AAdaM at SemEval-2024 Task 1: Augmentation and Adaptation for Multilingual Semantic Textual Relatedness

This repository contains code for our system paper: AAdaM at SemEval-2024 Task 1: Augmentation and Adaptation for Multilingual Semantic Textual Relatedness. If you find this repository useful, please consider citing our paper.

Contact: Miaoran Zhang (mzhang@lsv.uni-saarland.de)

Setup

Create your docker image or virtual environment based on the provided docker file (Dockerfile). Specifically,

python: 3.8
torch: 1.14.0
CUDA: 11.8.0

Code

Training

• train_mlm.py: MLM with full fine-tuning.
• train_ft.py: Supervised learning with full fine-tuning (supports warmup).
• train_lang_adapter.py: Train language adapters (supports continued pre-training).
• train_task_adapter.py: Train task adapters.
• train_warmup_adapter: Train task adapters on noisy data first, then on provided task data.

Evaluation

• eval_ft.py: Evaluate full fine-tuned models.
• eval_adapters.py: Evaluate adapters.
• eval_fasttext.py: Evaluate FastText embeddings.
• eval_overlap.py: Evaluate lexical overlap.
• eval_sentemb.py: Evaluate out-of-the-box embeddings extracted from pre-trained models.

Others

• prepare_data.py: Prepare data for various scenarios, e.g.,
  • Split task data to n-folds for cross-validation.
  • Extract unlabeled sentences for task-adaptive pre-training.
  • Prepare downloaded Leizig data for training language adapters.
• utils.py: Utilities functions, e.g., data processing, computing language similarites.

Please check scripts for experimentation details.