- This repository contains the code that was used to evaluate which categorization approach performed best at categorizing student implementations.
- To assess categorization quality, an ideal categorization of the "shirt-size"-task was crafted. It is shown in the "sample-data/shirt-size/optimal-categories" folder.
- In the "approaches" folder, all the tested approaches are listed. They are seperated by approach type, currently "jaccard", "llm", and "tsed".
- For each approach, there is one file that performs the offline clustering experiments and one that performs the online clustering experiments.
- Within those files, the relevant environment variables are set (e.g. data of which task is used. Since evaluation is only possible for the "shirt-size" task (because only this task has a ground-truth categorization), it is currently set everywhere).
- Running these files performs the clustering and outputs the result (as well as the evaluation if the "shirt-size" task data was used).
- The results are persisted in a file and stored in the "results" folder of the approach.
- Anonymized data of three old ACCESS tasks are contained in this repository ("arithmetic-expression", "invert-dictionary", and "shirt-size"). They can be found in the "sample-data" folder and can be used in the experiments.
- For each task, there's one .json file containing all the submissions that students made, and one .json file that contains only the first submissions students made.