Building Fact-Checking Systems: Catching Repeating False Claims Before They Spread

Automated fact-checking systems are crucial for combating the spread of misinformation online, especially for claims that are repeatedly shared. These systems often use a pipeline that includes detecting a claim, retrieving evidence, and predicting its veracity. A key optimization is previously fact-checked claim retrieval (PFCR), which uses a retriever-reranker architecture to find existing verifications for similar claims. To improve this retrieval process, an ensemble approach combining lexical models like BM25 with semantic models like E5 and BGE can be used. This ensemble method, particularly with Reciprocal Rank Fusion, leverages the strengths of different models to achieve faster and more accurate results.