Heuristic Backward Chaining Based on Predicate Tensorization

Inference methods for first-order logic or its fragments are inherently slow. Neural networks make it possible to rapidly approximate the truth values of ground atoms but the results are not explainable and not necessarily accurate. A hybrid neural-symbolic inference method is proposed in this paper. It is a framework for incorporating predicate tensorization techniques into backward chaining and its extensions. Heuristic functions for best-first search strategies for backward chaining are defined via tensor representations of predicates. These strategies speed up inference by reducing backtracking.