BERTScore é uma métrica empregada na área de processamento de linguagem natural (NLP) that leverages embeddings contextuais from the BERT (Bidirectional Encoder Representations from Transformers) model to evaluate the similarity between text segments. Unlike traditional metrics such as BLEU or ROUGE, which primarily rely on exact word matches, BERTScore considers the semantic meaning of words based on their context within the sentence.
The core idea behind BERTScore is to align the tokens of two texts—typically a reference text and a generated text—using the cosine similarity of their BERT embeddings. This means that BERTScore captures nuances in language and can better assess quality in tasks like tradução automática, text summarization, and paraphrase generation.
To compute BERTScore, each token in the generated text is compared to all tokens in the reference text by calculating the cosine similarity of their corresponding embeddings produced by the BERT model. The maximum similarity score for each token is taken, and the escore geral is derived by averaging these maximum scores. BERTScore can be calculated for precision, recall, or F1-score, depending on the specific needs of the evaluation.
Essa métrica ganhou popularidade devido a its ability to provide a more nuanced understanding of text quality, making it particularly valuable in scenarios where semantic meaning is crucial. By utilizing the advanced capabilities of BERT, BERTScore enhances the evaluation process in NLP applications, aligning it closer to human judgment.