Are there identifiable structural parts in the sentence embedding whole?
This work addresses the interpretability of sentence embeddings for NLP researchers, but it is incremental as it builds on existing embedding analysis methods.
The paper investigates whether sentence embeddings from transformer models contain separable overlapping layers of information, specifically about chunks and their structural/semantic properties, and demonstrates this using datasets with known chunk structures and linguistic intelligence tasks.
Sentence embeddings from transformer models encode in a fixed length vector much linguistic information. We explore the hypothesis that these embeddings consist of overlapping layers of information that can be separated, and on which specific types of information -- such as information about chunks and their structural and semantic properties -- can be detected. We show that this is the case using a dataset consisting of sentences with known chunk structure, and two linguistic intelligence datasets, solving which relies on detecting chunks and their grammatical number, and respectively, their semantic roles, and through analyses of the performance on the tasks and of the internal representations built during learning.