Learning Inference Rules with Neural TP-Reasoner

Most standard deep learning models do not perform logical rule-based reasoning
like human and are hard to understand. We present a novel neural architecture,
Tensor Product Reasoner (TP-Reasoner), for learning inference rules represented
with a structured representation. In TP-Reasoner, we aim to integrate symbolic
inference and deep learning: we utilize the ability of Tensor Product Representation
in a neural model for learning and reasoning inference rules, which extracts
intermediate representations of logical rules from a knowledge base reasoning
task. TP-Reasoner achieves comparable results with baseline models. Analysis of
learned inference rules in TP-Reasoner shows the interpretability of logical composition
via a strong neuro-symbolic representation, a novel model expressivity, and
an explicit tensor product expressions.