BriefGPT.xyz
Mar, 2024
自生成的重播记忆用于持续神经机器翻译
Self-generated Replay Memories for Continual Neural Machine Translation
HTML
PDF
Michele Resta, Davide Bacciu
TL;DR
通过利用编码器-解码器Transformers的生成能力,我们提出了一种新的方法来持续学习神经机器翻译系统,以有效地从包含不同语言的经验流中学习,通过使用模型作为生成器填充重播记忆来对抗灾难性遗忘,而无需明确记忆训练数据。
Abstract
Modern
neural machine translation
systems exhibit strong performance in several different languages and are constantly improving. Their ability to learn continuously is, however, still severely limited by the
catastroph
→