In this paper we provide a quantitative framework for the study of phonological networks (PNs) for the English language by carrying out principled comparisons to null models, either based on site percolation, randomization techniques, or network growth models. In contrast to previous work, we mainly focus on null models that reproduce lower order characteristics of the empirical data. We find that artificial networks matching connectivity properties of the English PN are exceedingly rare: this leads to the hypothesis that the word repertoire might have been assembled over time by preferentially introducing new words which are small modifications of old words. Our null models are able to explain the "power-law-like" part of the degree distributions and generally retrieve qualitative features of the PN such as high clustering, high assortativity coefficient, and small-world characteristics. However, the detailed comparison to expectations from null models also points out significant differences, suggesting the presence of additional constraints in word assembly. Key constraints we identify are the avoidance of large degrees, the avoidance of triadic closure, and the avoidance of large non-percolating clusters.
Emergent morpho-phonological representations in self-supervised speech models
Department of Neurological Surgery, University of California, San Francisco(加州大学旧金山分校神经外科系);Department of Linguistics, University of Southern California(南加州大学语言学系);Center for Computational Language Sciences, University of Southern California(南加州大学计算语言科学中心)
Probing for Phonology in Self-Supervised Speech Representations: A Case Study on Accent Perception
Department of Linguistics, University of Florida(语言学系,佛罗里达大学);Department of English Language and Linguistics, Heinrich Heine University Düsseldorf(英语语言与语言学系,海因里希-海涅大学杜塞尔多夫)
Layer-wise Probing of wav2vec 2.0 and Whisper for Consonant Cluster Reduction in African American English
Department of English Language and Linguistics, Institute of English and American Studies, Faculty of Arts and Humanities, Heinrich Heine University Düsseldorf, Germany(海因里希·海涅大学杜塞尔多夫分校,人文与艺术学院,英美研究学院,英语语言与语言学系);Department of Linguistics, University of Florida, United States of America(佛罗里达大学语言学系)
Quantifying Speaker Embedding Phonological Rule Interactions in Accented Speech Synthesis
Signal Analysis and Interpretation Lab, University of Southern California(南加州大学信号分析与解释实验室);Thomas Lord Department of Computer Science, University of Southern California(南加州大学托马斯·劳德计算机科学系);Department of Linguistics, University of Southern California(南加州大学语言学系)
本研究通过量化说话者嵌入与语音规则的相互作用,提出 PSR 度量标准,揭示口音生成中规则与身份的纠缠关系。
Multilingual LLMs Struggle to Link Orthography and Semantics in Bilingual Word Processing
Department of Electrical Engineering, Indian Institute of Technology Delhi(印度理工学院德里电气工程系);Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi(印度理工学院德里人工智能学院)
Neural Named Entity Recognition from Subword Units
Max Planck Institute for Informatics(马克斯·普朗克信息学研究所);Amazon Alexa(亚马逊Alexa)
命名实体识别(NER)是口语语言理解中的一项重要任务,其目的是识别文本中命名实体的提及,从转录的演讲中。现有的NER神经模型主要依赖于专用的单词级表示,这有两个主要缺点。首先,词汇量很大,产生了大量的记忆需求和训练时间。其次,这些模型不能学习形态或语音表示。为了弥补上述缺点,我们采用了基于双向LSTM和条件随机场的神经解决方案,其中我们依赖于子字单元,即字符,音素和字节。对于话语中的每个单词,我们的模型从每个子词单元中学习一个表示。我们在真实世界的大规模设置中进行了实验,用于语音控制设备的用例,覆盖四种语言,每种语言多达550万个话语。我们的实验表明:(1)随着训练数据的增加,仅在子字单元上训练的模型的性能变得更接近于具有专用字级嵌入的模型(英语为91.35 vs 93.92 F1),同时使用更小的词汇量(332 vs 74K),(2)子字单元增强了具有专用字级嵌入的模型,以及(3)组合不同的子字单元提高了性能。
Massachusetts Institute of Technology(麻省理工学院);Universitat Pompeu Fabra(庞培法布拉大学);Institut de Biologia Evolutiva (CSIC-UPF)(进化生物学研究所(西班牙国家研究委员会-庞培法布拉大学));Santa Fe Institute(圣塔菲研究所)