[MA/SA] ML-Based Detection of User Intent and Rhetorical Signals in Symptom Checker Chatbots

28.05.2026, Diplomarbeiten, Bachelor- und Masterarbeiten

This thesis investigates machine learning methods for detecting user intent and rhetorical signals in symptom checker chatbot messages. Using earlier chatbot study data and doctor–patient conversations, it compares rule-based, ML-based, and hybrid models for classifying messages as ethos, pathos, logos, or neutral, to support more adaptive and trustworthy healthcare chatbots.

Background:

Conversational agents in healthcare need to respond not only to user symptoms, but also to how users communicate. User messages may contain signals of ethos (credibility/authority), pathos (emotion/distress), logos (reasoning/information structure), or neutral communication. A rule-based rhetoric analysis framework for this purpose already exists and has been applied to earlier chatbot study data. The next step is to investigate whether this framework can be translated into a machine learning / neural network classifier.

Objective:

The goal of this thesis is to develop and evaluate an ML-based approach for automatic classification of user messages into ethos, pathos, logos, and neutral, using a hybrid dataset consisting of:

- user messages from earlier chatbot studies
- doctor–patient / health-related conversation open source dataset
The thesis should compare the new ML-based approach against the existing rule-based baseline and assess whether hybrid or learned models improve robustness and generalization.

Tasks:

- Review relevant work on rhetorical analysis, text classification, and medical dialogue datasets
- Refine the label schema for ethos, pathos, logos, neutral
- Prepare/annotate and harmonize a hybrid training dataset from existing study data and medical dialogue data
- Implement and compare two to three classifiers
- Evaluate performance quantitatively and qualitatively
- Analyze error patterns and implications for adaptive healthcare chatbots

Expected Outcome:

- A reproducible classification pipeline for ethos, pathos, logos, neutral detection
- A comparison of rule-based, ML-based, and possibly hybrid approaches
- An assessment of how rhetoric-aware classification could support adaptive chatbot response generation in healthcare


Please send your applications including your CV and current grade sheet to rutuja.joshi@tum.de

Highlight the relevant projects/courses related to model/classifier training in your application.

Kontakt: rutuja.joshi@tum.de