Modeling Persuasive Discourse to Adaptively Support Students' Argumentative Writing

Thiemo Wambsganss, Christina NiklausACL 2022 · DOI

Introduction

Persuasive writing is a cornerstone of professional success, particularly in entrepreneurship where the quality of a "pitch" can determine the fate of a business idea. However, students often struggle to master these skills because high-quality human feedback is slow, expensive, and difficult to scale. To bridge this gap, researchers have developed a novel argumentation mining (AM) framework specifically designed for student-written business model pitches in German. By creating a new corpus annotated for both discourse structure and qualitative persuasiveness, they built an adaptive writing support tool that provides instant, automated feedback.

The Challenge of Domain-Specific Argumentation

While Argumentation Mining has made strides in identifying claims and premises, most existing datasets focus on English essays or legal documents. There is a distinct lack of resources for the German language and the business domain. Furthermore, most AM tasks are purely structural; they identify parts of an argument but do not assess how "strong" or "relevant" those parts actually are.

Methodology: A Multi-Dimensional Scoring Framework

The authors adapted the classic Toulmin model of argumentation to the business pitch context. Text is decomposed into a hierarchy: Major Claims (the core business idea), Claims (value propositions), and Premises (supporting evidence or reasons).

Hierarchical discourse structure of a business model pitch, illustrating the support relations between major claims, claims, and premises.

Unlike standard AM tasks, this framework introduces a 1-to-5 multidimensional scoring system:

  • Major Claims are scored for Specificity and Evidence.
  • Premises are scored for Strength and Relevance.

The "Rebike" Corpus and Modeling

The study introduces a new corpus of 200 student-written business model pitches in German, comprising over 3,200 annotated sentences. This dataset is unique in its inclusion of qualitative labels alongside structural ones.

Annotated example of a 'Rebike' pitch, showing how specific sentences are classified and scored for persuasiveness.

To automate the feedback process, the researchers trained Long Short-Term Memory (LSTM) and BERT models. Interestingly, the LSTM model outperformed the transformer-based BERT (54.12% vs 47.50% accuracy), highlighting the challenges of applying massive pre-trained models to niche, low-resource domains like German business pitches.

Key Insights and Results

The research revealed that business pitches follow a unique hierarchy where authors focus almost exclusively on "support" relations to build their case, with "attack" relations being virtually non-existent.

In a field study with 25 students, the adaptive support tool—which highlights text and provides real-time scores—received positive feedback. Students reported a perceived ease-of-use score of 4.94/7 and a self-efficacy score of 4.98/7. This indicates that automated feedback, even with moderate accuracy, significantly boosts student confidence and their perceived ability to craft persuasive arguments.

Conclusion and Future Work

While the current system provides a strong foundation, the authors acknowledge that a ~54% accuracy rate leaves room for improvement. Future work will focus on developing more sophisticated NLP architectures and expanding the dataset. A key goal is to move beyond simple scoring to provide actionable feedback—concrete suggestions on how a student can rewrite a weak sentence to make it more persuasive.

Summary updated

BibTeX
@inproceedings{wambsganss2022modeling,
  author = {Thiemo Wambsganss and Christina Niklaus},
  title = {{Modeling Persuasive Discourse to Adaptively Support Students' Argumentative Writing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.599},
}
Figure