Variational Language Concepts for Interpreting Foundation Language Models.

Area of Science:

• Natural Language Processing
• Machine Learning Interpretability
• Artificial Intelligence

Background:

• Foundation Language Models (FLMs) like BERT excel in NLP tasks.
• Current interpretability methods using attention weights offer only word-level insights.
• Word-level interpretations lack higher-level structure, hindering readability and intuition.

Purpose of the Study:

• To formally define and introduce the concept of 'conceptual interpretation'.
• To develop a novel framework for achieving concept-level interpretability in FLMs.
• To overcome the limitations of existing word-level interpretation techniques.

Main Methods:

• Proposed a variational Bayesian framework named VAriational Language Concept (VALC).
• Developed theoretical analyses to demonstrate VALC's optimality in finding language concepts.
• Validated the framework on diverse real-world datasets.

Main Results:

• VALC successfully moves beyond word-level interpretations to provide concept-level insights.
• Theoretical analysis confirms VALC identifies optimal concepts for interpreting FLM predictions.
• Empirical results demonstrate the framework's effectiveness across multiple datasets.

Conclusions:

• The VALC framework offers a significant advancement in FLM interpretability.
• Conceptual interpretation provides more intuitive and readable explanations for model predictions.
• This approach enhances the trustworthiness and usability of foundation language models.