BERTScore is based on a greedy matching approach that calculates sentence similarity by aligning word pairs with the maximum cosine distance. However, in cases where the machine-translated text and the reference translation have different lengths, or when the machine-translated text is not fully translated, the word alignment results may not be entirely accurate. To enhance the performance of BERTScore by obtaining more precise word alignment results, this paper adopts the word similarity matrix iterative alignment method. Experiments are conducted on the MQM dataset, and the results demonstrate that the improved BERTScore algorithm significantly advances the correlation with the MQM dataset. This proves that the word similarity matrix iterative alignment method obtains superior word alignment results compared to the greedy matching approach, and can effectively improve the performance of the BERTScore algorithm.