Meta Language-Specific Layers in Multilingual Language Models

This repo contains the source codes for our paper

On Negative Interference in Multilingual Models: Findings and A Meta-Learning Treatment

Zirui Wang, Zachary C. Lipton, Yulia Tsvetkov

EMNLP 2020

Introduction

This repo contains code to train multilingual language models (XLM) that (1) contain language-specific layers, and (2) meta-learn these layers through gradient of gradient.

Language-specific layers are served as meta parameters, optimized using an iterative procedure. The goal is to remedy negative transfer in multilingual models through a meta training objective. Please see our paper for details.

Dependencies

• Python 3
• XLM
• NumPy
• PyTorch

Usage

The code is based on the official implementation of XLM. This repo only contains files that we modified from the original codebase. To train a model, please merge code with the source code of XLM, and then follow the standard preprocessing and training instructions there.