Self-Paced Learning for Neural Machine Translation

TL;DR

This paper introduces a self-paced learning approach for neural machine translation that automatically assesses training example difficulty and adjusts learning, resulting in improved translation quality and faster convergence compared to traditional curriculum methods.

Contribution

It proposes a novel self-paced learning framework for NMT that eliminates reliance on handcrafted curricula by enabling the model to self-regulate learning based on confidence measures.

Findings

Outperforms strong baselines in translation quality

Achieves faster convergence during training

Demonstrates robustness across multiple translation tasks

Abstract

Recent studies have proven that the training of neural machine translation (NMT) can be facilitated by mimicking the learning process of humans. Nevertheless, achievements of such kind of curriculum learning rely on the quality of artificial schedule drawn up with the handcrafted features, e.g. sentence length or word rarity. We ameliorate this procedure with a more flexible manner by proposing self-paced learning, where NMT model is allowed to 1) automatically quantify the learning confidence over training examples; and 2) flexibly govern its learning via regulating the loss in each iteration step. Experimental results over multiple translation tasks demonstrate that the proposed model yields better performance than strong baselines and those models trained with human-designed curricula on both translation quality and convergence speed.