View Full Microsoft AI-900 Exam Dumps and Practice Test Dumps.

 

Question 301

A company wants to identify whether each incoming support request should be routed to billing, technical support, or account management. Which machine learning task is most appropriate?

1. Regression
2. Clustering
3. Forecasting
4. Classification

Correct Answer: 4

Explanation

Classification is used when a model must assign an input to one of several predefined categories. In this scenario, support requests need to be routed to billing, technical support, or account management, so these categories can serve as the possible outputs. The model can learn from previously labeled support requests and then classify new requests based on their content and other relevant features. Regression predicts numerical values, clustering discovers groups without predefined labels, and forecasting focuses on future values. Therefore, classification is the appropriate machine learning task.

Question 302

Which Azure service provides capabilities for analyzing images, including extracting visual information from photographs?

1. Azure AI Vision
2. Azure AI Speech
3. Azure AI Language
4. Azure AI Search

Correct Answer: 1

Explanation

Azure AI Vision provides capabilities for analyzing visual information contained in images. Depending on the scenario, it can support workloads such as image analysis, optical character recognition, image tagging, object-related analysis, and image descriptions. Azure AI Speech focuses on spoken-language processing, Azure AI Language analyzes written language, and Azure AI Search provides search and indexing capabilities. When an application needs to understand or extract information from photographs or other visual content, Azure AI Vision is the relevant Azure service.

Question 303

A training dataset contains customer age, income, account type, and the known result of whether each customer renewed a subscription. Which item is the target label?

1. Customer age
2. Income
3. Account type
4. Renewal result

Correct Answer: 4

Explanation

The target label is the known outcome that a supervised machine learning model is trained to predict. In this scenario, the model uses customer age, income, and account type as input features and learns to predict whether the customer renewed the subscription. Therefore, the renewal result is the target label. Features provide information used to make predictions, while the label represents the expected output during training. Correctly identifying the target label is important when preparing a supervised learning dataset.

Question 304

A business wants to automatically translate customer messages written in German into English. Which AI capability is directly required?

1. Sentiment analysis
2. Machine translation
3. Object detection
4. OCR

Correct Answer: 2

Explanation

Machine translation converts text from one language into another. In this scenario, German customer messages need to be converted into English, so translation is the required language-processing capability. Sentiment analysis would determine the emotional tone of a message, object detection analyzes visual objects, and OCR extracts text from images. Translation can be used as part of a larger customer-service workflow, where translated messages may subsequently be analyzed for intent, sentiment, or other characteristics. The direct requirement here is changing written content from German into English.

Question 305

A data scientist divides available labeled data into training, validation, and test datasets. What is one purpose of keeping a separate test dataset?

1. To provide an unbiased final assessment of model performance
2. To increase the number of features automatically
3. To replace the training process
4. To convert categorical data into images

Correct Answer: 1

Explanation

A separate test dataset can provide a final assessment of how well a trained model performs on data that was not used during model development. Keeping test data separate helps provide a more realistic indication of generalization to unseen examples. Training data is used to learn model parameters, while validation data can support model selection or tuning. The test dataset should generally remain separate until final evaluation. It does not automatically increase features, replace training, or convert data into images.

Question 306

A voice-enabled application must recognize a spoken command and then respond using synthesized audio. Which combination of capabilities is required?

1. OCR and object detection
2. Image classification and translation
3. Speech-to-text and text-to-speech
4. Regression and clustering

Correct Answer: 3

Explanation

Speech-to-text can convert the user’s spoken command into written text so the application can interpret and process it. After generating an appropriate response, text-to-speech can convert that response back into spoken audio. Together, these capabilities support a conversational voice interface. OCR and object detection are visual workloads, while image classification does not process spoken commands. Regression and clustering are machine learning techniques that do not directly provide the required speech input and audio output functionality. Therefore, speech-to-text and text-to-speech form the appropriate combination.

Question 307

A company wants to find unusual transactions that differ significantly from normal transaction patterns without having examples of known fraud. Which machine learning approach may be useful?

1. Anomaly detection
2. Text summarization
3. Speech synthesis
4. Image captioning

Correct Answer: 1

Explanation

Anomaly detection is designed to identify observations that differ significantly from expected or normal patterns. It can be useful when an organization does not have a complete set of labeled examples showing every type of fraudulent transaction. A financial system can learn patterns associated with normal transactions and flag unusual activity for additional investigation. Text summarization creates shorter versions of written content, speech synthesis produces spoken audio, and image captioning describes images. Therefore, anomaly detection is an appropriate approach for identifying unusual transactions when labeled fraud examples are limited.

Question 308

A company wants to create a searchable collection of product manuals so users can quickly find relevant sections. Which capability is most directly associated with preparing content for search and retrieval?

1. Text-to-speech
2. Search indexing
3. Sentiment analysis
4. Regression

Correct Answer: 2

Explanation

Search indexing organizes content so that a search system can efficiently locate relevant information when users submit queries. Product manuals can be indexed to make their text searchable and support retrieval of relevant sections. Azure AI Search can provide indexing and search capabilities for applications that need to work with large collections of content. Text-to-speech converts text into audio, sentiment analysis evaluates opinions, and regression predicts numerical values. Therefore, search indexing is directly relevant when preparing a collection of manuals for efficient search and retrieval.

Question 309

A model is being tested for a safety application where failing to identify an actual dangerous event could have serious consequences. Which metric may be especially important?

1. Precision
2. Accuracy
3. Recall
4. Mean absolute error

Correct Answer: 3

Explanation

Recall measures the proportion of actual positive cases that a model successfully identifies. In a safety-related application, missing a real dangerous event can have serious consequences, so the team may place particular importance on identifying as many actual positive cases as possible. Precision instead measures how many predicted positive cases are actually positive. Accuracy provides an overall proportion of correct classifications, while mean absolute error is commonly used for numerical prediction. The appropriate metric ultimately depends on the application’s requirements, but recall is especially relevant when missed positives are costly.

Question 310

A developer wants to make a generative AI response more specific by telling the model its role, the expected format, and important constraints. What should the developer provide?

1. A clear prompt
2. A test dataset
3. A confusion matrix
4. A numerical feature

Correct Answer: 1

Explanation

A clear prompt can communicate the task, role, output format, constraints, and other relevant instructions to a generative AI model. Providing detailed and well-structured instructions can help the model understand what the application expects from the response. A test dataset is used to evaluate models, a confusion matrix summarizes classification results, and a numerical feature is an input variable used by a machine learning model. Therefore, when a developer wants to guide the content and format of a generated response, a clear prompt is appropriate.

Question 311

An organization wants to determine whether a customer review contains references to a specific company, product, or location. Which language capability can identify these entities?

1. Entity recognition
2. Image segmentation
3. Regression
4. Speech synthesis

Correct Answer: 1

Explanation

Entity recognition identifies meaningful entities within text and can categorize them according to supported entity types. A customer review may contain references to companies, products, locations, people, dates, or other recognizable entities. Identifying these entities can help organizations organize information, improve search, or perform further analysis. Image segmentation works with visual regions, regression predicts numerical values, and speech synthesis generates spoken audio. Because the requirement involves finding named entities within a written customer review, entity recognition is the appropriate natural language capability.

Question 312

A company needs to process a large number of documents every night and does not require results immediately for individual documents. Which inference pattern is appropriate for a model-based processing workload?

1. Online inference
2. Real-time speech recognition
3. Batch inference
4. Interactive translation

Correct Answer: 3

Explanation

Batch inference is designed to process large collections of inputs together, often according to a schedule. A company can use it to process documents overnight when immediate responses are not required for each individual document. This approach is different from online inference, where applications typically send individual requests and expect responses quickly. Real-time speech recognition and interactive translation address specific conversational or language scenarios rather than the general bulk-processing pattern described. Therefore, batch inference is the appropriate approach for scheduled overnight processing of many documents.

Question 313

An AI system is evaluated separately for different demographic groups to determine whether its performance differs significantly between them. Which responsible AI principle is being considered?

1. Fairness
2. Encryption
3. Regression
4. Speech processing

Correct Answer: 1

Explanation

Fairness involves considering whether an AI system provides equitable treatment and performance across relevant groups. Evaluating model results separately for different demographic groups can help identify disparities in outcomes or error rates that may require further investigation. Encryption protects data, regression is a machine learning task for numerical prediction, and speech processing handles audio. Fairness evaluation should be based on the context, relevant groups, and appropriate measures for the specific application. In this scenario, comparing performance across groups is directly related to the fairness principle.

Question 314

A company wants to use AI to determine whether photos uploaded by customers contain acceptable or inappropriate visual content. Which general AI workload is most relevant?

1. Natural language processing
2. Computer vision
3. Regression
4. Time-series forecasting

Correct Answer: 2

Explanation

Computer vision is the AI workload used to analyze and interpret visual information such as photographs and video. An application can use computer vision capabilities to examine uploaded images and support content moderation workflows. The exact implementation may involve image analysis or other visual classification techniques depending on the requirements and safety policies. Natural language processing focuses on text and language, regression predicts numerical values, and time-series forecasting predicts future values based on temporal patterns. Because the input consists of customer photographs, computer vision is the most relevant workload.

Question 315

A company wants to automatically extract structured fields from its own unique forms, which differ substantially from standard document types. Which approach may be required?

1. Custom Document Intelligence model
2. Speech-to-text
3. Sentiment analysis
4. Image captioning

Correct Answer: 1

Explanation

A custom Document Intelligence model can be appropriate when an organization needs to extract fields from specialized documents that are not adequately handled by available prebuilt models. The organization can provide representative documents and identify the fields it needs to extract, allowing a model to be tailored to the document structure and business requirements. Speech-to-text processes spoken language, sentiment analysis evaluates opinions in text, and image captioning generates descriptions of visual content. Therefore, a custom Document Intelligence model may be appropriate for specialized organizational forms.

Question 316

A model is designed to predict a numerical monthly revenue value. Which evaluation metric measures the average magnitude of prediction errors without considering whether the errors are positive or negative?

1. Recall
2. Precision
3. Accuracy
4. Mean absolute error

Correct Answer: 4

Explanation

Mean absolute error measures the average absolute difference between predicted and actual numerical values. Because the absolute value is used, positive and negative errors do not cancel each other out. This makes MAE useful for evaluating regression models that predict quantities such as monthly revenue. Precision and recall are generally classification metrics, while accuracy is also commonly used for categorical predictions. Therefore, when the requirement is to measure the average magnitude of numerical prediction errors, mean absolute error is the appropriate metric.

Question 317

An AI assistant should refuse requests that conflict with defined application rules while still following permitted user instructions. Which mechanism can establish high-level behavioral constraints?

1. System instructions
2. Image resolution
3. Training labels
4. Regression output

Correct Answer: 1

Explanation

System instructions can establish high-level rules, behavioral constraints, and application requirements for a generative AI assistant. They can specify the assistant’s role, limitations, response style, and other instructions that should guide its behavior. Training labels are associated with supervised machine learning, image resolution concerns visual data, and regression output represents a numerical prediction. System instructions do not eliminate the need for other safety controls, but they can provide an important layer of behavioral guidance. Therefore, they are appropriate for establishing application-level constraints.

Question 318

A company wants to identify the main topics discussed across thousands of customer comments without manually reading every comment. Which language capability can help extract important concepts?

1. Object detection
2. Key phrase extraction
3. Speech recognition
4. Image classification

Correct Answer: 2

Explanation

Key phrase extraction identifies important words and phrases that represent the main concepts in written content. When applied to thousands of customer comments, it can help an organization discover recurring topics without requiring every comment to be manually reviewed. The extracted phrases can support search, categorization, reporting, and further analysis. Object detection and image classification work with visual data, while speech recognition converts spoken audio into text. Because the requirement involves identifying important concepts within written comments, key phrase extraction is the appropriate capability.

Question 319

A generative AI application retrieves a relevant passage from an internal knowledge repository before generating an answer. What role does the retrieved passage primarily serve?

1. It acts as grounding information
2. It replaces the model’s tokenizer
3. It changes an image into audio
4. It performs numerical regression

Correct Answer: 1

Explanation

Retrieved information can provide grounding context for a generative AI model. By supplying a relevant passage from an internal knowledge repository, the application gives the model information that can help it produce a response based on a trusted or organization-specific source. This approach is commonly associated with retrieval-augmented generation. The retrieved passage does not replace tokenization, convert images into audio, or perform regression. Therefore, its primary role is to provide grounding information that supports a more contextually relevant generated answer.

Question 320

A model performs poorly during both training and evaluation because it does not have enough complexity to capture important patterns in the data. Which issue is most likely?

1. Overfitting
2. Data drift
3. Underfitting
4. Hallucination

Correct Answer: 3

Explanation

Underfitting occurs when a model is too simple or otherwise unable to learn important relationships in the training data. As a result, the model can perform poorly on both training data and unseen evaluation data. This differs from overfitting, where training performance is typically strong but performance on new data is significantly weaker. Data drift concerns changes in real-world input distributions after deployment, while hallucination describes unsupported or inaccurate generative AI output. Therefore, poor performance on both training and evaluation data due to insufficient model complexity suggests underfitting.