La Score Bleu Métrique, often abbreviated as BLEU, is a popular métrique d’évaluation used in the field of Traitement du langage naturel (NLP) to assess the quality of text produced by traduction automatique systems and other text generation models. Developed in the early 2000s, BLEU measures how closely the output of a model aligns with one or more reference texts, typically human-generated translations or summaries.
BLEU operates on the principle of comparing n-grams (contiguous sequences of n items) in the generated text with those in the reference texts. The basic formula for BLEU involves calculating the precision of n-grams, which is the ratio of the number of overlapping n-grams in the generated text to the total number of n-grams. BLEU also incorporates a brevity penalty to discourage short translations that might achieve high precision but fail to convey the full meaning of the source text.
La métrique renvoie un score allant de 0 à 1, où un score de 1 indique une correspondance parfaite avec les textes de référence. Cependant, le score BLEU présente certaines limites ; il se concentre principalement sur la précision et peut négliger des différences contextuelles ou sémantiques importantes. De plus, il peut être sensible à la longueur du texte de sortie, c’est pourquoi la pénalité de brièveté est incluse.
Despite its drawbacks, BLEU remains widely used because it provides a straightforward and quantitative way to evaluate and compare machine-generated text against human standards. It has been instrumental in benchmarking various NLP systems and continues to evolve with the advancement of les technologies d'IA.