Prioritization of tasks created from textual description using language models
Marek Bazan, Tomasz Gniazdowski, Maciej E. Marchwiany · Procedia Computer Science · 2024
In this article, we introduce a novel method to improve the quality of priority prediction on small datasets. Our approach combines data augmentation with automatic annotation using an open dataset (like Enron). We investigated various annotation methodologies, including Best-Worst scaling (BWS), Pair-wise annotation, and annotation using pairs created by comparing models. The new method increases precision by more than 10% with respect to the best baseline and reduces Mean Squared Error by about 40%. Additionally, we worked on texts in the Polish language (both native, real data or translated to Polish), enabling the benchmarking of two language models: XLM-RoBERTa and HerBERT for Polish. Through numerical experiments, we demonstrate the effectiveness of our method, showcasing its potential for real-world applications. Additionally, our results serve as a compelling example of transfer learning for contrastive learning.