Invariance and identifiability issues for word embeddings

Carrington, Rachel; Bharath, Karthik; Preston, Simon

Abstract

Word embeddings are commonly obtained as optimisers of a criterion function f of 1 a text corpus, but assessed on word-task performance using a different evaluation 2 function g of the test data. We contend that a possible source of disparity in 3 performance on tasks is the incompatibility between classes of transformations that 4 leave f and g invariant. In particular, word embeddings defined by f are not unique; 5 they are defined only up to a class of transformations to which f is invariant, and 6 this class is larger than the class to which g is invariant. One implication of this is 7 that the apparent superiority of one word embedding over another, as measured by 8 word task performance, may largely be a consequence of the arbitrary elements 9 selected from the respective solution sets. We provide a formal treatment of the 10 above identifiability issue, present some numerical examples, and discuss possible 11 resolutions.

Citation

Carrington, R., Bharath, K., & Preston, S. (2019, December). Invariance and identifiability issues for word embeddings. Presented at NeurIPS 2019, Vancouver, Canada

• Presentation Conference Type: Conference Paper (published)
• Conference Name: NeurIPS 2019
• Start Date: Dec 8, 2019
• End Date: Dec 14, 2019
• Acceptance Date: Sep 3, 2019
• Online Publication Date: Dec 14, 2019
• Publication Date: Dec 14, 2019
• Deposit Date: Oct 16, 2019
• Publicly Available Date: Feb 15, 2020
• Book Title: Advances in Neural Information Processing Systems 32 (NIPS 2019)
• Public URL: https://nottingham-repository.worktribe.com/output/2848777
• Publisher URL: https://papers.nips.cc/paper/9650-invariance-and-identifiability-issues-for-word-embeddings
• Related Public URLs: https://papers.nips.cc/book/advances-in-neural-information-processing-systems-32-2019
• Contract Date: Oct 16, 2019