| Scientific claim verification is a type of fact verification that aims to verify the authenticity of a proposed claim.Due to the rapid growth of the scientific literature,it is difficult for scientists to stay up-to-date with the latest findings.This challenge is especially acute during pandemics,such as the current Covid-19 pandemic,due to the risk of making decisions based on outdated or incomplete information.Therefore,the study of scientific claim verification systems has strong research significance and practical application value.Most existing work is dominated by pipeline models,which have the problem of error propagation.Therefore,this thesis proposes an approach called ARSJoint,which jointly learns modules for three tasks in a machine reading comprehension framework by including claim information.Meanwhile,to address the lack of information exchange between modules in the pipeline model,this thesis introduces a regularization term between the sentence attention score of abstract retrieval and the estimated output of rational selection to enhance the information exchange and constraints among tasks.Experimental results on the Sci Fact dataset show that the proposed joint learning approach has good performance and can effectively curb the error propagation problem.The performance of candidate abstract retrieval before claim verification determines the performance of subsequent claim verification models.Most existing candidate abstract retrieval methods are based on traditional machine learning methods,which have significant limitations.Based on this,this thesis trained an additional ARSJoint model for candidate abstract retrieval.However,neural network-based abstract retrieval methods often take too long,so the ARSJoint model for abstract retrieval was distilled into a smaller model to improve the retrieval speed.Experimental results demonstrate that the neural network-based abstract retrieval method achieves better performance than traditional machine learning methods.As the Sci Fact dataset is a small fact verification dataset,it has the problems of overfitting and poor generalization.To address this issue,this thesis proposes a method to pre-train the model on a similar large fact verification dataset called FEVER.The model is first pre-trained on the FEVER dataset to obtain generic knowledge,and then it is fine-tuned using the Sci Fact dataset for better handling of this specific dataset.Experimental results show that the pre-trained and then fine-tuned model achieves better performance on the SciFact dataset. |
