EXSEQREG: Explaining sequence-based NLP tasks with regions with a case study using morphological features for named...

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• State-of-the-art natural language engineering tasks rely heavily on machine learning.
• Existing systems lack robust methods for assessing prediction quality.
• Model interpretability is crucial for understanding and improving NLP systems.

Purpose of the Study:

• To introduce EXSEQREG, a universal method for explaining predictions in sequence-based NLP tasks.
• To provide a tool for assessing the quality and understanding the decision-making process of NLP models.
• To enable exploration of significant features contributing to model predictions.

Main Methods:

• Developed EXSEQREG, a method defining 'regions' linking predictions to important input features.
• Modeled prediction probability differences by systematically removing features.
• Applied the method to a neural network-based Named Entity Recognition (NER) tagger using Turkish and Finnish datasets.

Main Results:

• EXSEQREG generates importance values indicating feature impact on predictions.
• Qualitative analysis demonstrated reasonable explanations.
• Validation using mutual information scores confirmed the method's effectiveness.

Conclusions:

• EXSEQREG offers a reliable approach for explaining NLP model predictions across various tasks and models.
• The method facilitates assessing feature contributions to specific predictions.
• It aids in exploring globally significant features within a trained model.