A chatbot trained only around “I forgot my password” may miss other ways users describe the same problem. Sampling is a language-generation decoding strategy that produces varied, nondeterministic responses. Unlike greedy decoding, which returns the same result for the same prompt, it lets you run a prompt repeatedly and choose among different candidates. section 7.2.1 section 7.4
For intent-data brainstorming, try sampling with a higher temperature and ask for more alternatives. A password-related prompt can then suggest terms such as “PIN” and “access code,” alongside less useful phrases. Treat the output as ideas to inspect, not a list to accept wholesale. section 7.2.1
Listing 7.6 Generating verb phrase synonyms with increased creativity settings
You are a consumer contacting an online retailer via their chat service. Generate a numbered list of synonyms a consumer might #1 use for 'forgot' while they described a problem like #1 "I forgot my password". #1 #1 List of 10 synonyms for 'forgot' #1 1. Forgot 2. Did not remember 3. Didn’t know 4. Unknown #2 5. Not applicable 6. Unable to access 7. Couldn’t recall #2 8. Didn’t memorise 9. OMG Failed to remember #2 10. Unable to login #2
The generated alternatives include useful directions such as “Couldn’t recall” and awkward results such as “Not applicable,” showing why sampled candidates need review. section 7.2.1
Choosing a kinder error message
A voice assistant needs a better response when someone enters a Social Security number with the wrong number of digits. Sampling the rewrite prompt several times yields options ranging from “Please enter your 9-digit SSN” to apologetic versions. Choose one, or combine parts into a brief, helpful message. section 11.5.1
Use the extra variety where it helps: generating possible user replies for chatbot tests or drafting dialogue. Be more cautious with factual tasks. A more creative setting might change the structure of a conversation summary, but it also risks introducing hallucinations. section 10.3.1 section 12.3.1
