Multilingual alignment and transfer

This repository holds the code for the following papers:

• Multilingual Transformer Encoders: a Word-Level Task-Agnostic Evaluation, Félix Gaschi, François Plesse, Parisa Rastin, Yannick Toussaint. (IJCNN 2022)
• Exploring the Relationship between Alignment and Cross-lingual Transfer in Multilingual Transformers, Félix Gaschi, Patricio Cerda, Parisa Rastin, Yannick Toussaint. (Findings of ACL 2023)
• ALIGNFREEZE: Navigating the Impact of Realignment on the Layers of Multilingual Models Across Diverse Languages. Steve Bakos, Félix Gaschi, David Guzmán, Riddhi More, Kelly Chutong Li, En-Shiun Annie Lee. (NAACL 2025, link coming soon)

Architecture of the repository

• download_resources contain scripts to download necessary resources
• multilingual_eval contain the source code
• scripts contains launchables for reproducing experiments
• subscripts contain various scripts for using external dependencies (e.g. Stansford segmenter) and preparing data (sampling dataset, import results from wandb etc...)

How to use the repository

The reusable source code is found in multilingual_eval, while paper-specific scripts that allows to reproduce a specific experiments and figures from a given paper are found in dedicated subdirectories of scripts:

• scripts/2022_ijcnn for IJCNN 2022
• scripts/2023_acl for Findings of ACL 2023
• scripts/2024_naacl for NAACL 2025