人工智能培训

搜索

深度学习论文:用词嵌入技术平衡神经机器翻译中的性别偏见(Equalizing Gender Biases in Neural Machine Translati

[复制链接]
comceo 发表于 2019-1-11 11:21:41 | 显示全部楼层 |阅读模式
comceo 2019-1-11 11:21:41 339 0 显示全部楼层
深度学习论文:用词嵌入技术平衡神经机器翻译中的性别偏见(Equalizing Gender Biases in Neural Machine Translation with Word  Embeddings Techniques)神经机器翻译显着推动了该领域的质量。然而,翻译仍然存在重大问题,其中一个是公平。神经模型在大文本语料库上进行训练,其中包含偏见和刻板印象。因此,模型继承了这些社会偏见。最近的方法已经显示出减少其他自然语言处理应用(例如单词嵌入)中的性别偏差的结果。我们利用单词嵌入用于神经机器翻译来提出第一个脱映机器翻译系统这一事实。具体而言,我们提出,实验和分析了在变换器翻译体系结构中Glove嵌入的双向偏移技术的集成。我们在通用英语 - 西班牙语任务上评估我们提出的系统,显示最多一个BLEU点的收益。至于性别偏见评估,我们生成一组测试职业,我们表明我们提出的系统学习均衡了基线系统的现有偏差。
Neural machine translation has significantly pushed forward the quality ofthe field.However, there are remaining big issues with the translations andone of them is fairness.Neural models are trained on large text corpora whichcontains biases and stereotypes.As a consequence, models inherit these socialbiases.Recent methods have shown results in reducing gender bias in othernatural language processing applications such as word embeddings.We takeadvantage of the fact that word embeddings are used in neural machinetranslation to propose the first debiased machine translation system.Specifically, we propose, experiment and analyze the integration of twodebiasing techniques over GloVe embeddings in the Transformer translationarchitecture.We evaluate our proposed system on a generic English-Spanishtask, showing gains up to one BLEU point.As for the gender bias evaluation, wegenerate a test set of occupations and we show that our proposed system learnsto equalize existing biases from the baseline system.深度学习论文:用词嵌入技术平衡神经机器翻译中的性别偏见(Equalizing Gender Biases in Neural Machine Translation with Word  Embeddings Techniques) erRlX443R3X8xwRe.jpg
URL地址:https://arxiv.org/abs/1901.03116     ----pdf下载地址:https://arxiv.org/pdf/1901.03116    ----深度学习论文:用词嵌入技术平衡神经机器翻译中的性别偏见(Equalizing Gender Biases in Neural Machine Translation with Word  Embeddings Techniques)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

comceo当前离线
新手上路

查看:339 | 回复:0

快速回复 返回顶部 返回列表