AI Glossary / AI and machine learning / Beginner
Text classification
Definition
Text classification assigns one or more categories to a piece of text. An AI system might classify a message by topic, urgency, sentiment or another set of labels defined for a particular task.
What it means
The text can be a sentence, an email, a review or a full document. The category system determines what the model is being asked to recognize. A model trained to identify product topics is not automatically a model for identifying dissatisfaction.
Some tasks allow only one label per item, such as routing an enquiry to a primary department. Others allow several labels, such as marking a document as both technical and customer-facing.
Classification produces a category or associated scores. It differs from extracting an exact phrase, summarizing a document or generating a reply, although one workflow may combine all these tasks.
For a business reader, the first design question is what action follows the label. Categories should be clear enough for people to apply consistently and useful enough to support a real decision.
How it works
The team defines labels and assembles representative examples. Annotators need guidance for ambiguous cases, mixed topics and messages that fit none of the categories.
The system converts text into a representation a model can process. Approaches range from counts of words and phrases to learned language-model representations. A classifier maps the representation to labels or scores.
Training and evaluation must keep related material appropriately separated. Replies in the same conversation, repeated templates or duplicated documents can make a random split unrealistically easy.
The deployed workflow uses thresholds and routing rules appropriate to its purpose. Uncertain or unfamiliar text may go to a review queue. Monitoring checks changing language, new product names, multilingual inputs and changes in how people use the service.
Three practical examples
- Customer service: classify enquiries by topic to suggest an appropriate queue while allowing staff to correct mistakes.
- Knowledge management: tag documents with approved subject categories to improve filtering and discovery.
- Feedback analysis: classify comments by product theme and expressed sentiment, keeping mixed or ambiguous cases visible.
Think of it this way
Think of sorting incoming letters into labeled trays. The trays are useful only if their labels correspond to meaningful destinations and the sorter can handle a letter concerning several issues.
The analogy has a limit. A model may react to recurring words without understanding the sender's intent as a person would. Context, negation and unusual phrasing can change the correct classification, so a neat label should not be mistaken for a complete reading.
Worked example
Evaluating an urgent-message label
Suppose an invented evaluation contains 200 messages, of which reviewers mark 40 urgent. The classifier flags 50 messages as urgent, and 30 of those are genuinely urgent.
Precision is 30 / 50 = 60%: six out of ten flagged messages are urgent. Recall is 30 / 40 = 75%: the model finds three quarters of all urgent messages.
There are 20 unnecessary escalations and ten missed urgent messages. Whether that trade-off is acceptable depends on review capacity and the consequences of delay.
Changing the threshold may alter both counts, so threshold selection belongs in development rather than repeated tuning on the final test set. These figures are illustrative, not expected performance for any product.
Illustrative case study
A fictional travel service classifies messages containing the word cancel as cancellation requests. Customers asking whether they can avoid cancellation are repeatedly routed incorrectly, while indirect requests such as I cannot make the trip are missed.
The team adds varied reviewed examples, clarifies its labels and compares a learned classifier with its keyword rule. It keeps a fallback queue for mixed requests and examines errors by language and message length.
The lesson is that classification depends on the relationship between wording, context and label definitions. A list of obvious trigger words may provide a useful baseline without capturing the full task.
Common misconceptions and mistakes
- Sentiment analysis describes the customer's whole situation. It classifies expressed text under a particular definition and may miss context.
- One label is always enough. Some messages genuinely belong to several categories or require a hierarchy of labels.
- High accuracy means routing is reliable for every topic. Rare but important categories need their own evaluation and error review.
Business use
Text classification can reduce manual sorting and make large collections easier to analyze. It is most useful when categories are stable, examples are available and mistakes can be handled sensibly.
Start with an assisted workflow and measure the work saved after corrections. Include annotation consistency, review queues and category maintenance in the cost estimate. A model that produces labels quickly can still create extra work if the labels are poorly matched to operational responsibilities.
People also ask
Does it require a large language model?
No. Simpler statistical models can be effective, especially with clear labels and suitable examples.
Can one text receive several labels?
Yes. That is a multilabel classification setup, distinct from choosing exactly one class.
Can it classify text in several languages?
It can, if the model and evaluation support those languages. Performance should be checked rather than assumed from one language's results.
Check your understanding
A classifier flags 50 messages and 30 are genuinely urgent. Is its precision 75%?
Answer: No. Precision is 30 divided by 50, or 60%. The 75% figure would be recall if there were 40 urgent messages in total.
Related terms
- Classification
- Sentiment analysis
- Precision
- Recall
- Text span
- multilabel classification
Learn next
Read precision and recall. Draft a short label guide with difficult examples, then ask two reviewers to apply it independently before training a classifier.
More in AI and machine learning
- Tensor sketch - Tensor sketch is a randomized technique for representing a large tensor-product feature map with a much smaller vector.
- Termination condition - A termination condition is a rule that determines when a process stops.
- Test set - A test set is a collection of examples reserved for assessing a model after its development choices have been made.
- Three-factor learning - Three-factor learning is a family of synaptic learning rules in which changes depend on presynaptic activity, postsynaptic activity and an additional...
- Threshold (for decision trees) - A threshold in a decision tree is a value used in a comparison at a branching point.
- Time travel (for features) - Time travel for features is the ability to retrieve historical versions of model-input values.
Sources
scikit-learn, Feature extraction, text feature extraction: https://scikit-learn.org/stable/modules/feature_extraction.html#text-feature-extraction. AIMA glossary term context: https://github.com/aimacode/aima-glossary/blob/master/glossary.md. Examples are original; these links are further primary references, not a claim that every page was read in this pass.
Previous: Tetrachoric correlation Next: Text corpus