Luca Benedetto

Luca Benedetto

Postdoc researcher at the University of Cambridge, working on the application of Machine Learning and NLP to education.

Project proposals

Below you can find some project proposals. Feel free to email me to discuss any of the suggestions below, or any ideas that fit with my research interests.

Graph neural networks for knowledge tracing

In education, knowledge tracing is the task of modelling students’ knowledge level and how it evolves over time. Originally performed with psychometrics approaches, such as Item Response Theory, in recent years it witnessed the application of neural networks (e.g. DKT [1]) and Transformers (e.g. [2, 3]). Some research also employed Graph neural networks for the task, modelling students and questions as nodes and students’ answers as edges [4]. The goal of this project is building a knowledge tracing model that leverages a graph-like structure that contains not only information about students, questions, and their interactions, but also a taxonomy of the topics assessed by the questions, the reasoning types required by them, and their textual content.

Datasets:

References:

Question difficulty estimation from text with graph-structured knowledge bases

Question Difficulty Estimation (QDE) consists in estimating a numerical value, either continuous or discrete, that represents the difficulty of an exam question. Traditionally, it is performed with manual calibration or with pretesting, which both have some limitations: crucially manual calibration is subjective, and pretesting is slow and expensive to perform. Recent research has targeted these limitations by proposing the usage of Natural Language Processing (NLP) techniques for QDE, with the idea that question text is the only information that is available at the time question creation and, if we were capable of performing QDE from text, we might overcome the limitations of manual calibration and pretesting (see [1] for a survey and [2-4] to get an idea of the papers targeting this task). The goal of this project is to develop a model capable of performing question difficulty estimation from test in an unsupervised manner, leveraging the textual information of questions and the taxonomy of “skills” that are assessed by them.

Datasets:

References:

Modelling of students’ misconceptions

Given a set of student responses to a set of exam questions, the goal of this project is to develop an unsupervised approach to automatically detect common misconceptions, which are frequent across different students. Once such misconceptions are identified, they could be used to suggest content to target them (e.g. questions for formative assessment).

Datasets:

Difficulty-agnostic adaptive testing

Adaptive testing consists in customising the exam that a student is given, with the goal of providing her with the questions that are the most useful for her (estimated) skill level. It is generally performed by leveraging questions’ difficulty for item selection. The goal of this project is to develop a difficulty-agnostic ML-based approach to minimise the number of questions that are used in adaptive testing, while still reaching an accurate estimation of students’ skill level (according to a given threshold).

Datasets: