Explaining Contextualized Word Embeddings in Biomedical Research - A Qualitative Investigation

Area of Science:

• Computational linguistics
• Artificial Intelligence
• Philosophy of language

Background:

• Contextualized word embeddings are effective quantitative tools for natural language processing tasks.
• Explainable AI (XAI) aims to build confidence and trust in AI systems through interpretable methods.
• Saussurean sign theory offers a framework for understanding language's differential structure.

Purpose of the Study:

• To investigate the application of Saussurean sign theory as a qualitative XAI method for word embeddings.
• To explore how linguistic theory can provide deeper insights into the inner workings of natural language processing models.
• To bridge the gap between quantitative performance and qualitative understanding in AI.

Main Methods:

• Applying Saussurean sign theory concepts to analyze word embedding structures.
• Examining the mathematical properties of word embeddings in relation to linguistic principles.
• Qualitative analysis of word embedding behavior through the lens of semiotics.

Main Results:

• The differential structure of language, as described by Saussure, aligns with the additive and subtractive properties of word embeddings.
• Saussurean principles offer a qualitative framework for interpreting the relationships and meanings captured by word embeddings.
• Mathematical structures of embeddings reflect underlying natural language structures.

Conclusions:

• Saussurean sign theory provides a valuable qualitative approach to enhance the explainability of word embeddings.
• Integrating linguistic theories with AI methods can foster greater trust and understanding in NLP.
• This approach offers new perspectives on the relationship between language, meaning, and AI models.