View Code? Open in Web Editor
NEW
Code for KE-Blender, EMNLP 2021
ke-blender's Introduction
๐ Hi, Iโm @Nealcly, a senior researcher at Natural Language Processing Center, Tencent AI lab.
๐ Please email me ([email protected] ), if you would like to work with us.
ke-blender's People
Contributors
Watchers
ke-blender's Issues
Could not find any code to eval on test seen/test unseen
just concat dialog history with all retrieved_passages as input? It seems that it is too long.
With the help of the code and related papers provided by you, I reproduced it, but the implementation effect is not ideal and f1 scores are generally two percentage points lower. Do you have any experimental implementation details that need additional attention?
res = f_one (gold , decode_result )
Correct way should be:
res = f_one(decode_result, gold)
since reference is the second param
def f_one (hypothesis , references ):
Same problem exists in
b1 , b2 , b3 = bleu_metric (gold , decode_result )
Hello, I would like to know the specific location of the implementation code in the paper below