Hyper-X: A Unified Hypernetwork for Multi-Task Multilingual Transfer

Hyper-X is a single hypernetwork that unifies multi-task and multilingual learning with efficient adaptation. Hyper-X generates weights for adapter modules conditioned on both tasks and language embeddings to to fully leverage training data when it is available in different task-language combinations.

Getting Started

This repository is tested with Python 3.7. To install the the required python packages use requirements.txt

pip install -r requirements.txt

Data

POS (UD) and NER (WikiANN) data is used via huggingface/datasets library. To use masked language modelling as an auxiliary task, provide raw corpus into data/mlm with this format <lang_code>.txt.

Train Model

For training, provide following arguments:

--tasks: List of tasks to learn via the hypernetwork, e.g. 'ner' 'pos'

--languages: List of languages to learn via the hypernetwork, e.g. 'en' 'ar' 'tr'

--train_task_language_pairs: Task-language pairs for training, e.g. 'ner#en' 'pos#en' 'ner#ar' 'pos#tr'

--eval_task_language_pairs: Eval (during training) task-language pairs, e.g. 'ner#en' 'pos#en' 'ner#ar' 'ner#tr' 'pos#ar' 'pos#tr'

For other arguments see arguments.py

python3 train.py --model_name_or_path 'bert-base-multilingual-cased' \
                --max_train_steps 100000 \
                --preprocessing_num_workers 1 \
                --per_device_train_batch_size 32 \
                --per_device_eval_batch_size 32 \
                --output_dir $OUTPUT_DIR \
                --save_steps 5000 \
                --eval_steps 5000 \
                --tasks $TASKS \
                --languages $LANGS \
                --train_task_language_pairs $TRAIN_TRAIN_TASK_LANG_PAIRS \
                --eval_task_language_pairs $EVAL_TASK_LANG_PAIRS \
                --condition_to_layer_id \
                --project_source_embeddings \
                --projected_source_embedding_dim $HYPERNET_DIM \
                --adapter_dim $ADAPTER_DIM \
                --learning_rate 1e-4 \
                --warmup_steps 4000 \
                --freeze_params_regex '^(?!.*(hypernet|LayerNorm)).*' \
                --fp16

Evaluate Model Checkpoints

To evaluate a checkpoint with the test data provide the checkpoint folder, list of test task-language pairs and a batch size

--test_task_language_pairs: Test task-language pairs, e.g. 'ner#tr' 'pos#ar'

python3 evaluate.py --model_name_or_path 'bert-base-multilingual-cased' \
                --eval_ckpt $CKPT_FOLDER \
                --test_task_language_pairs $TEST_TASK_LANG_PAIRS \
                --per_device_eval_batch_size 32

Acknowledgement

This repository is initially based on a tutorial given by Jason Phang. For the implementation of adapter modules, Hyperformer repository has been used.

Citing This Research

@inproceedings{ustun-etal-2022-hyperx,
    title = {Hyper-X: A Unified Hypernetwork for Multi-Task Multilingual Transfer},
    author = {{\"U}st{\"u}n, Ahmet  and
      Bisazza, Arianna  and
      Bouma, Gosse  and
      van Noord, Gertjan and
      Ruder, Sebastian},
    booktitle = {Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    year = {2022},
    url = {https://arxiv.org/abs/2205.12148},
}

ahmetustun / hyperx