Extracting Latent Information from Datasets in CONSTRAINT 2021 Shared Task

Abstract

This paper introduces the result of Team Grenzlinie’s experiment in CONSTRAINT 2021 shared task. This task has two subtasks. Subtask1 is the COVID-19 Fake News Detection task in English, a binary classification task. This paper chooses RoBERTa as the pre-trained model, and tries to build a graph from news datasets. Finally, our system achieves an accuracy of 98.64% and an F1-score of 98.64% on the test dataset. Subtask2 is a Hostile Post Detection task in Hindi, a multi-labels task. In this task, XLM-RoBERTa is chosen as the pre-trained model. The adapted threshold is adopted to solve the data unbalanced problem, and then Bi-LSTM, LEAM, LaSO approaches are adopted to obtain more abundant semantic information. The final approach achieves the accuracy of 74.11% and weight F1-score of 81.77% on the test dataset.