| dc.description.abstract | In an era characterized by the rapid expansion of online information and the widespread dissemination of misinformation, automated fact-checking has emerged as an essential area of research. As digital platforms continue to proliferate, the necessity for accurate and efficient fact-checking mechanisms is attracting increasing interest. Automated fact-checking systems address two main tasks: claim detection and claim validation. Claim detection involves identifying sentences or text snippets containing assertions or claims potentially subject to fact-checking. Claim validation, a multifaceted endeavor, encompasses evidence retrieval and claim verification. During evidence retrieval, relevant information or evidence that may support or refute a given claim is obtained. Claim verification, on the other hand, entails assessing the veracity of a claim by comparing it against available evidence. Typically framed as a natural language inference (NLI) problem, claim verification requires the model to determine whether a claim is supported, refuted, or there is not enough information to reach a verdict. In this thesis, we explore challenges inherent in claim verification, with a focus on few-shot scenarios where limited labeled data and computational resources pose significant constraints. We introduce three innovative methods tailored to tackle these challenges: Semantic Embedding Element-wise Difference (SEED), Micro Analysis of Pairwise Language Evolution (MAPLE), and Active learning with Pattern Exploiting Training models (Active PETs). SEED, a novel vector-based approach, leverages semantic differences in claim-evidence pairs to perform claim verification in few-shot scenarios. By creating class representative vectors, SEED enables efficient claim verification even with limited training data. Comparative evaluations against previous state-of-the-art methods demonstrate SEED's consistent improvements in few-shot settings. MAPLE is another pioneering approach to few-shot claim verification, harnessing a small seq2seq model and a novel semantic measure to explore the alignment between claims and evidence. Utilizing micro analysis of pairwise language evolution, MAPLE achieves significant performance improvements over state-of-the-art baselines across multiple automated fact-checking datasets. Active PETs presents a novel ensemble-based active learning approach for data annotation prioritization in few-shot claim verification. By utilizing an ensemble of Pattern Exploiting Training (PET) models based on various pre-trained language models, Active PETs effectively selects unlabelled data for annotation, consistently outperforming baseline active learning methods. Its integrated oversampling strategy further enhances performance, demonstrating the potential of active learning techniques in optimizing claim verification workflows. Together, these methods represent significant advancements in claim verification research, offering scalable and practical solutions. Through extensive experimentation and comparative analysis, this thesis evaluates the effectiveness of each method on various dataset configurations and provides valuable insights into their strengths and weaknesses. Furthermore, by identifying potential extensions and areas for refinement, the thesis lays the groundwork for future research endeavors in this critical field of artificial intelligence. | en_US |
