Knowledge Graph Completion via Receptance-Weighted Encoding and Tucker Decomposition
A knowledge-graph-completion model that replaces static entity/relation embeddings with a Triple Receptance Perception encoder, adapting RWKV-style time-mixing and channel-mixing blocks so representations become dynamic and context-dependent. The encoded head-entity and relation vectors are scored by a Tucker tensor-decomposition decoder with a learnable core tensor, yielding expressive yet parameter-efficient link prediction and triple classification that rivals Transformer encoders at lower cost.
2501.00397
This paper introduces Triple Receptance Perception (TRP), a knowledge-graph-completion encoder that adapts the RWKV attention-free sequential architecture (time-mixing and channel-mixing blocks) so t…