Uncovering divergent linguistic information in word embeddings with lessons for intrinsic and extrinsic evaluation
This work addresses the challenge of evaluating and optimizing word embeddings for natural language processing researchers, but it is incremental as it builds on existing embedding models.
The paper tackled the problem of word embeddings capturing divergent linguistic aspects by showing that a linear transformation can adjust similarity order to improve performance in specific aspects, achieving better results without external resources.
Following the recent success of word embeddings, it has been argued that there is no such thing as an ideal representation for words, as different models tend to capture divergent and often mutually incompatible aspects like semantics/syntax and similarity/relatedness. In this paper, we show that each embedding model captures more information than directly apparent. A linear transformation that adjusts the similarity order of the model without any external resource can tailor it to achieve better results in those aspects, providing a new perspective on how embeddings encode divergent linguistic information. In addition, we explore the relation between intrinsic and extrinsic evaluation, as the effect of our transformations in downstream tasks is higher for unsupervised systems than for supervised ones.