What is text analytics?

What text analytics is

Text analytics is the use of software to turn unstructured text, such as customer reviews, survey comments, support tickets and emails, into structured information you can count, compare and act on. Instead of reading ten thousand reviews, you see how many mention billing, how that changed this month and which stores are affected, with the original sentences one click away.

Why businesses use it

Customers say a lot in their own words that ratings and scores do not capture. Text analytics makes that usable at volume: which topics drive complaints, how sentiment differs by location, what changed after a new policy. It also saves the time of reading and coding everything by hand.

Core tasks

• Classification: assign each text to categories, such as Billing or Staff.
• Sentiment analysis: decide whether text is positive, negative or neutral, overall or per topic.
• Entity extraction: find names of products, places, people or dates.
• Clustering: group similar texts without predefined categories.
• Keyword and phrase extraction: find frequent or unusual terms.
• Summarisation: produce a short summary of many texts.

Main methods

• Rules and keywords: if a review contains "queue" or "wait", tag it Wait time. Transparent, but misses phrasing and misfires on negation.
• Traditional machine learning: a model learns from labelled examples to predict categories or sentiment. Needs training data per category.
• Transformer models: neural models pre-trained on large amounts of text, then fine-tuned for a task. Better at context and multiple languages.
• Large language models: general models that can classify or summarise from instructions and examples. Flexible, but need checks for consistency.

A typical pipeline

1. Collect text with metadata: date, location, source, rating.
2. Clean: remove duplicates, handle languages and very short texts.
3. Classify into a taxonomy of themes.
4. Score sentiment, ideally per theme.
5. Aggregate by location, period and topic.
6. Link every number back to the texts behind it.

The taxonomy matters more than the method

Whatever the technique, results are only as useful as the categories. A taxonomy built from your own customers' language, with clear definitions and levels, is what makes outputs actionable.

Limits to know

• Sarcasm, mixed feelings and comparisons are hard for any method.
• Very short texts often lack enough information to classify.
• Languages and code-mixed text, such as Hinglish, need models that handle them.
• Accuracy varies by data and task; test on a sample of your own text rather than relying on a vendor's figure.

Checking quality

Read a random sample of classified texts every so often and check whether labels are right. Look especially at low-confidence labels and the Other category. Keep the supporting quote for each label so anyone can check why a text was tagged.

Common business uses

• Review analysis across locations.
• Survey open-ended comments.
• Support ticket categorisation.
• Product feedback from app reviews.
• Complaint trend reporting.

Text analytics and text mining

The terms are often used interchangeably. Where a distinction is made, text mining refers to discovering patterns in text and text analytics to turning text into measurable, reportable insight.

Common mistakes

• Starting with a tool before defining questions and categories.
• Trusting accuracy figures not tested on your data.
• Overall sentiment only, without topics.
• No way to see the text behind a number.

Common questions

Checklist

• Questions defined.
• Taxonomy from your own text.
• Method chosen.
• Sentiment per topic.
• Quality checks.
• Quotes linked to numbers.

In short

Text analytics turns free text into categories, sentiment and trends using rules, machine learning or language models. A good taxonomy, quality checks and links back to the original text matter more than the method.

Doing this with Falcon

Falcon applies these ideas to one source in particular: the Google Maps reviews of your locations and the Google Play reviews of your app, classified into themes, sub-themes and specific issues, with the quote behind every label. Negative share by topic, location and month shows where a problem is growing, and small samples are flagged rather than ranked.