On Dimensional Linguistic Properties of the Word Embedding Space
share
Word embeddings have become a staple of several natural language processing tasks, yet much remains to be understood about their properties. In this work, we analyze word embeddings in terms of their principal components and arrive at a number of novel conclusions. In particular, we characterize the utility of variance explained by the principal components (widely used as a fundamental tool to assess the quality of the resulting representations) as a proxy for downstream performance. Further, through dimensional linguistic probing of the embedding space, we show that the syntactic information captured by a principal component does not depend on the amount of variance it explains. Consequently, we investigate the limitations of variance based embedding post-processing techniques and demonstrate that such post-processing is counter-productive in a number of scenarios such as sentence classification and machine translation tasks. Finally, we offer a few guidelines on variance based embedding post-processing. We have released the source code along with the paper.
READ FULL TEXT