Ultradense Word Embeddings by Orthogonal Transformation
This work addresses efficiency and performance issues in NLP tasks for researchers and practitioners by providing a method to reduce embedding dimensionality without losing information.
The paper tackles the problem of high-dimensional embeddings in NLP by introducing DENSIFIER, a method that learns an orthogonal transformation to compress embeddings into an ultradense subspace 100 times smaller, achieving state-of-the-art results on a lexicon creation task and maintaining performance on sentiment analysis while improving training efficiency by an order of magnitude.
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.