Multi-agent organization of an incoming mail processing system using a metagraph model

The article proposes a hybrid intelligent information system for letter classification based on a metagraphic knowledge model and a holonic multi-agent architecture. The system combines statistical, lexico-semantic and deep semantic methods in a single structure, which allows for both high accuracy and transparency of solutions. A mechanism of self-organization based on user feedback is proposed. Experiments on a real corpus of university letters have shown that the hybrid approach can improve classification accuracy by 4–5 % compared to single-method systems. The results demonstrate the promise of using metagraphs to build adaptive and explicable information systems.