Argumentative Writing Support by means of Natural Language Processing

Persuasive essay writing is a powerful pedagogical tool for teaching argumentation skills. So far, the provision of feedback about argumentation has been considered a manual task since automated writing evaluation systems are not yet capable of analyzing written arguments. Computational argumentation, a recent research field in natural language processing, has the potential to bridge this gap and to enable novel argumentative writing support systems that automatically provide feedback about the merits and defects of written arguments. The automatic analysis of natural language arguments is, however, subject to several challenges. First of all, creating annotated corpora is a major impediment for novel tasks in natural language processing. At the beginning of this research, it has been mostly unknown whether humans agree on the identification of argumentation structures and the assessment of arguments in persuasive essays. Second, the automatic identification of argumentation structures involves several interdependent and challenging subtasks. Therefore, considering each task independently is not sufficient for identifying consistent argumentation structures. Third, ordinary arguments are rarely based on logical inference rules and are hardly ever in a standardized form which poses additional challenges to human annotators and computational methods. To approach these challenges, we start by investigating existing argumentation theories and compare their suitability for argumentative writing support. We derive an annotation scheme that models arguments as tree structures. For the first time, we investigate whether human annotators agree on the identification of argumentation structures in persuasive essays. We show that human annotators can reliably apply our annotation scheme to persuasive essays with substantial agreement. As a result of this annotation study, we introduce a unique corpus annotated with fine-grained argumentation structures at the discourse-level. Moreover, we present a novel end-to-end approach for parsing argumentation structures. We identify the boundaries of argument components using sequence labeling at the token level and propose a novel joint model that globally optimizes argument component types and argumentative relations for identifying consistent argumentation structures. We show that our model considerably improves the performance of local base classifiers and significantly outperforms challenging heuristic baselines. In addition, we introduce two approaches for assessing the quality of natural language arguments. First, we introduce an approach for identifying myside biases which is a well-known tendency to ignore opposing arguments when formulating arguments. Our experimental results show that myside biases can be recognized with promising accuracy using a combination of lexical features, syntactic features and features based on adversative transitional phrases. Second, we investigate for the first time the characteristics of insufficiently supported arguments. We show that insufficiently supported arguments frequently exhibit specific lexical indicators. Moreover, our experimental results indicate that convolutional neural networks significantly outperform several challenging baselines.