|
所在平台: Udemy |
课程主页: https://www.udemy.com/course/ahol-dl4nlp4/
课程评论：没有评论
Coursera 课程《深度学习在自然语言处理中的应用 - 第四部分》概述： 本课程是“深度学习在自然语言处理中的应用”系列课程的第四部分。课程旨在介绍跨语言基准测试和模型，为使用先进的深度学习模型进行多语言和跨语言的自然语言理解与生成提供基础。课程的动机来源于产品团队希望以经济高效的方式快速扩展到全球市场，以及同时向多个市场发布新功能的需求。 课程主要分为两个部分，都将涵盖跨语言模型和基准测试。 **第一部分：** * **跨语言基准数据集：** 重点介绍 XNLI 和 XGLUE。 * **跨语言模型：** 详细讲解 mBERT, XLM, Unicoder, XLM-R, 以及带适配器的 BERT 等模型。这些模型大部分是基于编码器的。 * **基础跨语言建模方法：** 探讨 translate-train, translate-test, multi-lingual translate-train-all, 以及 zero shot 跨语言迁移等技术。 **第二部分：** * **跨语言基准数据集：** 深入介绍 XTREME 和 XTREME-R。 * **跨语言模型：** 讲解 XNLG, mBART, InfoXLM, FILTER, 和 mT5 等模型。其中，InfoXLM 和 FILTER 是编码器独占模型，而 XNLG, mBART, 和 mT5 可用于编码器-解码器跨语言建模。 * **模型细节：** 对于每个模型，课程将讨论其特定的预训练损失、预训练策略、架构，以及在预训练和下游任务上取得的结果。
This course is a part of "Deep Learning for NLP" Series. In this course, I will introduce concepts like Cross lingual benchmarks and models. These concepts form the base for multi-lingual and cross-lingual processing using advanced deep learning models for natural language understanding and generation across languages.Often times, I hear from various product teams: "My product is in en-US only. I want to quickly scale to global markets with cost-effective solutions.", or "I have a new feature. How can I sim-ship to multiple markets?" This course is motivated by such needs. In this course the goal is to try to answer such questions.The course consists of two main sections as follows. In both the sections, I will talk about some cross-lingual models as well as benchmarks. In the first section, I will talk about cross-lingual benchmark datasets like XNLI and XGLUE. I will also talk about initial cross-lingual models like mBERT, XLM, Unicoder, XLM-R, and BERT with adaptors. Most of these models are encoder-based models. We will also talk about basic ways of cross-lingual modeling like translate-train, translate-test, multi-lingual translate-train-all, and zero shot cross-lingual transfer.In the second section, I will talk about cross-lingual benchmark datasets like XTREME and XTREME-R. I will also talk about cross-lingual models like XNLG, mBART, InfoXLM, FILTER and mT5. Some of these models are encoder-only models like InfoXLM or FILTER while others can be used for encoder-decoder cross-lingual modeling like XNLG, mBART and mT5. For each model, we will discuss specific pretraining losses, pretraining strategy, architecture and results obtained for pretraining as well as downstream tasks.
