Text Analysis practice questions

The document contains an equal number of uppercase and lowercase words.

Why it fails: Casing does not determine mixed sentiment.

The document length exceeds the 5,120 character limit.

Why it fails: Exceeding length limits results in a request error.

The language detection confidence score is below 0.5.

Why it fails: Low language confidence results in an error or fallback, not a mixed sentiment score.

LanguageService

Why it fails: LanguageService is not the correct deployment kind string.

TextUnderstanding

Why it fails: TextUnderstanding is not a valid resource kind.

CognitiveServices

Why it fails: CognitiveServices is a legacy multi-service resource kind.

documentLengthLimit

Why it fails: documentLengthLimit is not a configuration parameter.

sentenceCount

Why it fails: sentenceCount is used for extractive summarization, not abstractive length control.

confidenceThreshold

Why it fails: confidenceThreshold filters classification scores, not summary length.

1,000,000 characters

Why it fails: Azure AI Language synchronous document inputs are capped at 5,120 characters per document, so 1,000,000 exceeds the limit and the request would be rejected. The larger figure is tempting because it resembles asynchronous or batch limits, but synchronous single-document calls enforce the smaller 5,120-character ceiling.

50,000 characters

Why it fails: 50,000 exceeds the synchronous single document limit.

1,024 characters

Why it fails: 1,024 is too low; the limit is higher.

Switching from synchronous to asynchronous batch processing.

Why it fails: Batch processing changes execution mode, not data residency or model training policies.

Using the Free (F0) pricing tier.

Why it fails: The Free tier does not support enterprise security controls like CMK or private endpoints.

Enabling abstractive summarization with cross-region replication.

Why it fails: Cross-region replication sends data to other regions, violating strict regional residency mandates.

Increase the confidence score threshold to 0.99.

Why it fails: Threshold adjustments filter out lower-confidence scores but do not alter schema overlap handling.

Switch the project type from custom NER to custom text classification.

Why it fails: Classification categorizes documents; it does not extract entities.

Convert all contract texts to lowercase before tagging.

Why it fails: Case sensitivity settings can be adjusted, but lowercasing does not resolve overlapping structural tags.

/keyphrases

Why it fails: The keyphrases endpoint extracts key phrases.

/language/:analyze-text

Why it fails: The unified text analysis endpoint is used for tasks like sentiment and NER, but language detection has its own dedicated endpoint or task parameter.

/sentiment

Why it fails: The sentiment endpoint is for sentiment analysis.

summaryLengthLimit

Why it fails: summaryLengthLimit is not the correct parameter name.

topKeyPhrases

Why it fails: topKeyPhrases is not a summarization parameter.

minConfidenceScore

Why it fails: minConfidenceScore does not control extractive sentence counts.

Azure.AI.Language.Text

Why it fails: This package name is incorrect.

Microsoft.CognitiveServices.TextAnalytics

Why it fails: This is incorrect.

Microsoft.Azure.CognitiveServices.Language

Why it fails: This is the legacy v3 SDK NuGet package.

In the HTTP request headers of every API call.

Why it fails: Case sensitivity is a model training and project property, not an HTTP header.

In the Azure Monitor diagnostics logs.

Why it fails: Diagnostic logs record runtime telemetry.

In the Azure portal subscription access control blade.

Why it fails: Access control manages permissions, not NER case sensitivity.

1 document

Why it fails: Batch requests support multiple documents.

1,000 documents

Why it fails: 1,000 exceeds synchronous single-call payload document limits.

100,000 documents

Why it fails: The language detection endpoint accepts a maximum of 1,000 documents per request, so 100,000 exceeds the payload limit and the call would be rejected. Batching large volumes is tempting because it reduces API calls, but the correct ceiling is 1,000 documents per request.

Estimated reading time in minutes

Why it fails: Reading time is a derived presentation metric, not a field in the sentiment analysis response, which returns sentiment labels, confidence scores and sentence opinions. It is tempting because reading time appears in document-processing tools, and would be correct for a document summarisation or content-extraction service.

Detected language translation confidence score

Why it fails: Sentiment analysis does not translate text, so no translation confidence score is returned; the response contains sentiment and confidence values only. It is tempting because confidence scores do appear in the schema, and would be correct for the Translator service's language detection output.

Total word count of the document

Why it fails: Sentiment analysis returns confidence scores for positive, neutral and negative, plus sentence-level opinions; word count is not part of that JSON schema. It is tempting because word count is a common text-analytics statistic, and would be the correct output for a text summarisation or readability feature.

100 Continue

Why it fails: 100 is an informational status code, not an error.

200 OK

Why it fails: 200 indicates success.

Manually hardcode all possible classification outputs in client application code.

Why it fails: Hardcoding defeats the purpose of machine learning classification.

Delete all training data across every category.

Why it fails: Deleting all data destroys the model.

Set the 'model-version' parameter to the latest preview version to ensure Spanish is supported.

Why it fails: Spanish is supported by the generally available language detection model, so pinning a preview model version is unnecessary and can introduce instability. The routing decision depends on the returned language and score, not on forcing a preview model for a widely supported language.

Set the 'kind' property to 'LanguageDetection' and include the 'loggingOptOut' flag to suppress telemetry.

Why it fails: While the request identifies the task kind, the 'loggingOptOut' flag controls whether request data is logged for model improvement and has no bearing on detecting Spanish versus English. Enabling it does not improve detection, and it is not required to route documents to the correct analyzer.

Send the documents in batches of exactly one document per request to guarantee language detection accuracy.

Why it fails: Language detection accepts multiple documents per request, and batching one document at a time adds latency and cost without improving accuracy. Detection quality depends on text length and the optional country hint, not on forcing a single-document payload, so this action is unnecessary.

It is a numeric average of the sentence confidence scores converted into a single polarity band by fixed thresholds.

Why it fails: The output is not produced by averaging sentence scores against fixed numeric bands. The service returns a categorical label plus separate confidence scores for positive, neutral, and negative at both sentence and document level. Describing it as a thresholded average misrepresents how the label is generated.

It is the sentiment of the longest sentence, since that sentence is assumed to carry the most information.

Why it fails: Sentence length plays no role in selecting the document label. Every sentence is assessed and contributes to the document-level result. A short but emphatic complaint can outweigh a long neutral description, so tying the outcome to the longest sentence is incorrect.

It is always the sentiment of the first sentence, because the service anchors the document result to the opening statement.

Why it fails: There is no anchoring to the first sentence. The service evaluates the full document and its individual sentences, then produces a document-level label from the combined evidence. A document that opens with a complaint but is overwhelmingly positive later would not be labeled negative by this logic.

Sentiment analysis

Why it fails: Sentiment analysis returns a positive, negative or neutral confidence score for opinions expressed in text; it never identifies which language the ticket is written in. It is tempting because it also analyses free-text tickets, but it would be correct for triaging angry customers or prioritising complaints.

Named Entity Recognition (NER)

Why it fails: Named Entity Recognition extracts people, places and organisations from text, not the language itself, so it cannot determine which regional team should receive a ticket. It is tempting because it also processes multilingual text, but it would be the right call for pulling customer names or locations out of a ticket.

Key phrase extraction

Why it fails: Key phrase extraction returns salient terms from text; it does not identify which language the ticket is written in, so routing cannot be driven by its output. It is tempting because it analyses ticket text, but language detection is the feature that returns the detected language for routing decisions.

Switch the API endpoint to use the Free pricing tier.

Why it fails: Switching to the Free tier reduces the request quota further, so throttling persists or worsens; 429 signals rate limiting, which the application must handle by honouring the Retry-After header and backing off. The Free tier suits low-volume development and testing, not production throughput remediation.

Regenerate the subscription key and restart the client service.

Why it fails: A 429 signals rate limiting, so rotating the key does not restore throughput and restarting the service discards state. It is tempting because 401 and 403 errors do involve credentials, but the correct handling is to honour the Retry-After header and retry with exponential backoff.

Immediately re-submit the request with doubled payload size.

Why it fails: Doubling the payload increases resource consumption per call, sustaining or deepening the throttling that triggered the 429. Rate limiting requires honouring the Retry-After header with exponential backoff and jitter. Larger payloads suit batch document submission where per-request quotas are not the constraint.

Key phrase extraction

Why it fails: Key phrase extraction returns salient phrases without assigning entity categories, so medications and dosages are not distinguished. It suits surfacing topics from unstructured documents. Categorising medical terms, medications and dosages requires the healthcare domain entity recognition capability in Azure AI Language.

Standard Named Entity Recognition (NER)

Why it fails: Standard NER recognises general entities such as people, places and organisations, and does not categorise medications or dosages. It suits generic entity extraction from ordinary text. Medical intake forms require the healthcare domain model in Azure AI Language, which adds medication and dosage entity types.

Custom text classification

Why it fails: Custom text classification assigns whole-document or per-chunk labels from your trained taxonomy; it does not extract individual terms with categories. It suits routing documents into predefined classes. Identifying and categorising medical terms, medications and dosages requires entity extraction, not document-level labelling.

Personal Personally Identifiable Information (PII) detection

Why it fails: PII detection identifies and redacts personal data such as names and phone numbers; it does not surface topics or main ideas. It is tempting because both are Azure AI Language features, but key phrase extraction is the capability that returns salient concepts from long articles.

Language detection

Why it fails: Language detection returns the dominant language and confidence score for input text; it extracts no concepts. It is tempting because it is a foundational Azure AI Language feature, but key phrase extraction is the one that surfaces main ideas from long-form articles.

Summarization

Why it fails: Summarization condenses text but does not return the discrete key-concept entities and topics the task requires; that is key phrase extraction. It tempts because it genuinely handles long-form articles, and would be correct if the ask were an abstract or condensed overview rather than concept extraction.