Improving Word Embedding Factorization for Compression Using Distilled Nonlinear Neural Decomposition

TL;DR

This paper introduces a novel embedding compression method using low-rank matrix decomposition and knowledge distillation, significantly reducing memory usage while improving translation and language modeling performance.

Contribution

It proposes a simple, effective embedding compression technique that outperforms complex state-of-the-art methods in translation and language modeling tasks.

Findings

Higher BLEU scores in translation tasks

Lower perplexity in language modeling

Effective compression with a single parameter

Abstract

Word-embeddings are vital components of Natural Language Processing (NLP) models and have been extensively explored. However, they consume a lot of memory which poses a challenge for edge deployment. Embedding matrices, typically, contain most of the parameters for language models and about a third for machine translation systems. In this paper, we propose Distilled Embedding, an (input/output) embedding compression method based on low-rank matrix decomposition and knowledge distillation. First, we initialize the weights of our decomposed matrices by learning to reconstruct the full pre-trained word-embedding and then fine-tune end-to-end, employing knowledge distillation on the factorized embedding. We conduct extensive experiments with various compression rates on machine translation and language modeling, using different data-sets with a shared word-embedding matrix for both embedding and vocabulary projection matrices. We show that the proposed technique is simple to replicate, with one fixed parameter controlling compression size, has higher BLEU score on translation and lower perplexity on language modeling compared to complex, difficult to tune state-of-the-art methods.