Semantic Text Checking

Abstract

In this thesis, we developed a general framework to provide text editors with semantic checking service that helps identify in-text factual errors. The framework consists of transportable semantic checker within a Docker on server, a RESTful API for the communication between client and server, and a sample text-editor with method to call the semantic checking service. In order to find appropriate semantic checker, two fact extraction and verification models: FEVER and Athene, are examined in three steps. To begin with, quantitative tests are applied to show general performance of the two models. Subsequently, qualitative experiments are performed to identify both models’ capability with respect to different types of text. Finally runtime analysis shows efficiency and practicality of the models. The results of the three-step examination provide insights to select appropriate semantic checker model and indicate possible improvements in the future.