Reproduce “Traversing Knowledge Graphs in Vector Space” on OpenResearch

About this paper

Kelvin Guu, John Miller, Percy Liang

Researchers at Stanford University developed a compositional training methodology for knowledge graph embedding models, enabling them to accurately answer multi-hop path queries and simultaneously improve single-edge knowledge base completion. This approach reduced path query errors by up to 76.2% and enhanced KBC performance by up to 43.3% across different datasets.

Key insights

• Many existing KBC models can be generalized to perform compositional reasoning by reinterpreting their scoring functions into composable traversal and membership operations.
• Directly optimizing for multi-hop path accuracy during training effectively mitigates cascading errors, producing more robust intermediate representations.
• Compositional training provides a powerful form of structural regularization, surprisingly improving performance on the original single-edge Knowledge Base Completion task in addition to path queries.
• The models implicitly learn and exploit logical rules (Horn clauses) by aligning spatial relationships derived from paths with those of single relations.

Summary by alphaXiv.