AI-900 Azure AI Fundamentals - Set 3 - Part 1

Correct Answer: Unique face ID tracking

Fairness requirements directly address treating all people fairly and avoiding bias in AI systems, which is essential for ensuring equal treatment regardless of patient demographics in healthcare applications.

Option b is incorrect because Inclusiveness is about empowering everyone and designing for minorities, not specifically about preventing biased treatment recommendations based on demographic factors in AI systems.

Option c is incorrect because Transparency focuses on making AI systems understandable to users, not specifically on preventing demographic bias in treatment recommendations or ensuring equal outcomes.

Option d is incorrect because Privacy protocols focus on protecting user data and personally identifiable information, not on ensuring equal treatment or preventing bias in AI decision-making processes.

Correct Answer: Deploy custom trained models

Custom trained models allow extraction of text and data tailored specifically to unique company formats, requiring only five sample forms to start training for specialized document processing needs.

Option a is incorrect because prebuilt invoice models are designed for standard invoice formats and cannot handle unique company-specific formats or proprietary measurement systems found in technical drawings.

Option b is incorrect because receipt recognition templates are specifically built for sales receipts from retail environments and lack the capability to process technical drawings with specialized formats.

Option c is incorrect because business card extraction is designed only for contact information on business cards and cannot process technical specifications or proprietary measurement systems from drawings.

Correct Answer: Prebuilt invoice models

Prebuilt invoice models can extract data from invoices in various formats and return structured data that preserves relationships between fields like line items, totals, and vendor information.

Option a is incorrect because while custom models can be trained, prebuilt invoice models already support multiple formats and languages without requiring training time or sample documents.

Option b is incorrect because business card prebuilt models are specifically designed to extract contact information from business cards, not invoice data, line items, or financial information.

Option d is incorrect because receipt templates are designed for retail sales receipts and lack the comprehensive field extraction needed for supplier invoices with complex line items and vendor details.

Correct Answer: Intent classification system

Intent classification system in LUIS categorizes user utterances into predefined intents, allowing the bot to recognize and classify expressions of frustration as complaints rather than general inquiries.

Option a is incorrect because utterance examples are training data used to teach the model, not the actual component that performs the classification of user input into intent categories.

Option c is incorrect because entity extraction features identify specific data within utterances like names or dates, not classify the overall purpose or sentiment of user statements.

Option d is incorrect because schema generation tools create the structure for the LUIS application but do not actively classify or categorize user input during runtime interactions.

Correct Answer: Entity extraction

Entity extraction in LUIS identifies and extracts specific data elements like names, product codes, and quantities from user utterances to determine the detailed information needed for processing requests.

Option b is incorrect because intent classification determines what the user wants to do overall, not the specific data elements like names, product codes, or quantities within the message.

Option c is incorrect because None intent handler is used to explicitly train the model to ignore certain utterances, not to identify and extract specific data elements from messages.

Option d is incorrect because utterance training provides example inputs to teach the model but does not actively extract specific data elements during actual user interactions with the system.

Correct Answer: Document and URL import

QnA Maker can automatically build knowledge bases from existing documents like PDFs and websites, extracting question-answer pairs without manual entry, saving significant time and effort in setup.

Option a is incorrect because chitchat personality datasets provide prepopulated casual conversation responses for generic queries, not automated extraction of content from existing company documentation or websites.

Option b is incorrect because multi-turn conversation prompts enable follow-up questions and contextual conversations but do not import or extract content from existing documents or web pages automatically.

Option c is incorrect because active learning suggestions help improve existing questions based on user interactions at the endpoint, not import or extract initial content from documents or URLs.

Correct Answer: Same answer for questions

QnA Maker specifically handles cases where different users submit variations of the same question, returning the identical answer regardless of how the question is phrased.

Option a is incorrect because static information knowledge base refers to storing unchanging content but does not specifically address handling multiple variations of the same question with consistent responses.

Option b is incorrect because metadata filtering options help categorize and filter content based on tags but do not specifically ensure the same answer is provided for question variations.

Option d is incorrect because multi-turn conversation flow manages complex conversations requiring multiple exchanges but does not specifically address providing consistent answers to question variations submitted by different users.

Correct Answer: Multi-turn context prompts

Multi-turn context prompts in QnA Maker enable predetermined conversation flows where the bot provides follow-up questions based on previous answers, perfect for structured multi-step processes like account opening.

Option a is incorrect because active learning refinement suggests improvements to existing questions based on user interactions but does not create structured, predetermined conversation flows with sequential question dependencies.

Option c is incorrect because metadata tag filtering helps categorize and retrieve content based on attributes but does not establish sequential conversation flows where each answer determines the next question.

Option d is incorrect because chitchat content provides casual conversational responses for generic queries but does not create structured, step-by-step dialogue flows required for complex processes like account opening.

Correct Answer: Reliability and Safety standards

Reliability and Safety principles ensure AI systems perform reliably in critical applications where failures could cause harm, making it essential for emergency response scenarios with serious safety implications.

Option b is incorrect because while Inclusiveness ensures accessibility for all users, it does not specifically address the need for reliable performance during critical situations with potential safety consequences.

Option c is incorrect because Transparency focuses on making AI systems understandable to users but does not specifically ensure reliable performance during critical emergency situations where failures could cause harm.

Option d is incorrect because Accountability establishes human responsibility for AI systems but does not directly ensure the system performs reliably without failures during critical safety-dependent operations.

Correct Answer: Accountability framework

Accountability framework establishes clear human responsibility for AI systems and ensures decisions can be traced and audited, directly addressing the need for transparency in decision-making and error responsibility.

Option a is incorrect because Privacy protection focuses on securing user data and personally identifiable information, not on establishing audit trails or determining responsibility for AI decision-making errors.

Option b is incorrect because Fairness standards address treating all people fairly and avoiding bias but do not specifically ensure decisions can be audited or establish responsibility frameworks.

Option c is incorrect because Inclusiveness design focuses on empowering everyone and designing for minorities but does not address audit trails or establishing responsibility when AI systems make errors.

Correct Answer: Layout structure preservation

Form Recognizer's defining feature is preserving structure and relationships of form-like data during extraction, which is essential for maintaining field relationships in medical records during digitization processes.

Option a is incorrect because basic OCR text extraction only converts printed text to digital format without preserving the spatial relationships and structure between different fields in the documents.

Option b is incorrect because prebuilt model templates offer convenience for common document types but do not specifically ensure the preservation of structural relationships in complex handwritten medical records.

Option d is incorrect because while custom model training helps with accuracy, it does not specifically address the fundamental capability of preserving layout structure and field relationships during text extraction.

Correct Answer: Receipt prebuilt model

Receipt prebuilt models specifically support sales receipts from Australia, Canada, Great Britain, India, and United States, extracting fields like transaction date, items, quantities, and prices without requiring training.

Option b is incorrect because invoice prebuilt models are designed for vendor invoices with different field structures like purchase orders and billing addresses, not retail sales receipts with itemized purchases.

Option c is incorrect because business card templates extract contact information from business cards such as names and phone numbers, not transaction details or itemized purchases from receipts.

Option d is incorrect because custom trained models require sample documents and training time, while prebuilt receipt models already handle diverse formats from multiple countries without any training needed.

Correct Answer: Business card and ID models

Business card prebuilt models extract contact information from English business cards while ID prebuilt models handle worldwide passports and identity documents, providing the most efficient solution for both requirements.

Option a is incorrect because custom models require training with sample documents for each type, which is inefficient when prebuilt models already exist for both business cards and identity documents.

Option b is incorrect because receipt and invoice models are designed for financial transaction documents, not contact information from business cards or identity data from passports and identification documents.

Option c is incorrect because while layout models can extract text, they lack the specialized field extraction capabilities that business card and ID prebuilt models provide for specific information types.

Correct Answer: Various format support

Prebuilt invoice models can extract data from invoices in various formats and return structured data, making them ideal for handling supplier invoices with significant format variations across countries.

Option a is incorrect because fixed template matching would fail with invoices that vary significantly in format, whereas prebuilt invoice models are designed to handle diverse formats without templates.

Option c is incorrect because supporting only a single language would not address the needs of a logistics company working with international suppliers who send invoices in multiple languages and formats.

Option d is incorrect because manual field mapping would be inefficient for varying formats, while prebuilt invoice models automatically identify and extract fields across different invoice structures without manual configuration.

Correct Answer: Entity extraction component

Entity extraction in LUIS identifies specific data types within user utterances, recognizing that "two" represents quantity and "Toronto" represents a location for proper request processing.

Option a is incorrect because intent classification determines the overall purpose of the user's request like booking versus canceling, not identifying specific data elements like quantities or location names.

Option b is incorrect because utterance example sets are training data used to teach the model patterns, not the component that actively identifies and extracts specific data types during conversations.

Option d is incorrect because None intent category is used to train the model to ignore certain utterances that don't match any defined intents, not to extract specific data elements.

Correct Answer: Intent classification

Intent classification in LUIS categorizes different user utterances that express the same goal into a single intent category, enabling the system to understand varied expressions of the same request.

Option b is incorrect because entity extraction identifies specific data within utterances like subscription types or dates, not the overall goal or purpose of different user expressions.

Option c is incorrect because utterance examples are the training data that help teach intent patterns, not the classification mechanism that groups varied expressions under a single intent category.

Option d is incorrect because schema generation creates the application structure but does not actively categorize or group different user expressions under a single intent during runtime interactions.

Correct Answer: None intent

None intent in LUIS is specifically designed to explicitly train the model to recognize and ignore utterances that don't match any defined intents, perfect for handling irrelevant messages.

Option a is incorrect because entity extraction identifies specific data within relevant utterances but does not handle filtering out completely irrelevant messages that don't match any application intents.

Option b is incorrect because utterance training provides examples for defined intents but None intent is specifically needed to teach the model what messages to ignore or handle as out-of-scope.

Option c is incorrect because schema definition creates the application structure but does not specifically handle the recognition and appropriate processing of irrelevant messages that fall outside defined intents.

Correct Answer: Fifteen to thirty examples

LUIS documentation recommends fifteen to thirty example utterances per intent to explicitly train the model effectively, providing sufficient variety for the machine learning model to recognize patterns accurately.

Option a is incorrect because five to ten examples would not provide enough variety and patterns for the machine learning model to learn to recognize different ways users might express an intent.

Option b is incorrect because three to five examples are insufficient for training a machine learning model to recognize the natural variety in how users express intents, falling below recommended minimums.

Option d is incorrect because while fifty to hundred examples would improve accuracy, the recommended minimum range for effective training is fifteen to thirty utterances, making this unnecessarily high for initial training.

Correct Answer: Privacy and Security

Privacy and Security principle directly addresses protecting user data and personally identifiable information, making it the fundamental principle for handling sensitive information like social security numbers and medical records.

Option a is incorrect because Fairness focuses on treating all people equally and avoiding bias, not on protecting sensitive personal information or securing data against unauthorized access or breaches.

Option c is incorrect because Transparency ensures AI systems are understandable to users but does not specifically address the protection of sensitive data or implementation of security protocols for information.

Option d is incorrect because Inclusiveness focuses on empowering everyone and designing for minorities, not on securing sensitive user information or protecting personally identifiable data from unauthorized access.

Correct Answer: Inclusiveness principles

Inclusiveness principles focus on empowering everyone and designing specifically for minorities and people with disabilities, directly guiding the development of accessible systems that remove barriers for diverse needs.

Option b is incorrect because Transparency requirements focus on making AI systems understandable to users, not specifically on designing for accessibility or accommodating the needs of people with disabilities.

Option c is incorrect because Accountability framework establishes human responsibility for AI decisions but does not specifically guide the design process for accommodating diverse needs and removing accessibility barriers.

Option d is incorrect because Fairness standards address treating all people equally and avoiding bias but do not specifically focus on designing systems that actively empower people with disabilities or remove barriers.