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

 

Question 221

A company wants to identify whether a customer review expresses a positive, negative, or neutral opinion. Which capability should be selected?

1. OCR
2. Object detection
3. Sentiment analysis
4. Speech synthesis

Correct Answer: 3

Explanation

Sentiment analysis is designed to identify the opinion or emotional tone expressed in text. A customer review can be analyzed to determine whether it is positive, negative, or neutral, helping organizations understand customer satisfaction at scale. OCR extracts text from images, object detection identifies and locates objects in images, and speech synthesis converts written content into spoken audio. Because the input is a written review and the requirement is to determine its expressed sentiment, sentiment analysis is the most appropriate capability for this scenario.

Question 222

A business receives invoices from several suppliers and wants to extract fields such as supplier name, invoice date, and total amount automatically. Which Azure service is designed for this workload?

1. Azure AI Document Intelligence
2. Azure AI Speech
3. Azure AI Language
4. Azure AI Vision

Correct Answer: 1

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract structured information from them. Invoice processing is a common document scenario in which fields such as supplier names, invoice dates, totals, and other relevant information can be extracted automatically. This reduces the need for manual data entry and allows extracted information to be used by downstream business applications. Azure AI Speech focuses on audio, Language focuses on text analysis, and Vision focuses on visual workloads. Document Intelligence is therefore the appropriate Azure service for automated invoice extraction.

Question 223

A machine learning model receives information such as customer age, purchase frequency, and average order value to make a prediction. What are these inputs called?

1. Labels
2. Features
3. Predictions
4. Classes

Correct Answer: 2

Explanation

Features are the input variables used by a machine learning model to learn patterns and make predictions. In this scenario, customer age, purchase frequency, and average order value provide information that can help the model determine its output. Labels represent known target values during supervised training, predictions are outputs produced by a trained model, and classes are categories used in classification problems. Identifying and preparing appropriate features is an important part of machine learning because the quality and relevance of input data can influence model performance.

Question 224

A model predicts whether a machine component will fail within the next month. The possible outputs are “fail” and “not fail.” Which task does this represent?

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

Correct Answer: 3

Explanation

Classification is used when a model predicts one category from predefined classes. In this scenario, the possible outcomes are fail and not fail, making it a binary classification problem. Historical maintenance records can be used to train the model using relevant features and known outcomes. Regression would be appropriate for predicting a numerical value, while clustering discovers groups without predefined labels. Forecasting focuses on predicting future values over time and may be related to some predictive scenarios, but the specific output here is a categorical decision.

Question 225

Which technique can group similar customers together when no predefined customer segments are available?

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

Correct Answer: 1

Explanation

Clustering is an unsupervised learning technique that identifies groups of similar observations without requiring predefined labels. A business can provide customer attributes such as purchasing behavior, spending patterns, and product preferences, allowing the algorithm to discover natural groupings. Classification requires known categories during training, while regression predicts numerical values. OCR is used to extract text from images and is unrelated to customer segmentation. Therefore, clustering is the appropriate technique when customer segments need to be discovered from the available data rather than specified in advance.

Question 226

A company wants to estimate the resale price of a vehicle using mileage, age, model, and condition. Which machine learning task is appropriate?

1. Classification
2. Regression
3. Clustering
4. Entity recognition

Correct Answer: 2

Explanation

Regression is appropriate because the target is a numerical resale price. The model can use features such as vehicle age, mileage, model, and condition to learn relationships between these inputs and the target price. Once trained, the model can estimate prices for vehicles it has not previously seen. Classification predicts categories, clustering discovers groups without predefined labels, and entity recognition identifies entities in text. Since the desired output is a continuous monetary value, regression is the appropriate machine learning task for vehicle-price prediction.

Question 227

A developer wants to process thousands of records together during a scheduled overnight job. Which inference approach is most appropriate?

1. Online inference
2. Batch inference
3. Speech recognition
4. Text-to-speech

Correct Answer: 2

Explanation

Batch inference processes a collection of records together rather than responding individually to each request. It is suitable for scheduled workloads where immediate predictions are not required, such as processing thousands of transactions overnight or generating predictions for a large historical dataset. Online inference is better suited to applications that require immediate responses to individual requests. Speech recognition and text-to-speech are AI capabilities for spoken-language workloads rather than general inference patterns. Therefore, batch inference is appropriate for the described scheduled bulk-processing scenario.

Question 228

An application must return a prediction immediately after a user submits information through a web form. Which inference approach is appropriate?

1. Batch inference
2. Offline training
3. Online inference
4. Data preprocessing

Correct Answer: 3

Explanation

Online inference is designed for scenarios in which an application needs a prediction or result shortly after receiving an individual request. A web form can send user information to a deployed model, which can process the input and return a prediction to the application. Batch inference is more suitable for scheduled processing of large collections of records. Offline training and data preprocessing occur at different stages of the machine learning lifecycle. Since the user expects an immediate response after submitting the form, online inference is the appropriate approach.

Question 229

A model predicts the amount of rainfall expected tomorrow. Which evaluation metric is suitable for measuring the average absolute difference between predicted and actual rainfall amounts?

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

Correct Answer: 1

Explanation

Mean absolute error, or MAE, measures the average absolute difference between predicted numerical values and actual values. It is suitable for regression scenarios such as predicting rainfall amounts because the target is a continuous numerical value. Precision and recall are commonly used to evaluate classification models, while accuracy measures the proportion of correct classification predictions. MAE provides an interpretable indication of how far predictions are from actual values on average. Therefore, it is an appropriate metric for evaluating a numerical rainfall prediction model.

Question 230

A classification model correctly identifies 920 out of 1,000 total records. Which metric describes this result as a percentage of all predictions?

1. Recall
2. Precision
3. Accuracy
4. F1 score

Correct Answer: 3

Explanation

Accuracy represents the proportion of all predictions that are correct. If 920 out of 1,000 records are classified correctly, the model has an accuracy of 92 percent. Accuracy considers both positive and negative classes when calculating the overall proportion of correct predictions. Precision focuses on the correctness of predicted positive cases, while recall measures the proportion of actual positive cases identified. F1 score combines precision and recall. Therefore, the percentage of all predictions that were correctly classified is represented by accuracy.

Question 231

A fraud-detection model identifies most fraudulent transactions but also incorrectly flags many legitimate transactions. Which issue is occurring?

1. High false-positive rate
2. High true-negative rate
3. Low input size
4. Excessive image resolution

Correct Answer: 1

Explanation

A false positive occurs when a model predicts that a condition exists when it does not. In fraud detection, this happens when a legitimate transaction is incorrectly identified as fraudulent. If many legitimate transactions are flagged, the model is producing a high number of false positives. This can reduce precision and may create additional work for investigators or inconvenience customers. True negatives represent legitimate transactions correctly identified as legitimate. Image resolution and input size do not describe the classification error occurring in this fraud-detection scenario.

Question 232

Which metric is particularly useful when a classification application needs to minimize missed positive cases?

1. Accuracy
2. Recall
3. Mean absolute error
4. Image confidence

Correct Answer: 2

Explanation

Recall measures the proportion of actual positive cases that a model successfully identifies. It is particularly important when missing a positive case can have significant consequences. For example, a safety-monitoring system or certain fraud-detection scenarios may prioritize identifying as many actual positive cases as possible. Accuracy provides an overall measure of correct predictions, while mean absolute error is primarily associated with regression. Image confidence is not a standard classification metric for this purpose. Therefore, recall is particularly relevant when the goal is to reduce missed positive cases.

Question 233

A generative AI application produces inaccurate information that sounds convincing and presents it as factual. What issue should the development team recognize?

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

Correct Answer: 2

Explanation

A hallucination occurs when a generative AI model produces information that is inaccurate, unsupported, or fabricated while presenting it in a plausible manner. This can be problematic when users rely on generated responses for factual information. Developers can use techniques such as grounding, retrieval, output validation, suitable prompting, and evaluation to reduce the risk and impact of unsupported responses. Clustering, regression, and classification are machine learning concepts that do not describe this behavior. The generation of convincing but unsupported information is therefore an example of hallucination.

Question 234

A developer wants a language model to use current information from an organization’s knowledge base when answering questions. What approach can provide relevant retrieved information as context?

1. Retrieval-augmented generation
2. Image classification
3. Speech synthesis
4. Regression

Correct Answer: 1

Explanation

Retrieval-augmented generation combines retrieval from an external knowledge source with generative AI. When a user submits a question, the application can search an organization’s knowledge base, retrieve relevant information, and provide that content to the language model as context. This can help the model produce answers based on current or organization-specific information rather than relying only on its original training knowledge. Image classification, speech synthesis, and regression address different workloads and do not provide the described knowledge-retrieval architecture.

Question 235

Which prompt component can establish instructions that define how a generative AI assistant should behave throughout an interaction?

1. Test dataset
2. System message
3. Image label
4. Search result

Correct Answer: 2

Explanation

A system message provides high-level instructions that can define the role, behavior, constraints, and response requirements of a generative AI assistant. It can establish rules that should guide the assistant throughout an interaction. A user message typically contains a specific request, while a test dataset is used for model evaluation and an image label represents information associated with visual data. Search results can provide retrieved context but are not the primary mechanism for defining assistant behavior. Therefore, the system message is the appropriate prompt component for persistent behavioral instructions.

Question 236

A company wants to restrict generated responses that contain inappropriate or harmful material. Which capability should be considered?

1. Content safety
2. Regression
3. Clustering
4. Feature scaling

Correct Answer: 1

Explanation

Content safety capabilities help organizations detect and manage potentially harmful or inappropriate content in AI applications. They can be used to assess user prompts and generated responses and apply suitable policies, filtering, or review processes. This is particularly important for generative AI applications that interact directly with users. Regression and clustering are machine learning techniques, while feature scaling is a data-preparation method. None of these directly addresses content moderation. Therefore, content safety should be considered when an AI application needs controls for potentially harmful generated material.

Question 237

A model performs poorly because the training dataset contains very few examples of an important group that appears frequently in production. Which issue should the team investigate?

1. Data representativeness
2. Audio volume
3. Image file extension
4. Network speed

Correct Answer: 1

Explanation

Training data should adequately represent the situations, populations, and conditions that a model is expected to encounter after deployment. If an important group is underrepresented in training data but appears frequently in production, the model may perform poorly for that group. The team should therefore investigate the representativeness and coverage of the training dataset and determine whether additional suitable examples are needed. Audio volume, image file extensions, and network speed do not directly address the underlying model-training issue described in this scenario.

Question 238

An AI system is used to make recommendations that could significantly affect customers. What practice supports responsible accountability?

1. Allowing the model to operate without review
2. Assigning clear responsibility for the system
3. Removing all model records
4. Ignoring performance changes

Correct Answer: 2

Explanation

Accountability requires clear responsibility for how an AI system is developed, deployed, monitored, and used. Organizations should establish appropriate ownership and governance so that people can review system behavior, address problems, and make decisions about changes or continued use. Allowing an AI system to operate without oversight can make it difficult to identify and correct harmful outcomes. Removing records and ignoring performance changes also weaken governance. Assigning clear responsibility helps ensure that AI systems remain subject to appropriate human and organizational oversight.

Question 239

A team wants to explain to stakeholders what information and factors influence an AI system’s predictions. Which responsible AI principle is most closely related?

1. Privacy
2. Fairness
3. Transparency
4. Inclusiveness

Correct Answer: 3

Explanation

Transparency involves providing appropriate information that helps users and stakeholders understand how an AI system operates and produces results. Depending on the system, this can include information about inputs, limitations, intended use, or factors that influence predictions. Privacy focuses on protecting personal information, fairness addresses equitable treatment and outcomes, and inclusiveness focuses on supporting diverse users. When stakeholders need understandable information about how an AI system reaches or supports its results, transparency is the responsible AI principle most directly related to the requirement.

Question 240

A machine learning model performs poorly on both its training data and unseen evaluation data. Which issue may indicate that the model is too simple to learn the important relationships?

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

Correct Answer: 2

Explanation

Underfitting occurs when a model is not sufficiently capable of learning the important patterns in the training data. As a result, its performance can be poor on both training examples and unseen evaluation data. A model that is too simple, insufficiently trained, or based on inadequate features may underfit the problem. Overfitting produces a different pattern in which training performance is often high while performance on new data is poor. Hallucination relates to generative AI output, while data leakage concerns inappropriate information crossing into model training or evaluation.