Question regarding the Triplet Loss in your paper

[Blackteaxx]

Dear Authors,

I recently read your excellent paper and was very impressed by your work on an embedding model for multi-grained retrieval. Thank you for this valuable contribution.

I am writing to seek clarification on the triplet loss function utilized in your study. While the concept is familiar, I would be grateful if you could elaborate on its specific formulation and practical implementation within the context of your multi-grained retrieval task.

In particular, I would appreciate it if you could shed some light on:

1. The precise mathematical formula of the triplet loss as applied in your experiments.
2. Any specific strategies employed for selecting anchors, positives, and negatives, especially considering the multi-grained nature of the retrieval.
3. Any other practical considerations or modifications made to the standard triplet loss that were important for achieving your reported results.

Any insights you could offer would be immensely helpful for my understanding and ongoing research.

Thank you for your time and consideration.