The Effect of Domain and Diacritics in Yorùbá-English Neural Machine Translation
Saarland University(萨尔大学);Masakhane NLP(Masakhane自然语言处理团队);Max Planck Institute for Informatics(马克斯·普朗克信息学研究所);Alamoja Yoruba(阿拉莫贾约鲁巴组织);Yobamoodua Cultural Heritage (YMCH)(约巴穆杜阿文化遗产(YMCH));Defence Space Administration(国防太空管理局);University of Ibadan(伊巴丹大学);DFKI GmbH(德国人工智能研究中心(DFKI))
David I. Adelani, Dana Ruiter, Jesujoba O. Alabi, Damilola Adebonojo, Adesina Ayeni, Mofe Adeyemi, Ayodele Awokoya, Cristina España-Bonet
Massively multilingual machine translation (MT) has shown impressive capabilities, including zero and few-shot translation between low-resource language pairs. However, these models are often evaluated on high-resource languages with the assumption that they generalize to low-resource ones. The difficulty of evaluating MT models on low-resource pairs is often due to lack of standardized evaluation datasets. In this paper, we present MENYO-20k, the first multi-domain parallel corpus with a special focus on clean orthography for Yorùbá--English with standardized train-test splits for benchmarking. We provide several neural MT benchmarks and compare them to the performance of popular pre-trained (massively multilingual) MT models both for the heterogeneous test set and its subdomains. Since these pre-trained models use huge amounts of data with uncertain quality, we also analyze the effect of diacritics, a major characteristic of Yorùbá, in the training data. We investigate how and when this training condition affects the final quality and intelligibility of a translation. Our models outperform massively multilingual models such as Google ($+8.7$ BLEU) and Facebook M2M ($+9.1$ BLEU) when translating to Yorùbá, setting a high quality benchmark for future research.
YFACC: A Yorùbá speech-image dataset for cross-lingual keyword localisation through visual grounding
Stellenbosch University(斯泰伦博斯大学);University Politehnica of Bucharest(布加勒斯特理工大学)
视觉接地语音(VGS)模型在与未标记的口头字幕配对的图像上进行训练。这种模型可以用于在不可能获得标记数据的环境中构建语音系统,例如用于记录非书面语言。然而,大多数VGS研究都是用英语或其他高资源语言进行的。本文试图弥补这一缺陷。我们收集并发布了一个新的单扬声器数据集的音频字幕为6 k Flickr图像在Yorkebá-一个真正的低资源语言讲尼日利亚。我们训练一个基于注意力的VGS模型,其中图像自动标记有英语视觉标签,并与Yorkebá话语配对。这使得跨语言的关键字本地化:一个书面的英语查询被检测到,并在Yorkebá语音定位。为了量化较小数据集的影响,我们将其与在类似和更多数据上训练的英语系统进行了比较。我们希望这个新的数据集将刺激对真正的低资源语言使用VGS模型的研究。
Concept Direction Reliability Across Languages with Different Tokenizer Fertility
University of Cape Town(开普敦大学);African Institute for Mathematical Sciences(非洲数学科学研究所);Bayero University Kano(卡诺巴耶罗大学);Federal University Dutse(联邦杜策大学);University of Vienna(维也纳大学)
Efficient Convolutional Neural Networks for Diacritic Restoration
AWS, Amazon AI(亚马逊人工智能 AWS);The George Washington University(乔治华盛顿大学);Princess Nourah Bint Abdul Rahman University(诺拉·宾特·阿卜杜勒拉赫曼公主大学)
随着机器理解书面文本的需求不断增长,变音符号恢复变得越来越重要。该任务通常被建模为序列标记问题,目前双向长短期记忆(BiLSTM)模型提供了最先进的结果。最近,Bai et al.(2018)展示了时间卷积神经网络(TCN)在性能和计算资源方面优于递归神经网络(RNN)的序列建模优势。由于变音恢复的好处,从以前的以及随后的时间步,我们进一步应用和评估的TCN的一个变种,tuminusal TCN(A-TCN),它结合了上下文中的两个方向(以前和未来),而不是严格纳入以前的上下文的情况下,TCN。A-TCN在三种不同语言(阿拉伯语、约鲁巴语和越南语)的变音符号化方面比TCN有显著的改进。此外,A-TCN和BiLSTM具有相当的性能,使A-TCN成为BiLSTM的有效替代方案,因为卷积可以并行训练。A-TCN在推理时间上明显快于BiLSTM(每分钟变音文本量提高了270%-334%)。
Dialectal and Low-Resource Machine Translation for Aromanian
Tudor Vianu National College of Computer Science(图多尔·维亚努国家计算机科学学院);West University of Timișoara(蒂米什瓦拉西部大学);Human Language Technologies Research Center, University of Bucharest(布加勒斯特大学人类语言技术研究中心)
AFRIDOC-MT: Document-level MT Corpus for African Languages
Masakhane NLP;Saarland University(萨尔兰大学);DFKI GmbH;Barcelona Supercomputing Center(巴塞罗那超级计算中心);Inria(法国国家信息与自动化技术研究院);Mila, McGill University & Canada CIFAR AI Chair(Mila,麦吉尔大学及加拿大CIFAR人工智能主席);University of Ibadan(伊巴丹大学);National Institute of Informatics, Japan(日本国立信息研究所);Selcom Tanzania(坦桑尼亚Selcom);Imperial College London(伦敦帝国理工学院);University of KwaZulu-Natal(南非夸祖鲁-纳塔尔大学);Loughborough University(洛桑大学);University of Cambridge(剑桥大学)
YoNER: A New Yorùbá Multi-domain Named Entity Recognition Dataset
University of Ibadan(伊巴丹大学);Saarland University(萨尔大学);Atiba University(阿提巴大学);Mila - Quebec AI Institute(Mila - 魁北克人工智能研究所);McGill University(麦吉尔大学);Canada CIFAR AI Chair(加拿大CIFAR人工智能主席)
Measuring the Tokenization Premium: A Cost Audit for Underserved Language Communities
John Jay College of Criminal Justice, City University of New York(纽约城市大学约翰杰刑事司法学院);The City College of New York, City University of New York(纽约城市大学城市学院);University of Texas at San Antonio(德克萨斯大学圣安东尼奥分校)
Improved Visually Prompted Keyword Localisation in Real Low-Resource Settings
Electrical and Electronic Engineering(电气与电子工程);Stellenbosch University(斯泰伦博斯大学);National University of Science and Technology(国家科学与技术大学);\textsc{Politehnica} Bucharest(布加勒斯特理工大学)
Does Generative AI speak Nigerian-Pidgin?: Issues about Representativeness and Bias for Multilingualism in LLMs
Mila - Quebec AI Institute McGill University, Canada(魁北克AI研究所麦吉尔大学,加拿大);LT3, IDLab, Universiteit Gent(LT3,IDLab,根特大学);Bloomberg(彭博);University of Trento(特伦托大学)