WMT 2018 News (WMT 2018 News Translation Task)

新闻翻译是一项经常性的 WMT 任务。测试集是一组平行语料库，由大约 1500 个英语句子翻译成 5 种语言（中文、捷克语、爱沙尼亚语、德语、芬兰语、俄语、土耳其语）和 7 种语言中的每一种翻译成英语的另外 1500 个句子组成。这些句子选自数十家新闻网站，由专业翻译人员翻译。训练数据包括用于训练翻译模型的并行语料库、用于训练语言模型的单语语料库和用于调优的开发集。一些训练语料库与 WMT 2017 相同（Europarl、Common Crawl、SETIMES2、Yandex 提供的俄英平行数据、CMU 提供的 Wikipedia Headlines），一些是更新的（United Nations、CzEng v1.7、News Commentary v13、单语新闻数据）。此外，还添加了德语、芬兰语和爱沙尼亚语的欧盟新闻稿平行语料库。

News translation is a recurring WMT task. The test set consists of a parallel corpus, which includes approximately 1,500 sentence pairs where English sentences are translated into 7 languages (Chinese, Czech, Estonian, German, Finnish, Russian, Turkish), plus another 1,500 sentence pairs translated from each of the 7 aforementioned languages into English. All sentences are selected from dozens of news websites and translated by professional translators. The training data encompasses three categories: parallel corpora for training translation models, monolingual corpora for training language models, and a development set for model fine-tuning. Some of the training corpora are identical to those used in WMT 2017, including Europarl, Common Crawl, SETIMES2, Russian-English parallel data provided by Yandex, and Wikipedia Headlines provided by CMU. Others are updated versions, such as the United Nations corpus, CzEng v1.7, News Commentary v13, and monolingual news data. Furthermore, parallel corpora of EU press releases for German, Finnish and Estonian are added to the training data.

搜集汇总

背景与挑战

背景概述

WMT 2018 News数据集是一个用于新闻翻译任务的机器翻译数据集，其测试集包含约1500个英语句子翻译成5种语言以及7种语言翻译成英语的平行语料库，句子选自新闻网站并由专业翻译人员翻译。训练数据包括并行语料库、单语语料库和开发集，部分语料库与WMT 2017相同或更新，由爱丁堡大学等机构于2018年发布。