Decompose Embeddings Into Explainable Semantic Features

Supervisor(s): Andrianos Michail & Dr. Juri Opitz &Dr. Simon Clematide

The S3BERT method decomposes an embedding such that its parts reflect different semantic aspects. This way, we can not only say, “X and Y are similar”, but also, e.g., “X and Y are similar in that they share the same topic”. In this thesis we want to design simple metrics that measure similarity of your features (e.g. Named Entity Overlap) and apply the S³BERT method to learn your features with a SotA embedding model.

If interested, please send an email addressed to all three of us for maximum visibility.

• Machine Learning
• Python/PyTorch