Multihop Factual Claim Verification Using Natural Language Prompts

Abstract

Verifying a claim/statement using facts as evidence can be challenging, especially when the evidence consists of multiple sentences, making it difficult for NLP models to understand long-range dependencies. Most of the existing datasets provide claims that can be verified by single-hop reasoning i.e., relevant evidence to support or deny the claim can be found in a single evidence source. But the task becomes substantially challenging when it is required to mine evidence from multiple sources to correctly reach a verdict about the claim. Successful methods in single-hop verification task struggle to perform at a higher level when provided with a claim that requires multihop evidence in order to be verified. In light of the success of prompt learning in various NLP applications, this thesis introduces prompt learning for the multi-hop claim veri fication task. Through extensive experimentation, our proposed prompt-based method, which employs manually constructed prompts, has yielded promising results. By fine-tuning language models with prompts, we have achieved an accuracy of 83.9%, along with an enhanced cross-domain generalization performance. Additionally, we conducted experiments in few-shot and zero-shot settings, which demonstrated that prompt-based methods outperformed traditional supervised learning techniques that rely on the fine-tuning paradigm. These results underscore the effectiveness of prompt learning in the realm of claim verification