Reproduce “End-to-End Differentiable Proving” on OpenResearch

About this paper

Tim Rocktäschel, Sebastian Riedel

This work introduces Neural Theorem Provers (NTPs), a neural network architecture for end-to-end differentiable proving in knowledge bases, which combines symbolic multi-hop reasoning with subsymbolic representation learning. The proposed NTPs, particularly when enhanced with an auxiliary link prediction loss (NTPλ), demonstrate improved performance on multi-hop reasoning tasks for knowledge base completion and are capable of inducing interpretable logical rules.

Key insights

• Differentiable unification enables symbolic reasoning systems to leverage continuous, vector-based representations of symbols, allowing generalization based on semantic similarity.
• Jointly training NTPs with a neural link prediction model (ComplEx) provides strong local gradient signals for symbol embeddings, significantly accelerating learning and boosting performance on complex reasoning tasks.
• The recursive, modular design of `unify`, `or`, and `and` operations allows for the construction of a differentiable computation graph that mirrors multi-hop logical inference paths.

Summary by alphaXiv.