FactScore is an advanced real-time fact-checking system designed to verify political statements, specifically applied to candidates' remarks during the three debates of the 21st presidential election. The system operates on a foundational methodology that integrates real-time speech processing with automated fact verification.

At its core, FactScore utilizes a real-time Speech-to-Text (STT) component to transcribe spoken statements into textual scripts instantly. This STT model is engineered for low-latency processing to handle live debate environments, likely incorporating robust acoustic models and language models trained on political discourse and general conversational speech. The real-time generation of transcripts is crucial for immediate downstream fact-checking.

Following the real-time transcription, the textual statements are fed into a sophisticated fact-checking model. While the specific architecture of this model is not detailed, it would typically involve several stages:

1. Statement Extraction and Segmentation: Identifying distinct claims or factual assertions within the transcribed text. This might involve natural language processing (NLP) techniques such as sentence boundary detection and named entity recognition to identify subjects, predicates, and objects of claims.
2. Claim Normalization and Query Formulation: Transforming extracted claims into structured queries suitable for searching external knowledge bases or databases. This could involve semantic parsing or relation extraction to map natural language into a query language.
3. Evidence Retrieval: Querying a vast, curated knowledge base, a corpus of verified facts, or reputable external data sources (e.g., government statistics, academic studies, news archives from established media) to find supporting or refuting evidence for the extracted claims. This step might leverage information retrieval techniques, semantic search, or pre-indexed factual datasets.
4. Evidence Verification and Stance Detection: Analyzing the retrieved evidence to determine its veracity relative to the candidate's statement. This stage would involve natural language inference (NLI) or textual entailment models to assess whether the evidence *entails*, *contradicts*, or is *neutral* regarding the claim. Confidence scores for verification outcomes would likely be generated.
5. FactScore Assignment and Visualization: Based on the verification outcome and confidence, a "FactScore" is assigned to the statement. This score quantifies the truthfulness or factual accuracy of the utterance. The system then presents these verified statements and their corresponding FactScores on an interactive fact-check board. This board allows users to click on individual fact-checked cards to access detailed explanations of the verification process, including the specific claim, the evidence used for verification, and the source of that evidence.

The system's user interface is designed for accessibility, enabling users to explore the verification details and sources by interacting with the digital fact-check cards. The entire operation is overseen by "Team Underscore," which also provides further documentation on the model's operational principles and disclaimers, as well as contact information for media inquiries regarding content republication.