bert_score

Why we included this project

BERTScore evaluates text generation by comparing candidate and reference texts with pre-trained contextual embeddings and cosine similarity, so it captures meaning that word-overlap metrics like ROUGE or BLEU miss. It returns precision, recall, and F1 rather than a single opaque number, and its scores track human judgment closely, which is what you want when benchmarking a summarizer or translator and reporting results reviewers will take seriously. You can score batches from the command line or use the BERTScorer object inside a Python pipeline, and roughly 130 pretrained models are supported, with guidance on which ones align with human ratings best. The library also handles awkward edge cases like empty candidate strings and multiple reference sentences.