Introduction

IT support is a critical sector for operational efficiency, but remains challenging due to tedious textual inputs, subjective writing styles, and confusing classification systems. A new study identifies two families of approaches for prioritizing IT requests: pipeline-based models using embeddings that combine dimensionality reduction, clustering, and classical classifiers, and a fine-tuned multilingual model that processes both textual and numerical characteristics.

Comparison Results

Embedding-based methods exhibit limitations in generalizing across 30 configurations with clustering that fails to discover meaningful structures and supervised models are highly sensitive to the quality of embeddings. In contrast, the proposed model achieves superior performance with an average F1 score of 78.5% and weighted kappa values of almost 0.80, indicating strong alignment with real signals.

Technological Implications

These results highlight the limitations of generic embeddings for ITSM data and demonstrate the effectiveness of AI-dominated architectures for operational prioritization of IT requests. The proposed model represents a new generation of LLMs that use multi-lingual training to handle both textual and numerical characteristics, particularly effective in processing subjective and tedious inputs.

Conclusion

This study emphasizes the importance of fine-tuning AI-driven architectures to address the specific challenges of ITSM data, highlighting an unexplored path towards more effective and flexible prioritization of IT requests.