availability."
5. Avoid Overlaps:
- Ensure each intent description stands apart from others and doesn't overlap in meaning.
- Example: "Cancellation Request" should focus solely on requests to cancel services or orders, not support inquiries.
6. Define Clear Boundaries:
- Clearly define what is and is not included in the intent to avoid misclassification.
- Example: For "Feedback," clarify with "Customer opinions on products or services, NOT related to current transactions or issues."
7. Use Action-Oriented Language:
- Use verbs that indicate the action or purpose related to the intent.
- Example: For "Warranty Claim," use "Requests to claim warranty for purchased items."
8. Consider Audience Perspective:
- Write from the perspective of how the audience might phrase their queries.
- Example: For "General Inquiry," you might say, "General questions regarding services, business hours, locations."
9. Avoid Jargon:
- Use simple and easily understandable language, avoiding technical jargon unless the intent is specialized.
- Example: Instead of "Network downtime issues," use "Problems accessing internet or network."
10. Maintain Consistent Structure:
- Keep a consistent structure and format across all intent descriptions.
- Examples:
"Customer Service: Questions and issues related to customer service, such as help with orders or account problems."
"Product Review: Customer reviews and feedback about products they have used."
11. Avoid Special Characters:
- While writing the description, avoid using any special characters as it might interfere with the Intent Model prediction/performance.
- Example: Use "and" instead of "&" in the description.
By following these best practices, you can create clearer and more effective intent descriptions, improving the accuracy and reliability of the zero-shot intent classification model.