Need more traditional evaluation metrics

[3210448723]

Hope to add the evaluation of traditional evaluation metrics, such as' rouge ',' bleu ',' bert ',' meteor ',' sbert ', etc.

[Xinle-Deng]

Thank you for your suggestions. I will add them tomorrow.

[Xinle-Deng]

Hello, i have added token-level F1 score, ROUGE, BLEU, and BertScore.

It is worth noting that the computed scores may appear lower than those reported in prior work. This discrepancy is expected and can be attributed to differences in prompting.

Specifically, the prompt used for answering questions based on memory does not enforce concise responses, whereas the ground-truth answers are typically short phrases. As a result, the generated outputs tend to be longer and less aligned at the surface level with the references, leading to lower automatic metric scores.

To obtain more comparable results, you may customize the question-answering prompt to explicitly encourage concise responses. The corresponding prompt template can then be passed into the evaluation script via the --prompt-template argument.