ViNumFCR: A Vietnamese News Fact-Checking System for Numerical Reasoning

This article introduces the ViNumFCR project, a specialized fact-checking system for numerical reasoning scenarios in Vietnamese news. The system integrates Playwright data extraction, large language model fine-tuning, and complex reasoning chain evaluation technologies to provide a complete technical solution for detecting fake news involving percentages, absolute values, and time-series data. The project is open-sourced by developer cdmanh1108 and focuses on verifying numerical claims, distinguishing itself from general-purpose fact-checking systems.

In the era of information explosion, fake news spreads faster than it can be verified—especially news involving numerical claims in finance, statistics, etc., which requires professional analytical capabilities and verification processes. ViNumFCR was developed to address this challenge, focusing on numerical reasoning-based fact-checking for Vietnamese news, particularly handling complex scenarios such as percentage changes, absolute value comparisons, and time-series trend analysis.

ViNumFCR adopts a multi-stage pipeline architecture, including three core modules:

1. Data Extraction Layer: Uses the Python Playwright framework to automatically crawl web data (including dynamic tables/charts) to ensure traceable data sources;
2. Data Processing and Feature Engineering: Cleans structured data, unifies numerical units/formats, aligns text claims with extracted data, and identifies numerical relationships;
3. Model Reasoning and Verification: Based on the ViFactCheck benchmark dataset, uses open-source models like Google Gemma for fine-tuning, and evaluates complex reasoning chains (e.g., multi-step calculations to verify numerical claims).

ViNumFCR optimizes for three types of numerical reasoning scenarios:

• Percentage and Ratio Calculation: Accurately identify calculation benchmarks, handle year-on-year/sequential growth, and understand non-linear characteristics (e.g., a 50% increase followed by a 50% decrease does not return to the original point);
• Absolute Value Comparison: Precisely match values and understand statistical standards (e.g., gross income vs. net income);
• Time-Series Analysis: Identify seasonal patterns, outliers, and long-term trends, and verify whether trend descriptions match the data.

ViNumFCR is built on the ViFactCheck benchmark dataset, which covers numerical claims across multiple domains in Vietnamese news. In addition to traditional accuracy, evaluation metrics include numerical reasoning-specific indicators such as calculation correctness, unit consistency, and reasoning chain completeness, reflecting the system's real performance from multiple dimensions.

ViNumFCR can be applied in:

• Newsrooms: Quickly verify numerical claims in manuscripts;
• Financial institutions: Verify market rumors and financial data;
• Government departments: Check the accuracy of statistical data. Especially in the financial sector, it can reduce market impact and reputation risks caused by data errors, and improve the accuracy and timeliness of news releases.

The technical architecture of ViNumFCR is language-agnostic; replacing language-specific models and datasets can extend it to other languages. Combining Retrieval-Augmented Generation (RAG) technology, it can access real-time data sources to achieve instant verification of new news. As the capabilities of large language models improve, reasoning accuracy is expected to further increase.

ViNumFCR represents the trend of fact-checking towards specialization and refinement, demonstrating the idea of combining large language models with traditional data processing technologies to build practical automated systems. It provides valuable technical references for developers and researchers concerned with media credibility and information authenticity.