Abstract
The paper addresses the problem of short text request classification under conditions of significant class imbalance and high noise levels in real-world communication flows. The limited effectiveness of synthetic oversampling techniques when dealing with noisy labeling is demonstrated. A hybrid method is proposed, combining preliminary density-based data cleaning and multi-level model ensembling. The application of a density-based clustering algorithm enabled the exclusion of 16.5% of informational noise from the total sample volume. The final model features a two-level architecture and is optimized using Bayesian hyperparameter search. A Recall@3 (R@3) metric of 97.4% was achieved on a hold-out test set. The proposed method allows for the automation of the request distribution process, significantly reducing operator workload and decreasing dispatch time.