Adapter fusion for check-worthiness detection - combining a task adapter with a NER adapter

Detecting check-worthy statements aims to facilitate manual fact-checking efforts by detecting claims that fact-checkers should prioritize first. It can also be considered as the first step of a fact-checking system. In this paper, we present an adapter fusion model that combines a task adapter with a NER adapter achieving state-of-the-art results on two challenging check-worthiness benchmarks. Adapters are a resource-efficient alternative to fully fine-tuning transformer models. Our best performing model obtains an 𝐹1 score of 0.92 on the CheckThat! Lab 2023 dataset. Additionally, we interpret the fusion attentions, demonstrating the effectiveness of our approach. The quantitative analysis of the fusion attentions shows that named entities contribute significantly to the prediction of the adapter fusion model.