Anomaly-based large-scale analysis of style and genre reflected in the use of stylistic devices in medieval literature

这个项目的目标是开发一种新的远程阅读工具，用于定量分析修辞手段的使用。所开发的方法将被应用于对三种中世纪体裁的比较文体分析：三种中世纪体裁：中高德语改编的三种风格。从中世纪研究的角度来看，我们的目标是确定古代传奇、亚瑟王传奇和12世纪和13世纪改编的《诗经》之间的体裁差异。到目前为止，这些体裁在修辞手段使用方面的量化异同还没有得到分析。这样的分析对于讨论中世纪的“体裁”概念和中世纪作家的体裁意识，以及评估中世纪学校教授的修辞手法与白话作家实践的相关性都是有价值的。为此，我们打算分析修辞手段的选择、使用和频率是否取决于场景类型，以及特征差异是否可以归因于体裁，或者仅仅是不同的作者或起源日期。这种宽泛的分析需要用于远距离阅读的计算方法，而目前还不存在对文体手段进行令人满意的检测的计算方法，更不用说对于中高德语文本了。为此，我们试图基于异常检测和主动终身机器学习领域的方法来开发新的修辞检测方法。在这种情况下，我们认为修辞手段是“异常”，即对语料库大部分建立的数量规范的偏离，这可以用统计方法来检测(例如，交叉是非交叉结构中的异常)。为了将耗时的手动标注任务减少到最低限度，我们建议在半监督场景中使用主动学习技术，其中人在循环中：基于最初无监督的检测文体异常的过程，系统通过主动向用户索要几个信息性示例的标注来在交互过程中改进自己。此外，我们打算将专家提供的先验理论知识整合到异常检测过程中，以过滤出明显的假阳性，并将其引导到某些修辞手段上。在这个项目中，我们将重点介绍两种典型的文体手法：交叉作为重复的形象和隐喻作为比喻。后者也可以被视为一种反常现象，因为来自不同领域的词之间存在不寻常的相互作用，这反映在它们的词嵌入上。由于这样的嵌入需要从大型语料库中学习，而这对于资源较低的语言(如中高德语)是不可用的，因此我们还努力开发一种技术，将在新高德语语料库上学到的此类嵌入应用到中高德语中。

The goal of this project is the development of a novel distant-reading tool for the quantitative analysis of the use of rhetorical devices. The developed methods will be applied for a comparative stylometric analysis of three medieval genres: the Middle High German adaptations of the trois matières. From the viewpoint of medieval studies, we aim for identifying genre-specific stylistic differences between romances of antiquity, Arthurian romances, and chanson de geste adaptations of the 12th and 13th century. Up to now, quantifiable similarities and differences of these genres regarding their use of rhetorical devices have not been analysed. Such an analysis would be valuable for the controversial discussion about the medieval concepts of ‘genre’ and the genre-awareness of medieval authors as well as for assessing the relevance of rhetorical devices taught in medieval schools for the practice of vernacular authors. To this end, we intend to analyse whether selection, use, and frequency of rhetorical devices depend on the scene type and whether characteristic differences can be attributed to genres or just different authors or dates of origin. Such a broad analysis requires computational methods for distant reading, which do not yet exist for the detection of stylistic devices to a satisfactory degree, not to mention for Middle High German texts. For this purpose, we seek to develop novel methods for detecting rhetorical devices based on methods from the fields of anomaly detection as well as active and life-long machine learning. In this context, we consider rhetorical devices as ‘anomalies’, as deviations from the quantitative norm established by the greater part of the corpus, which can be detected using statistical methods (e.g., a chiasmus is an anomaly amidst non-chiastic constructions). To reduce time-consuming manual annotation tasks to a minimum, we propose to employ active learning techniques in a semi-supervised scenario with the human in the loop: Based on an initially unsupervised process of detecting stylistic anomalies, the system refines itself in an interactive process by actively asking the user for annotations for a few informative examples. Furthermore, we intend to integrate prior theoretical knowledge provided by experts into the anomaly detection procedure for filtering out obvious false positives and steering it towards certain rhetorical devices. In this project, we will focus on two exemplary stylistic devices: the chiasmus as a figure of repetition and the metaphor as a trope. The latter can be seen as an anomaly as well due to the unusual interactions between words from different domains, reflected in their word embeddings. Since such embeddings need to be learned from huge corpora, which are not available for low-resource languages such as Middle High German, we also strive for developing a technique to adapt such embeddings learned on New High German corpora to Middle High German.