An Empirical Analysis Towards Replacing Vocabulary-Rigid Embeddings by a Vocabulary-Free Mechanism

This paper addresses the limitations of subword based models in NLP by aligning the word embedding layer of a vocabulary-rigid transformer model to a vocabulary-free one. In order to do so, a CNN is trained to mimic the word embeddings layer of a BERT model, using a sequence of byte tokens as input. The study compares cosine-based and Euclidean-based loss functions for training the student network and finds better results with cosine-based metrics. The research contributes techniques for re-training transformer embedding layers and provides insights into loss function selection. The findings have implications for developing flexible and robust NLP models.