Text Analysis Applications

What it is

Azure Language, one of the Foundry Tools, offers ready made text analysis: sentiment, key phrases, entities, personal data detection, language detection and summarisation. A text analysis application sends text to the service with a client library and reads the structured results.

Because these features are pre built, you write very little code and need no model training.

How it works

Setting up

The Python package is azure-ai-textanalytics. The client needs the endpoint of your Language resource and a credential.

import os
from azure.core.credentials import AzureKeyCredential
from azure.ai.textanalytics import TextAnalyticsClient

endpoint = os.environ["LANGUAGE_ENDPOINT"]
key = os.environ["LANGUAGE_KEY"]

client = TextAnalyticsClient(endpoint=endpoint, credential=AzureKeyCredential(key))

The endpoint and key come from environment variables, not from the source file. Microsoft recommends Microsoft Entra ID authentication with managed identities for production, so there is no key to leak.

Calling a feature

Every feature takes a list of documents, where each document is a string, and returns a list of results in the same order.

documents = [
    "The food and service were unacceptable. The concierge was nice, however.",
    "Check in was quick and the room was spotless.",
]

result = client.analyze_sentiment(documents, show_opinion_mining=True)

for doc in result:
    if doc.is_error:
        print("Error:", doc.error)
        continue
    print(doc.sentiment)
    print(doc.confidence_scores.positive, doc.confidence_scores.neutral, doc.confidence_scores.negative)

Reading results and errors

• The results line up with the input documents, so the first result belongs to the first document.
• Each result has an is_error property. If a single document is invalid, for example it is empty, its result is an error entry with is_error set to True, while the other documents are still scored normally. The list keeps its length.
• Always check is_error before reading the other properties of a result.
• Confidence scores run from 0 to 1 and the three sentiment scores add up to 1.

Choosing between a dedicated service and a language model

You can also do text analysis by prompting a language model, for example "give the sentiment of this review as one word". The two approaches differ.

Good practice

• Send documents in batches, which is more efficient than one call per text.
• Keep each document within the service's size limits, and split long text if needed.
• Do not log personal data. Use PII detection to redact it before storing text.
• Handle failures and rate limits with retries.

A worked scenario

A retailer wants a daily report of customer feedback. A Python script reads yesterday's reviews and sends them in a list to analyze_sentiment with opinion mining. It skips any result with is_error set, counts positive and negative reviews, and uses extract_key_phrases on the negative ones to see the common complaints. A batch of 25 reviews contains one empty string. The service scores the other 24 normally and returns an error entry for the empty one, so the script still has 25 results, and it logs the failed document. The key is read from an environment variable.