QuanTOR - Quantitative Analysis of Textual Organisation across Registers

在语言学研究中，语域通常是在整个语篇的层面上进行分析的，尽管语域的一般定义都与情景语境有关。情景随着时间的推移而动态地展开，语言使用者在这个过程中的不同点会做出不同的选择。因此，语篇在开头、中间和结尾都是不同的，这种在次语篇层面上的动态组织需要整合到语域研究中。该项目旨在丰富语域研究，说明情况的动态性质，因此，语域以及研究这一现象的定量方法。我们专注于文本组织的时间动态需要一种能够从短文本片段中提取语言特征模式和潜在（潜在）变化维度的方法。它还需要自动识别和分类相关部分，以便扩展到非常大的语料库的分析。为了实现这些目标，该项目采用了三管齐下的语料库分析方法，利用其三个申请人各自的专业知识。我们的工作计划将语言解释，理论开发和手动注释与多变量定量分析以及无监督和监督机器学习相结合。为此，我们开发了一种新的贝叶斯版本的几何多变量分析，一种可靠的细粒度方法来调查语言变异，以及机器学习方法，用于分割和标记文本，应用最先进的神经和统计语言模型。在一个迭代的过程中，我们开发了一个理论的动态语言使用的情景背景下，手动分割和标记文本的黄金标准，贝叶斯GMA方法研究文本时间的多变量特征分布，以及易于应用的语言模型自动文本分割和标签。所有成分和定量结果都经过仔细评估和验证。该项目使用国际英语语料库（ICE）的组成部分，这不仅有利于跨一系列不同的寄存器的时间动态分析-在口语和书面语模式-但也使我们能够提交理论主张的实证检验，例如，关于体裁和寄存器之间的理论关系。计算支持的语料库分析的结果将提供一个新的视角，作为一个动态的语言学反映人类行为的情景语境中，基于定量的经验洞察力，这反过来将有助于我们理解的架构语言的理论寄存器。

In linguistic research, registers are usually analysed at the level of entire texts, despite the fact that common definitions of register are linked to the situational context. Situations unfold dynamically over time and language users make different choices at different points in this process. As a consequence, texts are linguistically different at the beginning, in the middle and at the end, and this dynamic organisation at the sub-textual level needs to be integrated in register studies. This project aims at enriching linguistic register studies with an account of the dynamic nature of situations and hence registers as well as quantitative methods for studying this phenomenon. Our focus on the temporal dynamics of text organisation requires an approach that is capable of extracting patterns of linguistic features and the underlying (latent) dimensions of variation from short text segments. It also necessitates automatic identification and classification of relevant segments in order to scale to the analysis of very large corpora. In order to achieve these goals, the project adopts a three-pronged approach to corpus analysis, leveraging the respective expertise of its three applicants. Our work programme combines linguistic interpretation, theory development and manual annotation with multivariate quantitative analysis as well as unsupervised and supervised machine learning. To this end, we develop a novel Bayesian version of Geometric Multivariate Analysis, a reliable and fine-grained approach to the investigation of linguistic variation, as well as machine-learning approaches for the segmentation and labelling of texts that apply state-of-the-art neural and statistical language models. In an iterative process, we develop a theory of the dynamics of language use in situational context, a gold standard of manually segmented and labelled texts, the BayesGMA approach for studying multivariate feature distributions in text time, as well as readily applicable language models for automatic text segmentation and labelling. All components and quantitative results are carefully evaluated and validated. The project uses components of the International Corpus of English (ICE), which not only facilitate the analysis of temporal dynamics across a range of different registers – in the spoken and written mode – but also allow us to submit theoretical claims to an empirical test, for instance, concerning the theoretical relationship between genre and register. The results of the computationally supported corpus analysis will provide a new perspective on a theory of register as a dynamic linguistic reflection of human behaviour in situational context, based on quantitative empirical insight, which, in turn, will feed into our understanding of the architecture of language.