Enhanced Knowledge Graphs Using Typed Entailment Graphs

Abstract

Constructing knowledge graphs from open-domain corpora is a crucial stage in question answering. Most previous works are based on open information extraction methods, which extract relations by parsing sentences into triples <e1, r, e2>. These methods lack inference ability and are limited by corpus. When the query is different from the relations in the text-based knowledge graph, it is hard to return correct answers. In this paper, we propose a method to enhance knowledge graphs by using typed entailment graphs to add missing links. We construct the enhanced knowledge graph in both dynamical and offline ways. The experiment shows that our method outperforms the pre-trained language models in zero-shot cloze-style question answering. Furthermore, we find entailment graphs can significantly improve the recall and F-score of knowledge graphs.

Tasks

Reproductions