Sentiment Analysis and Key Phrase Extraction

Sentiment analysis and key phrase extraction are two of the pre-built features offered by Azure AI Language (formerly Text Analytics). They are part of the Language service, which provides REST APIs and SDKs for text analysis. Sentiment analysis determines the emotional tone of a piece of text—whether it is positive, negative, mixed, or neutral—at both the document and sentence levels. Key phrase extraction identifies the main talking points or topics within the text, returning a list of phrases that capture the essence of the content.

These features are widely used in business scenarios such as social media monitoring, customer feedback analysis, and survey processing. They allow organizations to automatically process large volumes of unstructured text and gain actionable insights without manual effort.

Azure's sentiment analysis uses a deep learning model based on a transformer architecture, similar to BERT (Bidirectional Encoder Representations from Transformers). The model is trained on a large corpus of text with human-labeled sentiment scores at the sentence and document levels. Here is the step-by-step mechanism:

Key phrase extraction uses a different model that focuses on identifying noun phrases and other significant terms. The process involves:

To use sentiment analysis or key phrase extraction, you call the Azure AI Language REST API. The endpoint is typically:

https://<your-resource-name>.cognitiveservices.azure.com/text/analytics/v3.1/sentiment

For key phrase extraction:

https://<your-resource-name>.cognitiveservices.azure.com/text/analytics/v3.1/keyPhrases

You need a valid API key and a JSON payload containing the documents to analyze. Each document is an object with an "id", "text", and optionally "language". Here is an example request for sentiment analysis:

{
  "documents": [
    {
      "id": "1",
      "language": "en",
      "text": "The product is excellent. I am very satisfied."
    },
    {
      "id": "2",
      "language": "en",
      "text": "The service was terrible and slow."
    }
  ]
}

The response includes sentiment labels and scores for each document and each sentence. For key phrase extraction, the response contains a list of key phrases for each document.

Sentiment analysis and key phrase extraction are often used together. For example, you might first extract key phrases to identify topics, then analyze sentiment to understand how people feel about those topics. They can also be combined with other Language features like entity recognition and language detection. For instance, you could detect the language first, then apply sentiment analysis in that language.

Azure AI Language is billed per 1,000 text records. Each document counts as one record regardless of length (up to the character limit). The free tier (F0) offers 5,000 records per month. The standard tier (S) has no fixed limit but is pay-as-you-go. Throughput is limited by your resource's pricing tier; the S tier can handle up to 20 requests per second by default, but you can request a higher limit.

Common errors include: - InvalidLanguageCode: The specified language is not supported. - InvalidDocument: The document exceeds the size limit or is empty. - Forbidden: The API key is invalid or expired. - TooManyRequests: You have exceeded the rate limit.

Always check the HTTP status code and the error message in the response.

A large retail company monitors Twitter mentions to gauge public sentiment about its brand. They use Azure AI Language's sentiment analysis to automatically classify tweets as positive, negative, or neutral. Key phrase extraction identifies trending topics like "shipping delays" or "product quality." The system processes over 100,000 tweets daily. The API is called from an Azure Function triggered by a Twitter stream. Results are stored in Azure Cosmos DB and visualized in Power BI dashboards. When negative sentiment spikes, an alert triggers a manual review. A common misconfiguration is not specifying the language code, causing misclassification of tweets in mixed languages. Also, the free tier's 5,000 records per month limit is quickly exceeded; they needed the S tier. Performance considerations include API rate limits—at 20 requests per second, they batch tweets into groups of 10 to maximize throughput.

A telecom company analyzes post-call survey responses. Customers leave voice feedback transcribed by Azure Speech to Text. The transcribed text is then analyzed using sentiment analysis and key phrase extraction. This identifies common complaints (e.g., "billing issues") and overall satisfaction trends. The system processes 500,000 transcripts per month. They use the Language service's asynchronous operation for large batches. A common pitfall is not handling mixed sentiment correctly—a survey that says "The service was great, but the price is too high" is labeled "mixed" by the API, but their initial logic treated it as neutral, missing the negative aspect. They learned to use sentence-level scores to drill down into specifics.

A social media platform uses sentiment analysis to flag potentially harmful content. They combine sentiment analysis with other AI services like Content Moderator. For example, a post with very negative sentiment and aggressive key phrases like "hate" or "attack" is escalated for review. They process millions of posts daily, requiring high throughput. They implemented a caching layer to avoid redundant API calls for identical text. A scaling challenge was the API's 10 documents per request limit; they built a queue system to batch requests efficiently. Misconfiguration of the region (e.g., using a resource in a different geography) increased latency, so they deployed resources in multiple regions.