Transparent AI for mathematics: transformer-based large language models for mathematical entity relationship...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Mathematical text presents unique challenges for natural language understanding due to specialized entities and complex relationships.
• Existing methods struggle to accurately interpret the structure and meaning within mathematical problems.

Purpose of the Study:

• To formulate mathematical problem interpretation as a Mathematical Entity Relation Extraction (MERE) task.
• To develop and evaluate transformer-based models for automated MERE.
• To enhance model interpretability using Explainable Artificial Intelligence (XAI).

Main Methods:

• Formulation of MERE task: operands as entities, operators as relationships.
• Application of transformer-based models, specifically Bidirectional Encoder Representations from Transformers (BERT).
• Integration of Explainable Artificial Intelligence (XAI) techniques, notably Shapley Additive Explanations (SHAP).

Main Results:

• BERT achieved state-of-the-art performance with 99.39% accuracy in MERE.
• SHAP analysis provided insights into feature importance, explaining model predictions.
• Demonstrated the effectiveness of combining transformer models with XAI for mathematical text understanding.

Conclusions:

• The proposed MERE framework offers an effective and interpretable solution for mathematical text understanding.
• The approach supports advancements in automated problem solving, knowledge graph construction, and intelligent educational systems.
• Explainability enhances trust and transparency in AI models applied to complex domains like mathematics.