This repository contains the code for experiments from the ACL 2025 Findings paper "In the LLM era, Word Sense Induction remains unsolved" by Anna Mosolova, Marie Candito and Carlos Ramisch.

The dataset proposed in the paper is available at SemCor-WSI website.

To install the required dependencies, run:

pip install -r requirements.txt

To run the clustering algorithm, execute perform_clustering.py with its arguments:

python perform_clustering.py --output_file out

python perform_clustering.py --dataset datasets/semcorwsi_dataset_all_dev.csv --algorithm agglomerative --n_clusters_compute silhouette --constraint must-link --output_file out

• dataset : Path to a pandas dataframe containing the data for word sense induction
• algorithm : Clustering algorithm to use (agglomerative or x-means)
• n_clusters_compute : Method for determining the number of clusters (silhouette or const+silhouette if must-link is used)
• constraint : Whether to apply must-link or must-link/cannot-link constraints
• modelname : Name of the pre-trained model to use
• layer : The PLM's layer to extract embeddings from
• device : Which device to use
• add_data_path : Path to additional unlabeled data for data augmentation
• output_file : Path to save the output results

The script generates an out.pkl file containing a dictionary with metrics and labels for each example in the dataset.

The datasets folder contains:

• SemCorWSI dataset,
• Wiktionary dataset filtered on dev part of SemCorWSI,
• WikiBooks datasets with 10, 50, 100, 150 examples for dev part of SemCorWSI,
• GPT-4o dataset of generated examples for dev part of SemCorWSI,
• Llama 3.1 8B 4bit generated examples dataset for dev part of SemCorWSI.

The src folder contains the main codebase, including the WSIClustering class, which provides core functionality for clustering used in the paper.