Recommend an Alternative for Problematic Text Mitigation in a Data Source Feeding a GenAI Application GENERATIVE-AI-ENGINEER-ASSOCIATE Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   A customer support chatbot is trained on historical tickets that contain both PII and occasional toxic language. The team wants to minimize the risk of the chatbot leaking PII or producing toxic responses. Which mitigation approach is most suitable?

   • AFiltering out all tickets that contain PII or toxic language
   • BExcluding the entire dataset
   • CRedacting PII and filtering toxic tickets
   • DRewriting all tickets to remove PII and toxic language
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   A legal research firm is building a document-summarization tool that processes court rulings. The rulings contain sensitive personal information about litigants, such as names, addresses, and social security numbers. The firm wants to summarize the rulings for public access, but must comply with privacy laws. The firm also wants to preserve the legal reasoning in the summaries. Which mitigation strategy is most appropriate?

   • AExclude all rulings that contain personal information from the dataset.
   • BRedact the personal identifiers from the rulings before summarization, and then generate summaries from the redacted text.
   • CUse a prompt that instructs the model to omit personal information from the summaries.
   • DGenerate summaries from the original rulings and then filter out any summary that contains personal information.
   Reveal Solution Discussion Select an answer first
4. 13foundation · easy

   During a data quality review, a team identifies text that contains gender-based stereotypes. Which category of problematic text does this represent?

   • APersonally identifiable information (PII)
   • BBiased content
   • CToxic content
   • DOffensive language
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A financial institution is building a chatbot to answer customer questions about their accounts. The chatbot uses a retrieval-augmented generation (RAG) pipeline that pulls from a database of customer transaction notes. These notes contain account numbers and transaction details. The institution wants to prevent the chatbot from exposing account numbers, but the notes are the only source of information for the chatbot. Which mitigation strategy is most effective while maintaining the chatbot's functionality?

   • AExclude all notes that contain account numbers from the knowledge base.
   • BRedact the account numbers from the notes before indexing them for retrieval, and use a mapping table to re-identify the account when needed.
   • CUse a prompt that instructs the model to never reveal account numbers.
   • DAdd a response filter that blocks any output containing a 10-digit number.
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A social media analytics firm is training a sentiment-analysis model on a large corpus of public posts. The corpus contains a significant amount of profanity and offensive language, but the firm's clients expect the model to understand sentiment in informal text. The firm wants to reduce the model's generation of offensive language while retaining the ability to analyze sentiment. What is the most suitable mitigation?

   • AFilter out all posts containing profanity.
   • BRewrite the offensive posts to remove profanity while preserving the sentiment and meaning.
   • CRedact the profane words and replace them with a placeholder token.
   • DExclude all posts from users who use profanity.
   Reveal Solution Discussion Select an answer first