RUCAIBox/TextBox

Questions about the abulation study of the paper? (PTG)

Closed this issue · 1 comments

ssoyaavv commented

Hello. When posting an issue on GitHub, I have read the paper thoroughly and have developed some questions that arose during my reading. I would greatly appreciate it if you could provide me with your insights. Thank you in advance for your help

  1. In the ablation study, excluding [1]Prompt Pool and [2]Prompt Cluster, I couldn't find the relevant code to measure the performance. Could you please explain how the experiment was conducted?

  2. Does the experiment excluding [3]Multi-Key memory mean that MHA attention was not used? Can we understand that it had the same settings as SPOT?

StevenTang1998 commented

You can read the Section 5.4 Effectiveness of Core Designs for details.

[1] only training a sharing prompt for all source tasks.
[2] remove the step of grouping source prompts into different clusters and directly lookup source prompts based on queries.
[3] remove the learnable key vector associated with prompts and directly transfer the mean of the source prompts to the target task. It is similar with SPoT, but we average the source prompts.