社科网首页|客户端|官方微博|报刊投稿|邮箱 中国社会科学网

您现在的位置:今日语言学 → 语言学系 → 学位论文摘要 → 硕士

闻欣怡:语境对朗读语篇产出与加工研究的影响【语言学及应用语言学】

作者:闻欣怡 导师:贾媛  来源:科研处  时间:2019-10-13

摘  要

  随着语言学研究的深入,人们开始把眼光投入比句子更大的语言单位,即语篇。诸多学者将面向语篇的研究与句法、语用、社会语言学等其他方向相结合,丰富发展了语篇相关的理论和研究方法,并弥补了传统语法分析的不足。此外,随着人工智能领域的快速发展,句层面的语言学研究已无法满足应用需求。因此,开展语篇研究具有重要的理论与应用意义。语篇研究相较于传统研究的特点之一是密切结合语境,尽管目前学界已对语境如何影响语篇的产出与加工展开了一系列探讨,但在理论视角与方法论上仍具有一定的局限性。在此背景下,本文试图基于产出实验与感知实验,为语境对汉语朗读语篇产出与加工的影响提供新认识。在语言产出方面,本文考察了语境在小句和实体两个层面的作用:在小句层面重点探讨过渡关系对停顿时长的影响;在实体层面则将信息结构和回指现象共同纳入研究范畴,并进一步区分语言形式与指称的信息结构,系统地探讨语境对实体重音等级、实体后边界层级的影响。在语言加工方面,本文从神经-语言夹带现象入手,用反映不同程度语境作用的语言表征与脑电信号的夹带水平衡量语境对底层信息的促进作用,考察语境信息与底层声学信息在激活的脑区以及整体加工水平上的异同。

  产出实验中,本文从汉语普通话朗读语篇语料库(ASCCD)中选取了14篇朗读语篇(每篇由10名发音人朗读),对其开展以下标注工作:基于向心理论与RefLex信息结构标注体系标注过渡关系、信息结构与回指特征,基于C-ToBI韵律标注体系标注重音等级与边界层级。提取上述信息后,本文以过渡关系、信息结构与回指特征为自变量,小句间停顿时长、重音等级与停顿层级为因变量分别进行统计分析。感知实验中,本文收集了18名被试各自听取6篇选自ASCCD的朗读语篇时的脑电数据,将反映不同程度语境信息的语言表征作为刺激、将脑电信号作为反应,训练多变量时间反应函数,基于每个被试听取剩余5个语篇的模型参数以及当前语篇对应的刺激,预测该被试听取当前语篇的神经反应。在此基础上,本文用原始信号和预测得到的信号之间的相关系数评估预测的准确性,并将其视作评估神经夹带水平的因变量,以建模刺激、脑区为自变量进行统计分析。

  基于上述实证研究,本文得到了以下主要结论。语言产出过程中,小句层面的语境信息可在一定程度上影响小句间停顿时长,过渡关系的连贯程度最高,即语境的作用最强时,停顿时长最短;实体层面的语境信息对边界层级、重音等级等韵律特征均有显著影响,语境的作用越弱时,实体后越容易出现高层级的边界,且更易获得高等级重音;可及信息在词汇层和指称层范畴化的方式不同,因此考察语境在实体层面作用时有必要区分语言形式和指称。语言加工过程中,底层声学信息与涉及语境的信息主要激活的脑区均为右颞叶,而涉及语境的信息激活范围相对更广;涉及语境信息的组合表征加工水平略高于底层声学信号,说明语境对底层声学信息加工有一定的促进作用。通过同时考察语境在小句与实体层面的影响,本研究发现其在不同层面对停顿时长的影响呈一致趋势,强调了多层面互动研究的重要性。此外,由于本研究主要基于认知神经科学的手段开展,在此过程中也利用了自然语言处理研究的成果,跨学科研究的优势在本研究中得到了充分的体现。

  关键词:语境;朗读语篇;神经夹带;重音;停顿

 

Abstract 

  The development of linguistic studies has witnessed an extended research scope from sentence to discourse level. Researchers have combined discourse study to syntax, pragmatics, and other areas in linguistics, which to some extent fills in the gaps of traditional sentence-level studies. Besides, the rapid development of artificial intelligence has made it harder for sentence-level linguistic studies to meet the needs of society. Therefore, it is of both theoretical and applicational significance to conduct linguistic study on discourse level. Discourse study is featured by its close relationship to context, with several studies conducted to investigate the influence of context on discourse-level speech production and processing. Still, previous studies are more or less limited in their methodologies or theoretical scopes.

  Against the background, the present study aims to offer new insights to how context influences speech production and processing on discourse level, through both production and perception experiments. Concerning speech production, the present study investigates the influence of inter-clause transition status on pause durations, and the influence of entities’ information structure and anaphora features on stress and boundary levels. Specifically, information structure is studied on both lexical and referential levels. Concerning speech processing, speech-neural entrainment is investigated to represent the level of speech processing. Neural entrainment to both context-based information and sensory information are compared, in terms of activated regions and overall entrainment levels.

  In the production experiment, 14 reading texts selected from ASCCD (each read by 10 speakers) were annotated: transition status, information structure and anaphora features were annotated based on Centering Theory and RefLex annotation scheme, while stress and boundary levels were annotated based on C-ToBI. After extracting the aforementioned information using Praat, statistical analyses were conducted. In the perception experiment, EEG data of 18 participants were recorded while they were listening to 6 reading texts. Several multivariate Temporal Response Function models were trained with EEG signals as response, and context-based information and sensory information as stimuli respectively. The entrainment level was measured with the Pearson correlation coefficient between the reconstructed signal and the original signal, which was further submitted for statistical analysis.

  The following conclusions are drawn from the experiment results. (1) During speech production, context information on clause level (i.e., transition status) has a significant influence on pause durations: the stronger the contexts’ effects are, the shorter the pause durations are. (2) Context information on entity level (i.e., information structure and anaphora features) has a significant influence on stress and boundary levels: the stronger the contexts’ effects are, the lower the stress and boundary levels are. (3) Accessible information are categorized as different information on lexical and referential levels, so it is important to distinguish linguistic form and reference. (4) During speech processing, both context-based information and sensory information mainly activate right temporal area, although context-based information activates a slightly larger range of brain region. (5) The overall entrainment level of context-based information is slightly higher than that of sensory information, suggesting a marginal facilitating effect of context. By placing the influence of context on both clause and entity level together, the present study reveals a consistent pattern of how context influences speech production on both levels, hence emphasizing the significance of multi-level interface studies. Besides, speech processing on discourse level was investigated with the method of cognitive neuroscience in the present study, as well as natural language processing, thus highlighting the advantages of interdisciplinary studies.

  Key Words: context, reading texts, neural entrainment, stress, pause

友情链接

COPYRIGHT © 2017

中国社会科学院语言研究所版权所有

京ICP备17005063号-1

邮编:100732

地址:北京市东城区建国门内大街5号

电话:010-85195379

Mail:lingcass@yeah.net