Ultradense Word Embeddings by Orthogonal Transformation

TL;DR

This paper introduces DENSIFIER, a method that transforms word embeddings into an ultradense subspace, improving task-specific performance and training efficiency in NLP applications.

Contribution

DENSIFIER is a novel orthogonal transformation technique that creates ultradense embeddings, achieving state-of-the-art results and increased training efficiency.

Findings

Ultradense embeddings reach state-of-the-art on lexicon creation tasks.

No information loss in ultradense subspace for sentiment analysis.

Training is significantly more efficient with ultradense embeddings.

Abstract

Embeddings are generic representations that are useful for many NLP tasks. In this paper, we introduce DENSIFIER, a method that learns an orthogonal transformation of the embedding space that focuses the information relevant for a task in an ultradense subspace of a dimensionality that is smaller by a factor of 100 than the original space. We show that ultradense embeddings generated by DENSIFIER reach state of the art on a lexicon creation task in which words are annotated with three types of lexical information - sentiment, concreteness and frequency. On the SemEval2015 10B sentiment analysis task we show that no information is lost when the ultradense subspace is used, but training is an order of magnitude more efficient due to the compactness of the ultradense space.