It's a nice work giving me a lot of inspiration! One minor confusion for me is the model structure in the codes.
The context retriever and the score retriever in the figure both contain a unique KG information retrieval module.
But the codes show that only the score retriever has a KG information retrieval module:
|
score = self.kg_retriever(h_id, r_id, t_id, head_embeds, rel_embeds, graph, score_text_embs, all_index) |
while the context retriever only contains a text information retrieval module:
|
text_embs = super().forward(kgl_ids) |
It's a nice work giving me a lot of inspiration! One minor confusion for me is the model structure in the codes.
The context retriever and the score retriever in the figure both contain a unique KG information retrieval module.
But the codes show that only the score retriever has a KG information retrieval module:
MKGL/retriever.py
Line 115 in 04e347f
while the context retriever only contains a text information retrieval module:
MKGL/retriever.py
Line 94 in 04e347f