Aligning Knowledge Graphs Provided by Humans and Generated by Neural Networks

In this paper, an approach that extracts knowledge graphs (KGs) from neural networks (NNs) and aligns the generated KGs with human-provided ones is proposed for network optimization or transparency enhancement, which is achieved by leveraging Vector Symbolic Architectures (VSAs). The approach identifies entities and relations of NN’s knowledge along with the training process, which makes it a plug-and-play solution. Experiments on synthetic data showed that the matching method works on middle and small-size KGs, and tests on MNIST demonstrated that the aligned NN-generated KG could be very close to the human-provided ones. Further tests on Text2KGBench showed that the method could produce KGs from embedding generated by backbone large language models (LLM) that aligned well with human-provided labels as well.