Zhu, Xunjie. Assessing compositionality and linguistic regularities in pretrained language representation models. Retrieved from https://doi.org/doi:10.7282/t3-2gfz-r678
DescriptionWhile in computer vision, models pretrained on ImageNet have long been used for transfer learning, in NLP, there have only recently been important breakthroughs in learning pretrained representation models beyond the level of words. The most well-known of these are the BERT and XLNet ones, which rely on a language modeling objective. Despite the strong gains that they enable in a number of downstream tasks, it is not known to what extent the models can be used as sentence embeddings capturing important semantic phenomena such as compositionality and analogy. In this thesis, we propose a method to assess this based on relationships between vectors. We find that the models reflect only some of the properties we consider