Aller au contenu
Atlas de l'IA

Classification de texte et analyse de sentiment

Attribuer une étiquette prédéfinie à un texte

Langage et connaissancesDébutant #05
entréeTexteTexte

Le texte intégral est présenté en anglais ; le titre et le résumé sont localisés.

CE QUE DÉSIGNE CETTE CAPACITÉ

Takes a piece of text and outputs its category — sentiment polarity, topic, intent or policy violation. The label set is fixed at training time, so it answers which class rather than what is said. A single text can carry several labels, and the model can also emit a probability distribution over the label set.

Comment c'est fait

Early systems paired bag-of-words with linear classifiers; fine-tuning a pre-trained encoder such as the BERT family later became the standard when labelled data is scarce. A classification head sits on top of the encoder and is trained with cross-entropy. With many labels and few examples, it can be recast as sentence-pair entailment or done zero-shot by prompting a large model.

Produits représentatifs

5

Organisations concernées

Usages typiques

  • Sentiment polarity of reviews and public opinion
  • Spam and policy-violation filtering
  • Routing tickets by topic and priority
  • Intent detection to drive dialogue

Comment on l'évalue

F1
Harmonic mean of precision and recall; more trustworthy than accuracy under imbalance
ROC-AUC
Ranking quality, threshold-independent
Confusion matrix
Per-class error direction, to locate systematic bias

Limites et points difficiles

  • Under heavy class imbalance, minority classes are ignored while accuracy still looks high
  • Wording drift outside the training domain degrades predictions
  • Sarcasm, negation and double negation are often flipped, since the literal words point the other way

Concepts sous-jacents