EnFVe: An Ensemble Fact Verification Pipeline

Abstract

Recent years have seen exponential growth in the volume of data generated. An undesirable consequence of this rapid expansion is the alarming rise of misinformation that is spreading throughout social networks and publishing platforms. Fact verification is the act of verifying the correctness of a particular statement. Manual fact-checkers find it increasingly difficult to keep up with the rapid proliferation of fake news in the information ecosystem. Fact verification pipelines aim at assisting the fact-checkers. These pipelines involve retrieving potentially relevant documents for a given claim, checking the reliability of the document media sources, evaluating the significance of each document, and verifying the correctness of given claims. However, there are a few limitations to these pipelines. The first limitation is that they treat fact verification as a stance detection or textual entailment task and places less emphasis on pure reasoning as the basis of fact verification. The second limitation is the robustness of the models. The fact verification models do not perform well on real-world data due to the text constructed using a myriad of complex ways. The reason for low accuracy is due to the evaluation and training of models on synthetic data that does not resemble the real-world data. In this paper, we present EnFVe, an end-to-end pipeline that integrates various components of fact verification. EnFVe pipeline improves upon the existing pipelines by integrating: (1) a Data Acquisition module that routinely scrapes data for constructing and updating the ground truth, (2) Preprocessing modules such as Coreference Resolution and Sentence Simplification that improves the robustness of the verification, (3) a Continuous Retrieval Engine module that captures the semantic context of the text, and (4) an Ensemble module that consists of various state-of-the-art fact verification models equipped with hyperparameter optimization. EnFVe achieves an accuracy of (dev/test) 79.26%/73.28%, which is better as compared to the state-of-the-art fact verification pipelines.