Motivation

Dialogue is central to education, as learning emerges through interaction, discussion, and the co-construction of knowledge. One way to understand educational dialogue is through dialogic acts, which describe the communicative functions and meanings of utterances. In educational settings, dialogic acts provide a way to represent the pedagogical intent underlying interactions, offering valuable insights into teaching and learning processes.

However, manually annotating educational dialogues with dialogic acts is expensive and time-consuming. Researchers have therefore explored automating this process, particularly through large language models (LLMs), whose natural language understanding and in-context classification capabilities offer the potential to adapt to different dialogic act taxonomies without task-specific training. Yet, recent research has shown that even advanced LLMs struggle to reliably infer and predict dialogic acts.

In this project, we propose exploring a different approach to this challenge. LLMs trained for tool calling learn to select appropriate tools from a set of available options based on their descriptions, a capability that closely resembles in-context classification. By representing dialogic acts as tools and framing their inference as a tool-calling task, we aim to investigate whether LLMs trained for tool calling can transfer their tool-selection capabilities to the inference and prediction of pedagogical intent.

Your Task

In this thesis, you will first familiarize yourself with the theory of dialogic acts and their role in educational interactions. You will then extend an existing benchmark for educational dialogue understanding to support dialogic act prediction through tool calling.

Using this benchmark, you will systematically evaluate and analyze whether LLMs trained for tool calling can leverage their tool-selection capabilities to infer and predict dialogic acts, comparing their performance with conventional approaches. Building on these findings, you will explore fine-tuning strategies to further improve their performance and investigate how well these capabilities generalize across different educational contexts and dialogic act taxonomies.

Your Profile

• Strong programming skills with Python.
• Strong interest in large language models, NLP, and agentic AI.
• Experience with LLM inference, tool calling, or fine-tuning is a plus.
• Interest in educational AI and dialogue understanding is a plus.