MoverScore

MoverScore scores a candidate against a reference over contextual embeddings, but instead of BERTScore's greedy one-to-one token matching it finds the minimal Earth Mover's Distance (Word Mover's Distance) needed to transform one text's embeddings into the other's (Zhao et al., 2019). This soft, many-to-one alignment lets partial and distributed matches count, which tends to track human judgment better than hard matching, though it is still a single opaque similarity number that says nothing about factuality.