search

zfj1998 / M3NSCT5

the code base for our paper "Diverse Title Generation for Stack Overflow Posts with Multiple Sampling Enhanced Transformer"
4 stars 1 forks source link

about the ranking stage #2

Open AnhLD2610 opened 10 months ago

AnhLD2610 commented 10 months ago

Hi author, as I understand in your model when using ranking you choose K candidates ranked, then from K candidates that you choose the title with highest rouge with title to compare with BM25 and PLBART when K>1 right? The result in table 4, sorry for my bad english

AnhLD2610 commented 10 months ago

I mean how could you choose the best tile from top_k titles generate. Hope to see your response soon!

shaoyuyoung commented 8 months ago

If I understand it correctly, they use Maximal Marginal Ranking strategy to rank candidate titles. In the results, the 1st rank title is the so-called best title.

Of course, you can mail Fengji (fengji.zhang@my.cityu.edu.hk) to make it sense as he may ignore GitHub issues if you are still confused about this problem.

AnhLD2610 commented 8 months ago

@shaoyuyoung the Maximal Marginal Ranking seem to enhance the diversity not the rouge score and it just choose top_k titles from top_m titles, i think that the best title is when k = 1 or the initial title but the result in the table show the best title among k titles i confuse how to get the best of k title or they just choose the title with the highest score?