BERTScore was proposed by Tianyi Zhang et al. in 2019 (ICLR 2020), and the related paper is BERTScore: Evaluating Text Generation with BERT. It computes token-level semantic similarity using contextual embeddings from models like BERT and aggregates them into precision, recall, and F1 scores. It is more robust to synonymy and word order variations, typically reported as percentages or scores between 0 and 1 (where higher is better), serving as an important semantic evaluation supplement for text generation, summarization, and translation.