Question about BiAttention about kobe HOT 6 CLOSED

Hi @AbnerCode ,

Thanks a lot for being interested in this work and checking out the code!

Question 1

No, we use the second encoder to encode retrieved knowledge. As the baseline model doesn't incorporate external knowledge, there is only one encoder – title encoder.

Question 2

For KOBE, condition_context_attn is not used in transformer.py, but used in tensor2tensor.py. Make sure you have the latest code. See:

from kobe.

Hi @qibinc ,

It is such an honor to hear from you.

So, as you say, you only use the title encoder as your baseline, which does not use the BiAttention. Is it like this?

from kobe.

Hi @AbnerCode ,

The above figure displays the architecture of KOBE.

For baseline, which is without the knowledge encoder, looks like this:

Best,
Qibin Chen

from kobe.

Hi @qibinc ,

Thanks for your reply!
In addition, I wanna know why you use the 2 layer transformer as the decoder. Generally, the layer is set as 6.

Best wishes!

from kobe.

Hi @AbnerCode ,

In addition, I wanna know why you use the 2 layer transformer as the decoder. Generally, the layer is set as 6.

Please check out #13. It mentions the same problem.

from kobe.

Hi @qibinc ,

Thanks!

Best Wishes

from kobe.