WritingTutor: Improving a Conversational Learning Tool

Jan 6, 2023

• Python
• React
• TextBlob
• Hugging Face Transformers
• Google Translate API

Cover image generated with OpenAI's GPT-5 model.

• 26 students in field study
• 92.4% found interface intuitive
• 88.5% agreed with automated evaluation
• Grade: 5.5 / 6

TL;DR

• Extended WritingTutor chatbot to teach persuasive writing with interactive feedback.
• Rebuilt frontend in React for maintainability and flexibility.
• Added text analysis models: polarity, subjectivity, emotions, and summaries.
• Field study (n=26): >90% found interface intuitive; >80% found highlighted feedback helpful.

At a glance

• Role: Solo author (Bachelor’s thesis)
• Timeline: Sep 2022 – Jan 2023 (4 months)
• Context: EPFL B.Sc. Computer Science
• Users/Stakeholders: Students improving essay writing; educators scaling feedback
• Scope: Full-stack dev (frontend redesign + NLP integration) → Evaluation study → Report

Problem

Students often lack individualized feedback on writing due to large class sizes and MOOCs. Manual grading is slow and inconsistent, yet feedback is critical for learning. The challenge: Can a chatbot provide scalable, interactive writing support that is still helpful, intuitive, and trustworthy?

Solution overview

I improved the WritingTutor chatbot, focusing on both UX and feedback depth:

1. Frontend redesign in React for flexibility and feature extensibility.
2. Switchable interfaces: chat (theory), writing, and evaluation.
3. Interactive feedback: sentence highlighting for subjectivity/polarity, keyword emphasis, and summaries.
4. New analyses: emotion detection and automatic summarization.
5. Evaluation study: 26 students tested the tool, providing structured feedback.

Architecture

• Frontend: React-based, modular interfaces (chat, writing, evaluation).
• Backend: Chatomatic bot + NLP models (TextBlob/TextBlobDE, Hugging Face summarization/emotion).
• Flow: Chat → Essay writing → Evaluation & feedback (highlighted text, keywords, emotions).

Data

• Essays: ~250–300 words written by 26 participants.
• Language: German essays, auto-translated to English for analysis (due to limited German NLP models at the time).
• Analysis models:
  • TextBlob: subjectivity/polarity (English).
  • TextBlobDE: polarity (German).
  • Hugging Face: summarization (T5-Base GNAD) & emotion classification (DistilRoBERTa).
• Survey data: Usability & feedback ratings from participants.

Method

• Iterative development: Scrum-like cycles (design → implement → test → evaluate).
• Feature prioritization: based on earlier survey results (need for switching interfaces, richer feedback).
• Evaluation:
  • Continuous testing after each increment.
  • Final field study with 26 participants (essay writing task + survey).

Experiments & Results

• User experience:
  • 92.4% found the UI intuitive and enjoyed using WritingTutor.
• Feedback quality:
  • 88.5% agreed with the automated evaluations.
  • >80% found highlighted sentences helpful to understand results.
• Learning experience: Participants valued control over pace and interactive theory inputs.

Product & UX

• Primary journey: learn theory via chatbot → write essay → receive interactive feedback.
• Key features:
  • Switch between chat and writing without losing progress.
  • Highlights decisive sentences for polarity/subjectivity.
  • Emotion analysis + auto-summaries.
  • Keywords highlighted for synonym suggestions.

Impact

• Showed chatbots can scale feedback in writing education.
• Improved engagement and trust with interactive, explainable feedback.
• Provided design principles for educational conversational agents (control, transparency, interactivity).

What I learned

• How to design usable NLP-powered educational tools.
• Importance of explainability (highlighting decisive sentences increased trust).
• Trade-offs between language availability (German vs English models) and translation quality.
• Running a user study: balancing usability, learning outcomes, and data collection.

Future Work

• Multilanguage support (native models, beyond German).
• Support for different text types (reports, news, emails).
• More detailed feedback (e.g., argumentation mining, rewrite suggestions).
• Mobile UX improvements for small screens.

References

• Wambsganss & Käser — Conversational Tutoring Systems (2023).
• Hartmann — Emotion DistilRoBERTa (2022).
• Hugging Face T5-Base GNAD summarizer.
• TextBlob & TextBlobDE documentation.