Are All Good Word Vector Spaces Isomorphic?
This addresses a key problem in multilingual NLP by identifying practical factors affecting alignment, though it is incremental as it builds on existing assumptions about isomorphism.
The paper investigates whether poor performance in cross-lingual word vector alignment is due to non-isomorphic spaces, finding that variance is largely attributed to monolingual resource size and training properties like under-training, rather than just typological differences.
Existing algorithms for aligning cross-lingual word vector spaces assume that vector spaces are approximately isomorphic. As a result, they perform poorly or fail completely on non-isomorphic spaces. Such non-isomorphism has been hypothesised to result from typological differences between languages. In this work, we ask whether non-isomorphism is also crucially a sign of degenerate word vector spaces. We present a series of experiments across diverse languages which show that variance in performance across language pairs is not only due to typological differences, but can mostly be attributed to the size of the monolingual resources available, and to the properties and duration of monolingual training (e.g. "under-training").