Kala soocidda qoraalka

Text classification assigns categories to documents, messages, or other text sequences. Binary classification selects between two labels, such as spam and not spam. Multi-class classification assigns exactly one category from a larger set, while multi-label classification permits several categories for the same example. An article can carry both politics and economy tags without requiring those topics to exclude each other. Sentiment and intent can also be classification targets when the task defines their labels. Representations differ across systems: sparse word-count features and dense learned embeddings are two possibilities. A classifier converts those features into scores associated with the defined labels. Evaluation should match the intended decisions. Accuracy is the fraction of correct predictions, but it can conceal failures on a rare category. If almost every message is ordinary, always predicting ordinary can score highly while missing every relevant spam case. Recall for a designated positive class is true positives divided by all actual positives. Precision instead divides true positives by all predicted positives. In an illustrative evaluation with six true positives and two false positives, precision is six divided by eight, or 75 percent. Those quantities answer different questions. A model can have a useful recall yet flag many irrelevant messages. A further distinction concerns output scores: single-label softmax scores sum to one across mutually exclusive alternatives, whereas a separate sigmoid score for each label permits several labels to score highly together. Neither transformation alone establishes that scores are calibrated estimates of real-world frequency.