The Low-Dimensional Linear Geometry of Contextualized Word Representations.
Evan HernandezJacob AndreasPublished in: CoRR (2021)
Keyphrases
- low dimensional
- linear subspace
- appearance manifolds
- high dimensional
- high dimensional data
- dimensionality reduction
- low dimensional manifolds
- data points
- manifold learning
- three dimensional
- input space
- n gram
- co occurrence
- underlying manifold
- principal component analysis
- linear systems
- feature space
- word sense disambiguation
- feature representation
- dimension reduction
- graph embedding
- multidimensional scaling
- nonlinear manifold
- geometric structure
- euclidean space
- closed form
- pairwise
- pattern recognition
- similarity measure