Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 45

Question 440

The app needs lower bandwidth audio for a call-center line. Which configuration should be considered?

  1. Pin a supported voice identifier and test voice changes before rollout
  2. Configure supported source-language identification or the correct input speech locale
  3. Select a supported compressed audio output format suitable for the playback channel
  4. Compare the returned media format against mobile decoder support

Explanation

Correct Answer

C:

Synthesis quality and file size depend on the output encoding; text normalization cannot replace audio-format selection. In the hands-free multilingual service desk, confirm the fix by retesting the affected behavior after selecting a supported compressed audio output format suitable for the playback channel.

Incorrect Answers

A:

Although useful elsewhere in the hands-free multilingual service desk, voice choice affects audible identity; uncontrolled defaults may shift pronunciation and style; the present case is different and calls for the team to select a supported compressed audio output format suitable for the playback channel.

B:

Translation quality depends on accurate source-speech recognition before target rendering in its own workflow, but that is a different correction from the present requirement to select a supported compressed audio output format suitable for the playback channel. The reported issue would remain.

D:

That change addresses another failure mode: platform codec support varies; a valid synthesis stream may still be unplayable by a particular client; it does not address this case, for which the relevant intervention is to select a supported compressed audio output format suitable for the playback channel.

Question 441

One rare intent has only two labeled examples and performs poorly. What data work is useful?

  1. Gather additional correctly labeled examples representing that intent's real usage
  2. Investigate overfitting and mismatched or insufficient training examples
  3. Split the broad intent into actionable categories if the workflows require different handling
  4. Validate required extracted entities and ask for missing parameters before tool execution

Explanation

Correct Answer

A:

A class with very sparse representative data often generalizes poorly relative to common intents. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after gathering additional correctly labeled examples representing that intent's real usage.

Incorrect Answers

B:

Although useful elsewhere in the service-request virtual assistant, memorizing known examples cannot guarantee correct predictions on new conversations; the present case is different and calls for the team to gather additional correctly labeled examples representing that intent's real usage.

C:

An intent taxonomy should correspond to meaningful application decisions, not just an umbrella department name in its own workflow, but that is a different correction from the present requirement to gather additional correctly labeled examples representing that intent's real usage. The reported issue would remain.

D:

That change addresses another failure mode: entity predictions do not guarantee the full set of business-required inputs is present; it does not address this case, for which the relevant intervention is to gather additional correctly labeled examples representing that intent's real usage.

Question 442

A user says “move the meeting to Thursday,” and the bot recognizes rescheduling but loses the new date. What training annotation is needed?

  1. Define distinct intents corresponding to the different user actions
  2. Label the date expression as an entity in representative CLU utterances
  3. Evaluate entity extraction separately from intent classification
  4. Deploy the intended trained model version to the prediction endpoint

Explanation

Correct Answer

B:

Intent identifies the action; entities supply variable parameters such as dates. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after labeling the date expression as an entity in representative CLU utterances.

Incorrect Answers

A:

CLU intents represent the purpose of an utterance; conflating different actions prevents reliable routing in its own workflow, but that is a different correction from the present requirement to label the date expression as an entity in representative CLU utterances. The reported issue would remain.

C:

Although useful elsewhere in the service-request virtual assistant, cLU has distinct intent and entity predictions; success at one task does not prove the other works; the present case is different and calls for the team to label the date expression as an entity in representative CLU utterances.

D:

That change addresses another failure mode: training and local tests do not necessarily replace the deployed runtime model; it does not address this case, for which the relevant intervention is to label the date expression as an entity in representative CLU utterances.

Question 443

A development team compares model versions using different held-out samples and declares the higher score a winner. How should it compare fairly?

  1. Align the prediction request with an existing named CLU deployment
  2. Inspect intent predictions and confusion examples in held-out evaluation
  3. Use the same versioned evaluation split and metrics for both model candidates
  4. Apply a consistent entity-labeling policy across representative training utterances

Explanation

Correct Answer

C:

Comparable evaluation controls test-set variation before attributing differences to training improvements. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after using the same versioned evaluation split and metrics for both model candidates.

Incorrect Answers

A:

Runtime calls must identify a deployed model; arbitrary version strings are not automatically valid deployment names in its own workflow, but that is a different correction from the present requirement to use the same versioned evaluation split and metrics for both model candidates. The reported issue would remain.

B:

That change addresses another failure mode: intent confusion must be examined across real phrasings; adding random entities will not correct a mislabeled purpose; it does not address this case, for which the relevant intervention is to use the same versioned evaluation split and metrics for both model candidates.

D:

Although useful elsewhere in the service-request virtual assistant, inconsistent annotations teach the model unstable extraction boundaries and meanings; the present case is different and calls for the team to use the same versioned evaluation split and metrics for both model candidates.

Question 444

A test set contains paraphrases copied from the same training conversations. What evaluation risk exists?

  1. Use the same versioned evaluation split and metrics for both model candidates
  2. Apply a consistent entity-labeling policy across representative training utterances
  3. Create held-out conversations and speakers to avoid near-identical training leakage
  4. Inspect intent predictions and confusion examples in held-out evaluation

Explanation

Correct Answer

C:

Highly correlated examples inflate intent accuracy and conceal failures on unseen input forms. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after creating held-out conversations and speakers to avoid near-identical training leakage.

Incorrect Answers

A:

Although useful elsewhere in the service-request virtual assistant, comparable evaluation controls test-set variation before attributing differences to training improvements; the present case is different and calls for the team to create held-out conversations and speakers to avoid near-identical training leakage.

B:

That change addresses another failure mode: inconsistent annotations teach the model unstable extraction boundaries and meanings; it does not address this case, for which the relevant intervention is to create held-out conversations and speakers to avoid near-identical training leakage.

D:

Intent confusion must be examined across real phrasings; adding random entities will not correct a mislabeled purpose in its own workflow, but that is a different correction from the present requirement to create held-out conversations and speakers to avoid near-identical training leakage. The reported issue would remain.

Question 445

The bot references deployment `prod-v2`, but the CLU resource exposes only `prod`. What should change?

  1. Create held-out conversations and speakers to avoid near-identical training leakage
  2. Inspect intent predictions and confusion examples in held-out evaluation
  3. Apply a consistent entity-labeling policy across representative training utterances
  4. Align the prediction request with an existing named CLU deployment

Explanation

Correct Answer

D:

Runtime calls must identify a deployed model; arbitrary version strings are not automatically valid deployment names. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after aligning the prediction request with an existing named CLU deployment.

Incorrect Answers

A:

Highly correlated examples inflate intent accuracy and conceal failures on unseen input forms in its own workflow, but that is a different correction from the present requirement to align the prediction request with an existing named CLU deployment. The reported issue would remain.

B:

Although useful elsewhere in the service-request virtual assistant, intent confusion must be examined across real phrasings; adding random entities will not correct a mislabeled purpose; the present case is different and calls for the team to align the prediction request with an existing named CLU deployment.

C:

That change addresses another failure mode: inconsistent annotations teach the model unstable extraction boundaries and meanings; it does not address this case, for which the relevant intervention is to align the prediction request with an existing named CLU deployment.

Question 446

A class performs well on training utterances but poorly on validation. What failure pattern is likely?

  1. Validate required extracted entities and ask for missing parameters before tool execution
  2. Retain known-good deployment metadata and manage promotion and rollback explicitly
  3. Investigate overfitting and mismatched or insufficient training examples
  4. Gather additional correctly labeled examples representing that intent's real usage

Explanation

Correct Answer

C:

Memorizing known examples cannot guarantee correct predictions on new conversations. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after investigating overfitting and mismatched or insufficient training examples.

Incorrect Answers

A:

Entity predictions do not guarantee the full set of business-required inputs is present in its own workflow, but that is a different correction from the present requirement to investigate overfitting and mismatched or insufficient training examples. The reported issue would remain.

B:

That change addresses another failure mode: versioned deployment management allows reverting application routing without reconstructing past training samples; it does not address this case, for which the relevant intervention is to investigate overfitting and mismatched or insufficient training examples.

D:

Although useful elsewhere in the service-request virtual assistant, a class with very sparse representative data often generalizes poorly relative to common intents; the present case is different and calls for the team to investigate overfitting and mismatched or insufficient training examples.

Question 447

A service-request virtual assistant must extract variable details such as a date or tracking number from a request. Reviewers will accept the implementation when recognized responses include the required parameter span. Which approach best fulfills that requirement?

  1. Define an intent in a CLU project
  2. Train and evaluate a CLU model
  3. Define and label an entity in CLU
  4. Deploy the trained CLU model

Explanation

Correct Answer

C:

CLU entities capture parts of utterances that provide parameters relevant to the predicted intent. For this service-request virtual assistant, the intended evidence is that recognized responses include the required parameter span. Entity annotation does not by itself specify which action the user wants.

Incorrect Answers

A:

An intent describes a category of user purpose that a trained conversational model predicts from utterances; it may be used to represent the distinct user action that the dialog should recognize in a service-request virtual assistant, but the current requirement is instead to extract variable details such as a date or tracking number from a request. That separate capability does not prove that recognized responses include the required parameter span; An intent is not the name of a person extracted from the request, so selecting that alternative would not meet the stated acceptance check.

B:

Model training fits learned patterns from labeled examples, while evaluation compares predictions on held-out data; it may be used to assess whether its intent and entity predictions generalize before exposing it to clients in a service-request virtual assistant, but the current requirement is instead to extract variable details such as a date or tracking number from a request. That separate capability does not prove that recognized responses include the required parameter span; Publishing a project with unreviewed labels does not establish reliable intent detection, so selecting that alternative would not meet the stated acceptance check.

D:

A CLU deployment publishes an available trained model version for runtime use through the appropriate project endpoint; it may be used to expose a named deployment for application prediction requests in a service-request virtual assistant, but the current requirement is instead to extract variable details such as a date or tracking number from a request. That separate capability does not prove that recognized responses include the required parameter span; A local training job alone cannot answer live client requests, so selecting that alternative would not meet the stated acceptance check.

Question 448

A customer says “track my package” but the bot interprets it as “create shipment.” What diagnostic helps?

  1. Align the prediction request with an existing named CLU deployment
  2. Inspect intent predictions and confusion examples in held-out evaluation
  3. Create held-out conversations and speakers to avoid near-identical training leakage
  4. Use the same versioned evaluation split and metrics for both model candidates

Explanation

Correct Answer

B:

Intent confusion must be examined across real phrasings; adding random entities will not correct a mislabeled purpose. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after inspecting intent predictions and confusion examples in held-out evaluation.

Incorrect Answers

A:

Runtime calls must identify a deployed model; arbitrary version strings are not automatically valid deployment names in its own workflow, but that is a different correction from the present requirement to inspect intent predictions and confusion examples in held-out evaluation. The reported issue would remain.

C:

That change addresses another failure mode: highly correlated examples inflate intent accuracy and conceal failures on unseen input forms; it does not address this case, for which the relevant intervention is to inspect intent predictions and confusion examples in held-out evaluation.

D:

Although useful elsewhere in the service-request virtual assistant, comparable evaluation controls test-set variation before attributing differences to training improvements; the present case is different and calls for the team to inspect intent predictions and confusion examples in held-out evaluation.

Question 449

A tracking-number entity is labeled in one sample but inconsistently omitted from similar samples. What should the trainer do?

  1. Apply a consistent entity-labeling policy across representative training utterances
  2. Use the same versioned evaluation split and metrics for both model candidates
  3. Create held-out conversations and speakers to avoid near-identical training leakage
  4. Align the prediction request with an existing named CLU deployment

Explanation

Correct Answer

A:

Inconsistent annotations teach the model unstable extraction boundaries and meanings. In the service-request virtual assistant, confirm the fix by retesting the affected behavior after applying a consistent entity-labeling policy across representative training utterances.

Incorrect Answers

B:

Comparable evaluation controls test-set variation before attributing differences to training improvements in its own workflow, but that is a different correction from the present requirement to apply a consistent entity-labeling policy across representative training utterances. The reported issue would remain.

C:

That change addresses another failure mode: highly correlated examples inflate intent accuracy and conceal failures on unseen input forms; it does not address this case, for which the relevant intervention is to apply a consistent entity-labeling policy across representative training utterances.

D:

Although useful elsewhere in the service-request virtual assistant, runtime calls must identify a deployed model; arbitrary version strings are not automatically valid deployment names; the present case is different and calls for the team to apply a consistent entity-labeling policy across representative training utterances.

Question 450

A release for a service-request virtual assistant is blocked until the application can expose a named deployment for application prediction requests. The tester will verify that an application can submit utterances to the deployment and receive predictions. What should be implemented?

  1. Add varied labeled training utterances
  2. Define and label an entity in CLU
  3. Train and evaluate a CLU model
  4. Define an intent in a CLU project
  5. Deploy the trained CLU model

Explanation

Correct Answer

E:

A CLU deployment publishes an available trained model version for runtime use through the appropriate project endpoint. For this service-request virtual assistant, the intended evidence is that an application can submit utterances to the deployment and receive predictions. A local training job alone cannot answer live client requests.

Incorrect Answers

A:

Labeled utterances supply examples of the language associated with intents and entities for model training; it may be used to teach a CLU model how real users phrase each supported intent in a service-request virtual assistant, but the current requirement is instead to expose a named deployment for application prediction requests. That separate capability does not prove that an application can submit utterances to the deployment and receive predictions; Adding labels to a deployment endpoint is not a replacement for training data, so selecting that alternative would not meet the stated acceptance check.

B:

CLU entities capture parts of utterances that provide parameters relevant to the predicted intent; it may be used to extract variable details such as a date or tracking number from a request in a service-request virtual assistant, but the current requirement is instead to expose a named deployment for application prediction requests. That separate capability does not prove that an application can submit utterances to the deployment and receive predictions; Entity annotation does not by itself specify which action the user wants, so selecting that alternative would not meet the stated acceptance check.

C:

Model training fits learned patterns from labeled examples, while evaluation compares predictions on held-out data; it may be used to assess whether its intent and entity predictions generalize before exposing it to clients in a service-request virtual assistant, but the current requirement is instead to expose a named deployment for application prediction requests. That separate capability does not prove that an application can submit utterances to the deployment and receive predictions; Publishing a project with unreviewed labels does not establish reliable intent detection, so selecting that alternative would not meet the stated acceptance check.

D:

An intent describes a category of user purpose that a trained conversational model predicts from utterances; it may be used to represent the distinct user action that the dialog should recognize in a service-request virtual assistant, but the current requirement is instead to expose a named deployment for application prediction requests. That separate capability does not prove that an application can submit utterances to the deployment and receive predictions; An intent is not the name of a person extracted from the request, so selecting that alternative would not meet the stated acceptance check.

Designing and Implementing a Microsoft Azure AI Solution (AI-102) Practice Tests