Symbolic Semantic Memory in Transformer Language Models

This paper demonstrates how transformer language models can be improved by giving them access to relevant structured data extracted from a knowledge base. The methods for doing so include identifying entities in a text corpus, sorting the entities using a novel attention-based approach, linking entities to a knowledge base, then extracting and filtering the knowledge to create a knowledge-augmented dataset. We evaluate these methods with the WikiText-103 corpus using standard language modeling objectives. These results show that even simple additional knowledge augmentation leads to a reduction in validation perplexity by 81.04%. These methods also significantly outperform common ways of improving language models such as increasing the model size or adding more data.