Investigating semantic subspaces of Transformer sentence embeddings through linear structural probing

Abstract: The question of what kinds of linguistic information are encoded in different layers of Transformer-based LLMs is of considerable interest for the NLP community. Existing work, however, has overwhelmingly focused on word-level representations and encoder-only LLMs with the masked-token training objective. In this paper, we present experiments with semantic structural probing, a method for studying sentence-level representations via finding a subspace of the embedding space that provides suitable task-specific pairwise distances between data-points. We apply our method to LLMs from different families (encoder-only, decoder-only, encoder-decoder) and of different sizes in the context of two tasks, semantic textual similarity and natural-language inference. We find that model families differ substantially in their performance and layer dynamics, but that the results are largely model-size invariant.