Near-lossless Binarization of Word Embeddings

Abstract : Word embeddings are commonly used as a starting point in many NLP models to achieve state-of-the-art performances. However, with a large vocabulary and many dimensions, these floating-point representations are expensive both in terms of memory and calculations which makes them unsuitable for use on low-resource devices. The method proposed in this paper transforms real-valued embeddings into binary embeddings while preserving semantic information, requiring only 128 or 256 bits for each vector. This leads to a small memory footprint and fast vector operations. The model is based on an autoencoder architecture, which also allows to reconstruct original vectors from the binary ones. Experimental results on semantic similarity, text classification and sentiment analysis tasks show that the binarization of word em-beddings only leads to a loss of ∼2% in accuracy while vector size is reduced by 97%. Furthermore, a top-k benchmark demonstrates that using these binary vectors is 30 times faster than using real-valued vectors.
Document type :
Conference papers
Complete list of metadatas

Cited literature [30 references]  Display  Hide  Download

https://hal-ujm.archives-ouvertes.fr/ujm-02010043
Contributor : Christophe Gravier <>
Submitted on : Wednesday, February 6, 2019 - 5:47:22 PM
Last modification on : Friday, February 22, 2019 - 1:18:41 AM
Long-term archiving on : Tuesday, May 7, 2019 - 3:36:06 PM

File

1803.09065.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : ujm-02010043, version 1

Citation

Julien Tissier, Christophe Gravier, Amaury Habrard. Near-lossless Binarization of Word Embeddings. 33rd AAAI Conference on Artificial Intelligence (AAAI-19), Jan 2019, Honolulu, HI, United States. ⟨ujm-02010043⟩

Share

Metrics

Record views

45

Files downloads

28