I modified a file to use a BERT model for the conv_model on the redial dataset, and I obtained a dist@k score > 1 (this result is more higher than the scores reported in the papers below, which raises some doubts). Is this normal? It seems that the range values for dist@k, similar to bleu@k, should both fall within the [0,1] range. When I compared it with other papers, the dist@k for the redial dataset mostly seemed to fall within the [0,1] range. Since I don't have a clear understanding of the dist@k code, could someone help me with this?
My results:
Paper: (https://arxiv.org/pdf/2007.04032.pdf), (https://arxiv.org/pdf/2206.09363.pdf)
Bert Model:
Dist@k score: I add a hyp_token = hyp.split()
I modified a file to use a BERT model for the conv_model on the redial dataset, and I obtained a dist@k score > 1 (this result is more higher than the scores reported in the papers below, which raises some doubts). Is this normal? It seems that the range values for dist@k, similar to bleu@k, should both fall within the [0,1] range. When I compared it with other papers, the dist@k for the redial dataset mostly seemed to fall within the [0,1] range. Since I don't have a clear understanding of the dist@k code, could someone help me with this?
My results:
Paper: (https://arxiv.org/pdf/2007.04032.pdf), (https://arxiv.org/pdf/2206.09363.pdf)
Bert Model:
Dist@k score: I add a hyp_token = hyp.split()
