Explainable AI for NLP Dataset Debugging

Project Overview

This research introduces X-Deep, a Human-in-the-Loop (HitL) framework for systematically debugging NLP datasets using Explainable AI (XAI) techniques. The work was published in Nature Scientific Reports, demonstrating the practical value of data-centric AI approaches.

The X-Deep framework combines human expertise with XAI techniques for systematic dataset error analysis

Key Contributions

• Novel Framework: Designed a data-centric approach that iteratively improves datasets rather than just models
• XAI Integration: Combined LIME and SHAP to uncover patterns in model misclassifications
• Systematic Analysis: Identified spurious correlations, bias patterns, and other anomalies in Arabic emotion detection datasets
• Multi-Classifier Evaluation: Analyzed misclassifications across Naive Bayes, Logistic Regression, GRU, and MARBERT

Explainable AI in Action

The framework uses LIME and SHAP to understand why models make incorrect predictions:

Example: LIME and SHAP outputs revealing how repeated words influence emotion classification

Example: Detecting spurious correlations where specific words (e.g., “Olympics”) trigger incorrect labels

Impact

• Published in Nature Scientific Reports (Impact Factor: 4.6)
• Practical mitigation strategies for informed data collection and augmentation
• Reproducible framework applicable to other NLP tasks beyond emotion detection

Technical Details

Publication

“A Data Centric HitL Framework for Conducting a Systematic Error Analysis of NLP Datasets using Explainable AI”
Ahmed El-Sayed, Aly Nasr, Youssef Mohamed, Ahmed Alaaeldin, Mohab Ali, Omar Salah, Abdullatif Khalid & Shaimaa Lazem
Scientific Reports 15, 30406 (2025) | DOI: 10.1038/s41598-025-13452-y