Scalable knowledge base completion with superposition memories

TL;DR

HMem is a scalable neural architecture for knowledge base completion that models entities as superpositions of neighbor-relation pairs, enabling dynamic representation of unseen entities and achieving state-of-the-art results.

Contribution

Introduces Harmonic Memory Networks, a novel model that efficiently handles unseen entities and scales with knowledge graph growth without retraining.

Findings

Achieves state-of-the-art performance on benchmark datasets.

Supports dynamic generation of representations for unseen entities.

Demonstrates scalability and flexibility in evolving knowledge graphs.

Abstract

We present Harmonic Memory Networks (HMem), a neural architecture for knowledge base completion that models entities as weighted sums of pairwise bindings between an entity's neighbors and corresponding relations. Since entities are modeled as aggregated neighborhoods, representations of unseen entities can be generated on the fly. We demonstrate this with two new datasets: WNGen and FBGen. Experiments show that the model is SOTA on benchmarks, and flexible enough to evolve without retraining as the knowledge graph grows.