Efficient Non-parametric Estimation of Multiple Embeddings per Word in Vector Space
Abstract
There is rising interest in vector-space word embeddings and their use in NLP, especially given recent methods for their fast estimation at very large scale. Nearly all this work, however, assumes a single vector per word type ignoring polysemy and thus jeopardizing their usefulness for downstream tasks. We present an extension to the Skip-gram model that efficiently learns multiple embeddings per word type. It differs from recent related work by jointly performing word sense discrimination and embedding learning, by non-parametrically estimating the number of senses per word type, and by its efficiency and scalability. We present new state-of-the-art results in the word similarity in context task and demonstrate its scalability by training with one machine on a corpus of nearly 1 billion tokens in less than 6 hours.
Cite
@article{arxiv.1504.06654,
title = {Efficient Non-parametric Estimation of Multiple Embeddings per Word in Vector Space},
author = {Arvind Neelakantan and Jeevan Shankar and Alexandre Passos and Andrew McCallum},
journal= {arXiv preprint arXiv:1504.06654},
year = {2015}
}
Comments
In Conference on Empirical Methods in Natural Language Processing, 2014