An Evaluation of Strategies to Train More Efficient Backward-Chaining Reasoners

Knowledge bases traditionally require manual optimization to ensure reasonable performance when answering queries. We build on previous work on training a deep learning model to learn heuristics for answering queries by comparing different representations of the sentences contained in knowledge bases. We decompose the problem into issues of representation, training, and control and propose solutions for each subproblem. We evaluate different configurations on three synthetic knowledge bases. In particular we compare a novel representation approach based on learning to maximize similarity of logical atoms that unify and minimize similarity of atoms that do not unify, to two vectorization strategies taken from the automated theorem proving literature: a chain-based and a 3-term-walk strategy. We also evaluate the efficacy of pruning the search by ignoring rules with scores below a threshold.