WnQinm / 2025final

data:

• entity, desc
• link (head, link, tail)

pre-process:

• entity passage(pos, k*neg) sim_entity(k)
• head-id link tail-id (以link的index作为Dataset索引)
• head(pos, k*neg, passage(pos, k*neg)) link tail(pos, k*neg, passage(pos, k*neg))

loss:

• entity reconstruction cross entropy(entity, passage)
• infoNCE(head+link, tail, neg_tail) (bs, embed_dim),(bs, embed_dim),(bs, k, embed_dim)
• infoNCE(tail-link, head, neg_head)

$$ -\log\frac{e^{s(q)/\tau}}{e^{s(q+)/\tau}+\sum e^{s(q-)/\tau}} $$