This repository contains the code, data, plots, and results accompanying our paper submission to COLING 2025.
- All scripts in the
CODEfolder are executable. You must set the language to either'FR'or'EN'(or'fr','en') at the beginning of each script. - Ensure that the correct local paths are set before executing the code.
- The code allows for reproduction of results and can be adapted to test additional models and/or a different corpus using the provided input format.
Contains all input data used for the analysis:
- Stylometry reports (based on EnzoFleur's repository).
- Embeddings from French and English datasets in .pkl files.
Contains the output data from the analysis:
- Clustering Results: Includes NMI, Purity, and Mean S scores, as detailed in Section 4.1 of the paper.
- Dispersion Results: Contains condition-wise analysis, centroid means, and statistical significance, as discussed in Section 4.2.
- Embedding Style Interpretability: Results for Section 4.3.
Contains the main scripts to reproduce the results:
- Clustering Analysis: Code for Section 4.1.
- Dispersion Analysis: Full experimental setup for Section 4.2, covering all dimensionalities and 30 seed iterations.
- Style Ground Truth Analysis: Code for Section 4.3, including analysis of style and dispersion deltas and MSE additional experiment per class.
- Text Generation: Code for generating additional classes using GPT-4.
Contains all major plots for each section:
- Section 4.1: 2D PCA clustering projections, along with class frequency bar plots.
- Section 4.2: Contour plots for dispersion results (using sentence-transformers/all-MiniLM-L12-v2) in 2D UMAP, including sensitivity plots across different iterations for decision making.
- Section 4.3: Correlation plots comparing dispersion deltas with stylometric ground truth deltas, and analysis of dispersion results across four classes.
For any support or additional information, please contact : [email protected]