Semantic Ax

Semantic axes research aims to uncover the underlying interpretable structures within high-dimensional word embeddings, primarily using Independent Component Analysis (ICA) to identify independent semantic features. Current work focuses on improving the reliability and interpretability of these axes, including methods to optimize their order and demonstrate consistency across languages and modalities. This research enhances our understanding of how meaning is represented in computational models, with implications for improving natural language processing tasks and providing insights into the evolution of language and social biases.