Talk · Ural Symposium on Biomedical Engineering, Radioelectronics and Information Technology · May 14, 2026
Cross-Model Analysis of Token-Level Explanations in Tatar Text Classification
Fail Mubarakovich Gafarov, Viliuza R. Gafarova, Madehur M. Ayupov
Abstract
The aim of this work is to evaluate the performance, interpretability, and consistency of explanations generated by different Explainable AI (XAI) methods for different machine learning models trained on a Tatar language text classification task. Explanations were generated by using Integrated Gradients and LIME methods, and aggregated to obtain global importance scores. Inter-model similarity of explanations was assessed by using Spearman rank correlation and Jaccard overlap, while cross-model consensus was identified through bootstrap confidence intervals and Kendall’s coefficient of concordance. Although classifiers demonstrated comparable predictive performance, their global explanatory patterns differed substantially. But a subset of architectureindependent tokens showed statistically stable and directionally consistent influence across models. The proposed framework enables systematic evaluation of explanation stability and is applicable to low-resource NLP languages requiring transparent and robust model analysis. The work contributes both to the field of computational linguistics for Turkic languages and to the methodology of interpreting different machine learning models for text data.
More by Fail Gafarov
- Predicting Changes in Personality Behavior Under the Influence of an InterlocutorRussian Digital Libraries Journal · 2026
- RNN-Based Approach to Cognitive Architecture of Subject Behavior in Digital EnvironmentInternational Conference on Computer Science and Engineering · 2026
- СРАВНИТЕЛЬНЫЙ АНАЛИЗ МОДЕЛЕЙ АРХИТЕКТУРЫ TRANSFORMER В ЗАДАЧЕ УПРОЩЕНИЯ ТЕКСТАВысшая школа: научные исследования · 2025
- Explainable Artificial Intelligence Methods in Text Classification of Machine Learning Models for Tatar LanguageInternational Conference on Computer Science and Engineering · 2025 · Citations: 6
