Using Text-Based Representations of Knowledge Graphs to Improve the Consistency of Generated Text

Abstract

Knowledge-enhanced language models in Natural Language Processing (NLP) have demonstrated in recent years that they are effective in many applications. This work attempts to show that knowledge-enhanced language models can be used to produce outputs that are not only more consistent with the information that is already known, but are also more plausible given other known information. In order to do so, this work introduces a modified Bidirectional and AutoRegressive Transformers (BART) model\cite{Lewis2020BARTDS}, which is modified to use a Text-based Representation of a Knowledge Graph (BART-TRKG). Experimental results demonstrate that this approach, when evaluated on a custom dataset, is able to improve upon the performance of a plain BART model in consistency and plausibility, as well as on other metrics according to a variety of human evaluators.