المفاهيم الأساسية
Explainable Active Learning (XAL) enhances text classification by integrating rationales and explanations into the active learning process.
الملخص
The content introduces XAL, a novel Explainable Active Learning framework for text classification tasks. It addresses the limitations of traditional active learning methods by incorporating explanations to improve model performance. The framework consists of training and data selection processes, utilizing pre-trained encoders and decoders to generate explanations. Experimental results demonstrate the effectiveness of XAL in improving model performance across various text classification tasks.
Introduction
- Active learning efficiently acquires data for annotation.
- Traditional AL methods rely on model uncertainty or disagreement.
- XAL integrates rationales and explanations into AL for improved performance.
Methodology
- XAL framework includes training with encoder-decoder models.
- Data selection combines predictive uncertainty and explanation scores.
- Experiments show consistent improvement over baselines.
Results and Discussion
- XAL outperforms other AL methods in various text classification tasks.
- Ablation study confirms the effectiveness of each component in XAL.
- Human evaluation shows high consistency between generated explanations and model predictions.
الإحصائيات
XAL achieves consistent improvement over 9 strong baselines in experiments on six datasets.