The dissemination of erroneous or misleading information, known as misinformation, has become a major societal concern due to its potential for catastrophic consequences. For instance, false rumors about medical advice could result in detrimental outcomes, while the spread of biased news has been linked to political polarization. Yet, given the sheer volume of information generated and shared, verifying its accuracy in a timely and accurate manner is not practically feasible through human labor alone. Consequently, it is imperative to develop efficient and effective methods for automatically mitigating misinformation.

One effective means of automatically addressing the issue of misinformation involves utilizing natural language processing (NLP) for fact-checking. This approach involves...[ Read more ]

The dissemination of erroneous or misleading information, known as misinformation, has become a major societal concern due to its potential for catastrophic consequences. For instance, false rumors about medical advice could result in detrimental outcomes, while the spread of biased news has been linked to political polarization. Yet, given the sheer volume of information generated and shared, verifying its accuracy in a timely and accurate manner is not practically feasible through human labor alone. Consequently, it is imperative to develop efficient and effective methods for automatically mitigating misinformation.

One effective means of automatically addressing the issue of misinformation involves utilizing natural language processing (NLP) for fact-checking. This approach involves retrieving relevant evidence from a knowledge base and then verifying the accuracy of the information against this evidence. It is considered reliable because it is justified with facts and evidence. Many explorations and advancements have been made in fact-checking methods, however, several critical limitations remain: (1) A substantial amount of expensive data is required to properly train a fact-checking model. (2) For rapidly evolving misinformation, there is often no reliable evidence available to do proper fact-checking; this is especially the case for misinformation related to political and economic events. (3) Fact-checking fails to properly capture misinformation involving framing bias. This is because framing bias is not always incorrect but instead involves subjective or nuanced wording that deliberately conveys misleading impressions about what really happened. In this thesis, we aim to improve and go beyond existing fact-checking solutions to more exhaustively cover misinformation.

Firstly, we introduce novel zero-shot fact-checking techniques that leverage LLMs to verify information in a data-efficient manner. When LLMs are trained on a vast corpus of text, they can extract rich world knowledge that makes them an effective tool for fact-checking purposes. Our research demonstrates that the likelihood (and perplexity) of LLMs can serve as a reliable proxy metric for the factual accuracy of information. We propose techniques to leverage such ability of LLMs to efficiently fact-check misinformation.

Secondly, we show that misinformation without evidence can still be effectively tackled by relying on the stylistic features. We propose to learn a generalizable and discriminative features, such as stylistic and linguistic patterns, associated with the misinformation itself, and leverage them to detect misinformation. To achieve this, we jointly trained LLM on a variety of different domains of misinformation and empirically verified the effectiveness of our model in detecting misinformation in previously unseen events.

Lastly, we propose to de-bias the framing bias in misinformation by fine-tuning the LLM to generate a neutralized alternative. Since articles from conflicting news outlets still share the “same set of underlying facts” [28], we propose a new task of generating a neutral summary from a group of polarized and biased news articles. For this new task, we establish a new benchmark by collecting the dataset, designing set of evaluation metrics, and proposing strong baseline models.