Non-Autoregressive Neural Machine Translation

[flrngel]

https://arxiv.org/abs/1711.02281

Abstract

Features

• Non-Autoregressive (means output selves doesn't have dependency)
• Parallel outputs

How

• Knowledge distillation
• Input token fertilities
• Policy Gradient

1. Introduction

Paper model uses CNN and SAN (Transformer) to avoid autoregressive

2. Background

2.1. Autoregressive Neural Machine Translation

• Transformer's masking is better than CNN

2.2. Non-Autoregressive decoding

Problems of beam-search

• suffers from diminishing returns with respect to beam size
• limits search parallelism

They made output length variable T as probabilistic variable

2.3. The multimodality problem

Multimodality problem is problem of "high multimodal distribution of target translation"

3. The non-autoregressive transformer

3.3. Modeling fertility to tackle the multimodality problem

Used IBM Model 2 to use fertilities.

Definition of fertilities and it's benefit

• Definition: number of input word has been copied
• Provides natural factorization that dramatically reduces mode space
• Allows decoder easier

3.4. Translation predictor and the decoding process

• Argmax decoding
• Average decoding
• Noisy parallel decoding

4. Training

I didn't like this section

4.2. Fine-Tuning

Uses KL Divergence, RL, backpropagation

Word-level knowledge distillation (Teacher)

External fertility inference model

Todo

• (3.4) Search about average decoding and noisy parallel decoding