Extracting RTA Location and Time from Russian News: Combining Traditional Methods with Llama LLM Validation

A combined approach for extracting spatio-temporal attributes of Road Traffic Accidents from Russian-language news texts is presented. This approach integrates traditional methods, specifically rule-based techniques for temporal data and gazetteer-based methods supplemented with rules for spatial data, with validation procedures utilizing the Llama Large Language Model. The study focuses on processing texts related to Road Traffic Accidents occurring within St. Petersburg, Russia. The central hypothesis, positing that the incorporation of Large Language Model validation enhances extraction effectiveness compared to baseline methods lacking such validation, was investigated through experimental evaluation. A manually annotated dataset derived from regional news publications was used for this purpose. The results confirmed the hypothesis, indicating improvements in extraction performance, particularly in precision, for both temporal and spatial attributes when the Large Language Model validation step was included. Consequently, it was demonstrated that the proposed approach enables the reliable extraction of Road Traffic Accident location and time information directly from narrative text, even in the absence of explicit metadata. This capability facilitates the use of previously challenging open-source textual data for Road Traffic Accident monitoring and analysis, allowing for potential data integration via spatio-temporal proximity assessment.