Inquiry on Combining Hybrid O2M and Relation Attention Mask

[Anchor1566]

Question

Thank you for your remarkable work and elegant implementation !

While reproducing and experimenting with your approach, I encountered a question regarding the combination of the hybrid O2M (One-to-Many) mechanism and the proposed relation attention mask. Specifically, I observed that when implemented separately, both components contribute to performance improvements. However, when combined, the results show a noticeable decline. I have tried adjusting the weights of the O2M component, but the issue persists.

Could there be any subtle details or considerations that I might have overlooked when integrating these two components? Any insights or suggestions you could provide would be greatly appreciated.

Thank you in advance for your time and response. Also, I’d like to take this opportunity to wish you an early Happy Chinese New Year! May the coming year bring even more wonderful things to your life and research endeavors.

Additional

No response

[xiuqhou]

Hi @Anchor1566 Thank you for your kind words and thoughtful message. I truly appreciate your interest in our work.

Regarding your question about O2M and relation attention mask, I found only when O2O branch adopts Relation and O2M branch skips Relation, performance is improved. If both branches use Relation or not use Relation, the performance may be degraded. You can refer to the implementation in this repo to integrate the two modules.