融合词簇约束的汉越跨语言词嵌入

针对传统跨语言词嵌入方法在汉越等差异较大的低资源语言上对齐效果不佳的问题,提出一种融合词簇对齐约束的汉越跨语言词嵌入方法。通过独立的单语语料训练获取汉越单语词嵌入,使用近义词、同类词和同主题词3种不同类型的关联关系,充分挖掘双语词典中的词簇对齐信息以融入到映射矩阵的训练过程中,使映射矩阵进一步学习到不同语言相近词间具有的一些共性特征及映射关系,根据跨语言映射将两种语言的单语词嵌入映射至同一共享空间中对齐,令具有相同含义的汉语与越南语词嵌入在空间中彼此接近,并利用余弦相似度为空间中每一个未经标注的汉语单词查找对应的越南语翻译构建汉越对齐词对,实现跨语言词嵌入。实验结果表明,与传统有监督及无监督的跨语言词嵌入方法Multi_w2v、Orthogonal、VecMap、Muse相比,该方法能有效提升映射矩阵在非标注词上的泛化性,改善汉越低资源场景下模型对齐效果较差的问题,其在汉越双语词典归纳任务P@1和P@5上的对齐准确率相比最好基线模型提升了2.2个百分点。     

基于预训练与音素字节对编码的越南语识别

基于无监督预训练技术的wav2vec 2.0在许多低资源语种上获得了良好的性能,成为研究的热点。本文在预训练模型的基础上进行越南语连续语音识别。将语音学信息引入到基于链接时序分类代价函数（Connectionist temporal classification,CTC）的声学建模中,选取音素与含位置信息的音素作为基础单元。为了平衡建模单元数目以及模型的精细程度,采用字节对编码（Byte-pair encoding,BPE）算法生成音素子词,将上下文信息结合到声学建模过程。实验在美国NIST的BABEL任务低资源的越南语开发集上进行,所提算法相对wav2vec 2.0基线系统有明显改进,识别词错误率由37.3%降低到29.4%。     

Data augmentation for low-resource languages NMT guided by constrained sampling

Data augmentation (DA) is a ubiquitous approach for several text generation tasks. Intuitively, in the machine translation paradigm, especially in low-resource languages scenario, many DA methods have appeared. The most commonly used methods are building pseudocorpus by randomly sampling, omitting, or replacing some words in the text. However, previous approaches hardly guarantee the quality of augmented data. In this study, we try to augment the corpus by introducing a constrained sampling method. Additionally, we also build the evaluation framework to select higher quality data after augmentation. Namely, we use the discriminator submodel to mitigate syntactic and semantic errors to some extent. Experimental results show that our augmentation method consistently outperforms all the previous state-of-the-art methods on both small and large-scale corpora in eight language pairs from four corpora by 2.38–4.18 bilingual evaluation understudy points.     

Maintenance of quality of life improvements in diverse rural older adults.

The maintenance of effects from home-delivered cognitive-behavioral therapy (CBT) was examined. One hundred thirty-four participants, predominantly African American and primarily rural, low-resource, and physically frail, were randomly assigned to either immediate or delayed CBT. The six-month follow-up assessments indicated that among those who remained in the study, participants evidenced significantly improved quality of life and reductions in psychological symptoms at follow-up, relative to pretreatment levels. Posttreatment gains were maintained at follow-up. These data suggest that treatment effects can be achieved and perhaps maintained with a disadvantaged sample of older adults and suggest that evidence-based treatments delivered through nontraditional means can have effects beyond posttreatment. (PsycINFO Database Record (c) 2011 APA, all rights reserved)     

GODZILA: A Low-resource Algorithm for Path Planning in Unknown Environments

We propose a novel sensor-based path-planning and obstacle avoidance algorithm GODZILA for navigation in unknown environments. No prior knowledge of the environment is required. The path-planning algorithm follows a purely local approach using only the current range sensor measurements at each sampling instant and requiring only a small number of stored variables in memory. No map of the environment is built during navigation. This minimizes the memory and computational requirements for implementation of the algorithm, a feature that is especially attractive for small autonomous vehicles. The algorithm utilizes three components: an optimization algorithm, a local straight-line path planner to visible targets, and random navigation. It is proved, for navigation in any finite-dimensional space, that the path-planning algorithm converges in finite time with probability 1. The performance of the algorithm is demonstrated through simulations for path-planning in two-dimensional (2D) and three-dimensional (3D) spaces. It is seen that a relatively small number of range sensor measurements is sufficient even in complex unknown environments. This work is supported in part by the ARO under contract #W911NF-04-C-002, by the ONR under contract #N00014-06-C-0051, and by IntelliTech Microsystems, Inc. An earlier version of this paper was presented at the 2005 American Control Conference, Portland, OR.     

面向神经机器翻译的枢轴方法研究综述

低资源语言的神经机器翻译（neural machine translation,NMT）一直是机器翻译领域研究的难点和热点,基于枢轴的方法为其性能的提升和改进提供了思路。针对枢轴思想在低资源语言神经机器翻译中的应用,从枢轴翻译、基于枢轴的伪平行数据生成和基于枢轴的模型构建三方面,对不同方法的国内外研究现状、主要问题和趋势进行了分析和比较,为该领域的研究提供参考和借鉴。     

低资源语音识别中融合多流特征的卷积神经网络声学建模方法

针对卷积神经网络（CNN）声学建模参数在低资源训练数据条件下的语音识别任务中存在训练不充分的问题,提出一种利用多流特征提升低资源卷积神经网络声学模型性能的方法。首先,为了在低资源声学建模过程中充分利用有限训练数据中更多数量的声学特征,先对训练数据提取几类不同的特征;其次,对每一类类特征分别构建卷积子网络,形成一个并行结构,使得多特征数据在概率分布上得以规整;然后通过在并行卷积子网络之上加入全连接层进行融合,从而得到一种新的卷积神经网络声学模型;最后,基于该声学模型搭建低资源语音识别系统。实验结果表明,并行卷积层子网络可以将不同特征空间规整得更为相似,且该方法相对传统多特征拼接方法和单特征CNN建模方法分别提升了3.27%和2.08%的识别率;当引入多语言训练时,该方法依然适用,且识别率分别相对提升了5.73%和4.57%。     

融合句法解析树的汉-越卷积神经机器翻译

神经机器翻译是目前应用最广泛的机器翻译方法,在语料资源丰富的语种上取得了良好的效果.但是在汉语-越南语这类缺乏双语数据的语种上表现不佳.考虑汉语和越南语在语法结构上的差异性,提出一种融合源语言句法解析树的汉越神经机器翻译方法,利用深度优先遍历得到源语言的句法解析树的向量化表示,将句法向量与源语言词嵌入相加作为输入,训练翻译模型.在汉-越语言对上进行了实验,相较于基准系统,获得了0.6个BLUE值的提高.实验结果表明,融合句法解析树可以有效提高在资源稀缺情况下机器翻译模型的性能.     

基于数据增强的藏文改写检测研究

该文针对藏文语料稀缺的问题,在藏汉双语、藏文单语文本改写检测任务中使用数据增强的方法,在一定程度上解决了低资源语言训练语料规模小的问题。在藏汉跨语言文本改写检测任务中,该文使用数据增强方法,有效利用目前公开的藏汉平行语料,扩充藏汉跨语言文本改写检测训练语料,当扩充至20万句对时,藏汉改写检测模型的皮尔森系数(pearson correlation)达到0.547 6,比基线系统的皮尔森系数提升了0.397 1,表明藏汉改写检测模型检测出的句对相似度值与人工标注的相似度值已达到中等程度相关。在藏文单语言任务中,该文采用训练藏文音节向量的方法,以缓解语料稀缺带来的词向量稀疏问题。实验结果表明,基于藏文音节向量的藏文改写检测模型的皮尔森系数可达到0.678 0,比相应的基于藏文词向量实验的结果提升了0.1,使得藏文单语言文本改写检测模型的检测结果与人工标注的结果达到了强相关程度。     

探索低资源的迭代式复述生成增强方法

复述生成旨在同一语言内将给定句子转换成语义一致表达不同的句子。目前,基于深度神经网络的复述生成模型的成功依赖于大规模的复述平行语料,当面向新的语言或新的领域时,模型性能急剧下降。面对这一困境,提出低资源的迭代式复述生成增强方法,最大化利用单语语料和小规模复述平行语料迭代式训练复述生成模型并生成复述伪数据,以此增强模型性能。此外,提出了句子流畅性、语义相近性和表达多样性为基准设计的伪数据筛选算法,选取高质量的复述伪数据参与每轮模型的迭代训练。在公开数据集Quora上的实验结果表明,提出的方法仅利用30%的复述语料在语义和多样性指标上均超过了基线模型,验证了所提方法的有效性。