Directed Graph Networks for Logical Entailment

EasyChair Preprint 2185, version 2

Abstract

We introduce a novel neural model for detecting propositional entailment, a benchmark task for learning on logical structures, based upon learned graph convolutions on directed syntax graphs. The model removes some inflexible inductive bias found in previous work on this domain, while producing competitive results on the benchmark datasets. Model performance on larger problems surpasses all previous work. We also introduce a similar first-order learning problem and show good performance of the same model on this task. Such models have many applications for learned guidance of first-order theorem provers.

Keyphrases: Graph Neural Network, Logical Entailment, automated reasoning, directed acyclic graph

@booklet{EasyChair:2185,
  author    = {Michael Rawson and Giles Reger},
  title     = {Directed Graph Networks for Logical Entailment},
  howpublished = {EasyChair Preprint 2185},
  year      = {EasyChair, 2020}}