<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[liuy9803的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/liuy9803</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; liuy9803]]></copyright><item><title><![CDATA[论文阅读——“Bilingual Expert” Can Find Translation Errors]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/107227433</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/107227433</guid><author>liuy9803</author><pubDate>Mon, 13 Jul 2020 16:25:37 +0800</pubDate><description><![CDATA[https://www.researchgate.net/publication/326621822_Bilingual_Expert_Can_Find_Translation_Errors

“双语专家”可以发现翻译错误

Github：https://github.com/lovecambi/qebrain



Abstract




























































...]]></description><category></category></item><item><title><![CDATA[论文阅读——A Simple and Effective Approach to Automatic Post-Editing with Transfer Learning]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/106949383</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/106949383</guid><author>liuy9803</author><pubDate>Fri, 26 Jun 2020 18:13:17 +0800</pubDate><description><![CDATA[https://www.aclweb.org/anthology/P19-1292.pdf

一个简单有效的迁移学习自动后编辑方法

Github：https://github.com/deep-spin/OpenNMT-APE








]]></description><category></category></item><item><title><![CDATA[论文阅读——用于自动后编辑的单语和双语神经机器翻译模型的对数线性组合]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/106264605</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/106264605</guid><author>liuy9803</author><pubDate>Fri, 22 May 2020 00:21:05 +0800</pubDate><description><![CDATA[Log-linear Combinations of Monolingual and Bilingual Neural Machine Translation Models for Automatic Post-Editing

https://arxiv.org/pdf/1605.04800.pdf





本文介绍了AMU (Adam Mickiewicz University) 团队提交的 WMT 2016 自动后编辑 APE 任务，探索神经翻译模型在 APE 中的应用，使用不同的模型作为对数线性.]]></description><category></category></item><item><title><![CDATA[论文阅读——Pushing the Limits of Translation Quality Estimation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/106151821</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/106151821</guid><author>liuy9803</author><pubDate>Tue, 19 May 2020 10:23:20 +0800</pubDate><description><![CDATA[https://www.aclweb.org/anthology/Q17-1015.pdf

挑战翻译质量评估的极限

Github：https://github.com/Unbabel/OpenKiwi



翻译质量评估可以极大减少后编辑工作，本文使用单词级质量评估和自动后编辑之间的协同，将一个新模型堆叠到具有丰富特征的单词级质量评估系统中，使用自动后编辑系统的输出作为一个额外的功能，取得了非常好的效果，在 WMT16 单词级 F1^MULT 为57.47%（比 sota 增加了7.95%），句子级 .]]></description><category></category></item><item><title><![CDATA[论文阅读——Predictor-Estimator using Multilevel Task Learning with Stack Propagation for Neural QE]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/106130905</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/106130905</guid><author>liuy9803</author><pubDate>Fri, 15 May 2020 11:33:59 +0800</pubDate><description><![CDATA[https://www.aclweb.org/anthology/W17-4763.pdf

使用多任务学习与堆栈传播的质量估计方法Predictor-Estimator

Github：https://github.com/Unbabel/OpenKiwi



两段神经网络 QE 模型，多任务学习估计句子、单词和短语的翻译质量，基于端到端的堆叠神经模型 Predictor-Estimator，训练时使用 stack propagation，可以在单一模式下联合训练单词预测模型和 QE 模型。

Pr..]]></description><category></category></item><item><title><![CDATA[OpenKiwi 使用的 QE 系统]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/106047616</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/106047616</guid><author>liuy9803</author><pubDate>Tue, 12 May 2020 00:40:19 +0800</pubDate><description><![CDATA[QUETCH

QUality Estimation from scraTCH

http://www.statmt.org/wmt15/pdf/WMT37.pdf

模型输入：对于 MT 中的每个单词，围绕该单词固定大小的窗口，与源句中对齐单词的窗口相连接，得到该目标单词的双语上下文向量。

模型结构：单隐层 MLP，lookup 得到词向量（训练中更新），激活函数为 tanh。

模型输出：softmax 计算每个单词的 OK/BAD 的概率，该模型单独训练以预测源句标签、gap 标签和目标标签。


]]></description><category></category></item><item><title><![CDATA[论文阅读——Automatic Testing and Improvement of Machine Translation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105933219</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105933219</guid><author>liuy9803</author><pubDate>Sat, 09 May 2020 01:32:45 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1910.02688.pdf

机器翻译的自动测试和改进

Github：https://github.com/zysszy/TransRepair（无代码）



本文提出一种测试和修复机器翻译系统一致性的自动方法——TransRepair，在没有标准翻译的情况下，结合 metamorphic testing 蜕变测试的 mutation 变化检测不...]]></description><category></category></item><item><title><![CDATA[论文阅读——OpenKiwi: An Open Source Framework for Quality Estimation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105963461</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105963461</guid><author>liuy9803</author><pubDate>Thu, 07 May 2020 15:17:04 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1902.08646.pdf

OpenKiwi：一个用于质量评估的开源框架

Github：https://github.com/Unbabel/OpenKiwi



OpenKiwi 是一个用于翻译质量评估的开源 Pytorch 框架，支持单词级和句子级质量评估系统的训练和测试，实现了 WMT 2015-18 中一些获奖的质量评估系统。使用 WM...]]></description><category></category></item><item><title><![CDATA[论文阅读——Structure-Invariant Testing for Machine Translation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105914175</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105914175</guid><author>liuy9803</author><pubDate>Tue, 05 May 2020 10:11:49 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1907.08710.pdf

机器翻译的结构不变测试

Github：https://github.com/PinjiaHe/StructureInvariantTesting



神经机器翻译 NMT 模型的复杂性和难解性使得它们远远不够健壮，这个问题也没有得到充分的研究。因此作者引入 structure-invariant testing 结构...]]></description><category></category></item><item><title><![CDATA[论文阅读——A Call for Clarity in Reporting BLEU Scores]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105897292</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105897292</guid><author>liuy9803</author><pubDate>Sat, 02 May 2020 23:29:46 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1804.08771v1.pdf

Github：https://github.com/awslabs/sockeye/tree/master/contrib/sacrebleu



由于在不同的环境下经常产生新的测试集，机器翻译领域面临着新的挑战，而且在如何进行度量评价方面缺乏共识。常用的 BLEU score 会随着 parameterizati...]]></description><category></category></item><item><title><![CDATA[论文阅读——A Study of Cross-Lingual Ability and Language-specific Information in Multilingual BERT]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105760348</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105760348</guid><author>liuy9803</author><pubDate>Tue, 28 Apr 2020 18:33:57 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/2004.09205.pdf

多语言BERT中跨语言能力和语言特定信息的研究（未开源）



Multilingual BERT (m-BERT Devlin et al., 2019) 在跨语言迁移任务中表现很好，优于静态非上下文词嵌入。比较相同数据的非上下文和上下文模型的跨语言性能，发现影响迁移性的关键因素是数据集和上下文窗口大小。观察多语 BE...]]></description><category></category></item><item><title><![CDATA[论文阅读——Testing Machine Translation via Referential Transparency]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105787024</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105787024</guid><author>liuy9803</author><pubDate>Tue, 28 Apr 2020 00:55:11 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/2004.10361.pdf

通过引用透明性测试机器翻译

Github：https://github.com/ReferentialTransparency/RTI（无代码）



由于神经网络的复杂性和难解性，现代机器翻译软件还远远不够健壮，因此作者引入一种简单而广泛用于验证机器翻译软件的方法 referentially transparent ...]]></description><category></category></item><item><title><![CDATA[论文阅读——Topic Modeling in Embedding Spaces]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105670825</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105670825</guid><author>liuy9803</author><pubDate>Thu, 23 Apr 2020 18:21:08 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1907.04907.pdf

嵌入空间中的主题建模

Github：https://github.com/adjidieng/ETM
















































]]></description><category></category></item><item><title><![CDATA[论文阅读——Learning Contextualized Sentence Representations for Document-Level Neural Machine Translation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105518599</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105518599</guid><author>liuy9803</author><pubDate>Wed, 15 Apr 2020 10:40:36 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/2003.13205.pdf

学习文档级神经机器翻译的上下文句子表征（未开源）
































]]></description><category></category></item><item><title><![CDATA[论文阅读——Aligned Cross Entropy for Non-Autoregressive Machine Translation]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105505965</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105505965</guid><author>liuy9803</author><pubDate>Tue, 14 Apr 2020 18:04:45 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/2004.01655.pdf

对齐交叉熵的非自回归机器翻译（未开源）



非自回归机器翻译模型对整个目标序列进行并行预测，极大地提高了解码速度。然而由于模型中缺少自回归因子，难以对词序建模，且交叉熵损失会严重影响词序的微小变化，因此作者提出使用 aligned cross entropy 对齐交叉熵 AXE 作为非自回归模型的损失函数，AXE 使用...]]></description><category></category></item><item><title><![CDATA[论文阅读——Unsupervised Neural Machine Translation with Indirect Supervision]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105485949</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105485949</guid><author>liuy9803</author><pubDate>Mon, 13 Apr 2020 15:49:07 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/2004.03137.pdf

带有间接监督的无监督神经机器翻译（未开源）












































]]></description><category></category></item><item><title><![CDATA[论文阅读——面向任务的对话中无监督度量的相关性对自然语言生成的评估]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105124390</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105124390</guid><author>liuy9803</author><pubDate>Tue, 07 Apr 2020 17:50:06 +0800</pubDate><description><![CDATA[Relevance of Unsupervised Metrics in Task-Oriented Dialogue for Evaluating Natural Language Generation

https://arxiv.org/pdf/1706.09799.pdf

实现地址：https://github.com/Maluuba/nlg-eval













...]]></description><category></category></item><item><title><![CDATA[论文阅读——Transformer++]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/105083730</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/105083730</guid><author>liuy9803</author><pubDate>Thu, 26 Mar 2020 17:28:35 +0800</pubDate><description><![CDATA[Transformer++

https://arxiv.org/pdf/2003.04974.pdf






























]]></description><category></category></item><item><title><![CDATA[论文阅读——使用聚类评估LDA的稳定性并提高其可靠性]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/104969156</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/104969156</guid><author>liuy9803</author><pubDate>Thu, 19 Mar 2020 17:12:05 +0800</pubDate><description><![CDATA[Improving Reliability of Latent Dirichlet Allocation by Assessing Its Stability Using Clustering Techniques on Replicated Runs

https://arxiv.org/pdf/2003.04980.pdf

实现地址：https://github.com/JonasRiege...]]></description><category></category></item><item><title><![CDATA[论文阅读——DYNAMIC EVALUATION OF NEURAL SEQUENCE MODELS]]></title><link>https://blog.youkuaiyun.com/liuy9803/article/details/104904405</link><guid>https://blog.youkuaiyun.com/liuy9803/article/details/104904405</guid><author>liuy9803</author><pubDate>Tue, 17 Mar 2020 00:48:01 +0800</pubDate><description><![CDATA[https://arxiv.org/pdf/1709.07432.pdf

神经序列模型的动态评估方法

实现地址：https://github.com/benkrause/dynamic-evaluation



作者提出一种使用动态评价改进神经序列模型的方法，模型使用梯度下降来适应最近的序列，从而使它们为再次发生的序列模式 patterns 分配更高的概率。动态评估将 Penn Treeb...]]></description><category></category></item></channel></rss>