Closed cheungdaven closed 5 years ago
Another quick question about the head and tail mode, are there any differences for Complex and DistMult in these two different modes?
Hi, this is the paper for L3 regularization, which proves to be effective for DistMult and ComplEx: Lacroix, Timothée, Nicolas Usunier, and Guillaume Obozinski. "Canonical tensor decomposition for knowledge base completion." arXiv preprint arXiv:1806.07297 (2018).
The difference in head and tail mode is that in the head-batch mode, we predict the head entity given the tail entity and relation, while in the tail-batch mode, we predict the tail entity given the tail entity and relation.
Using these two modes helps save memory and computation.
I am curious about the L3 regularization you use for complEx and DistMult? Can you give any references for it?