bertscore

BERTScore uses BERT embeddings to measure semantic similarity between generated and reference text. **How it works**: Encode candidate and reference sentences with BERT, compute pairwise cosine similarity between token embeddings, greedily match tokens, aggregate into precision, recall, F1. **Advantages over BLEU/ROUGE**: Captures semantic similarity not just n-gram overlap. Same meaning, different words gets credit. **Calculation**: For each candidate token, find most similar reference token (and vice versa). Precision = avg best match for candidate tokens. Recall = avg best match for reference tokens. **IDF weighting**: Optionally weight tokens by inverse document frequency (rare words matter more). **Layer selection**: Different BERT layers capture different features. Later layers often better for semantics. **Use cases**: Machine translation, summarization, text generation evaluation. **Limitations**: Still a proxy (not human judgment), can be fooled by adversarial examples, computationally heavier than BLEU. **Variants**: RoBERTa-based, multilingual versions available. **Best practice**: Use alongside other metrics, validate correlation with human judgment for your task.