AI-901 Text Analysis Techniques: Keyword Extraction, Entity Detection, Sentiment, and Summarization
Image: AI executing text analysis
Text analysis turns raw language into structured data a system can act on. The AI-901 exam expects you to recognize and differentiate four common techniques: keyword extraction, entity detection, sentiment analysis, and summarization. These capabilities are accessible via Azure AI Language and the language features in Microsoft Foundry.
The shortcut for this objective is to realize that each technique returns a different kind of output. Keyword extraction returns topics. Entity detection returns known types of values. Sentiment analysis returns a tone label. Summarization returns shorter text. Match the output the scenario needs to the correct technique.
What the Exam Expects You to Know
This article covers the Microsoft AI-901 objective: describe common text analysis techniques, including keyword extraction, entity detection, sentiment analysis, and summarization.
You should be able to recognize each technique from a given scenario, know exactly what each one returns, and cleanly separate them from one another.
Expect scenario-based questions: the exam might describe a batch of unstructured reviews, support tickets, or notes, and ask which technique fits the business goal.
The Four Techniques by Output Type
The fastest way to keep these distinct is by memorizing what they return.
| Technique | The Question it Answers | What it Returns |
|---|---|---|
| Keyword extraction | What is this text mainly about? | Important words and phrases. |
| Entity detection | Which known names, places, dates, or values appear? | Labeled entities by type. |
| Sentiment analysis | How does this text sound? | Positive, negative, neutral, or mixed. |
| Summarization | What is the short version? | A condensed version of the text. |
Keyword Extraction Finds the Main Topics
Keyword extraction (also called key phrase extraction) pulls the important words and phrases out of text. It does not only judge tone or write a summary, it surfaces what the text keeps mentioning.
Example input: "The delivery arrived late, but the packaging was excellent and support replied quickly."
Likely key phrases: delivery, packaging, support.
Common uses:
Finding repeated topics across thousands of reviews.
Tagging support tickets by the issues customers mention.
Surfacing common themes in survey comments.
Exam Scenario: A product team collects 20,000 survey comments and wants to know the main topics customers discuss (e.g., battery life, shipping delays). Keyword extraction fits because the goal is extracting important phrases from unstructured text.
Entity Detection Finds Known Types of Values
Entity detection (also called Named Entity Recognition, or NER) locates specific, known categories of information in text and labels them by type—such as person, location, organization, date, quantity, email, or phone number.
Example input: "Contoso will meet Maria Chen in Chicago on May 20."
Entities found: Contoso (Organization), Maria Chen (Person), Chicago (Location), May 20 (Date).
Exam Scenario: A travel company receives customer emails and wants to identify destination cities, travel dates, and traveler names. That is entity detection, because the system is finding specific entity types, not general topics.
Sentiment Analysis Estimates Tone
Sentiment analysis evaluates the tone of text and returns a specific label: positive, negative, neutral, or mixed. Mixed is used when a comment contains both praise and criticism.
Examples:
"The app is fast and easy to use." → Positive
"The order was late and support never replied." → Negative
"The package arrived yesterday." → Neutral
"The room was clean, but check-in took forever." → Mixed
Exam Scenario: A hotel chain wants a dashboard showing how many guest reviews are positive, negative, neutral, or mixed. That is sentiment analysis. (Keep in mind the service is reading language patterns, not reading minds, so heavy sarcasm may still require human review).
Summarization Produces a Shorter Version
Summarization condenses long text into a shorter version. For the AI-901, you should distinguish between two high-level styles:
Extractive summarization: Selects the most important existing sentences and stitches them together. The words come directly from the source.
Abstractive summarization: Writes a new, shorter version using fresh wording, much like a human paraphrasing a document. This style heavily overlaps with Generative AI.
Exam Scenario: A team wants a two-sentence summary of each long support thread so agents can catch up quickly. That is summarization. If they specifically want exact key sentences pulled from the thread, it leans extractive. If they want a fresh paraphrase, it leans abstractive.
Where These Run
These techniques are core language capabilities. Azure provides them as text analytics features via Azure AI Language, and you can add these same capabilities to custom applications built in Microsoft Foundry.
Exam Tip: Match the output to the technique.
Topics and themes = Keyword extraction.
Names, dates, places, and amounts = Entity detection.
Positive/negative/neutral/mixed = Sentiment analysis.
A shorter text block = Summarization.
If the question asks for fresh, paraphrased wording, that is abstractive summarization (which leans generative).
Common Trap: Do not confuse keyword extraction with summarization. Keyword extraction returns important phrases that already appear in the text. Summarization produces a logical shorter version of the document.
Also, do not select entity detection just because the text has nouns. "Slow checkout process" is a topic for keyword extraction. "Seattle", "Microsoft", and "April 15" are entities for entity detection.
Quick Knowledge Check
A retailer wants to know the most common topics mentioned across customer reviews. Which technique fits best?
A system needs to find people, organizations, and dates inside news articles. Which technique fits best?
A team needs to label support chats as positive, negative, neutral, or mixed. Which technique fits best?
A tool produces a short, fresh-worded version of a long report. Which technique fits best, and which specific style is it using?
Answers
Keyword extraction. The goal is to find important topics and phrases.
Entity detection. People, organizations, and dates are specific entity types.
Sentiment analysis. The system is estimating the overall tone.
Summarization (Abstractive). It is abstractive because it writes a new, shorter version using fresh wording.
Microsoft References For Further Study
Cross-reference this guide with the official Microsoft Learn documentation to familiarize yourself with their exact terminology:
