bonadossou/multitask_model_fon_False_multiplicative

• Paper: FonMTL: Toward Building a Multi-Task Learning Model for Fon Language, accepted at WiNLP co-located at EMNLP 2023

• Official Github: https://github.com/bonaventuredossou/multitask_fon

• Build Multi-task Learning Model: For the shared layers (encoders) we used the following language model heads:

  • AfroLM-Large

    • AfroLM: A Self-Active Learning-based Multilingual Pretrained Language Model for 23 African Languages (Dossou et.al., EMNLP 2022)

  • XLMR-Large:

    • Unsupervised Cross-lingual Representation Learning at Scale (Conneau et.al., ACL 2020)

• Evaluation:

  • The goal primarily is to explore whether multitask learning improves performance on downstream tasks for Fon. We try two settings: (a) training only on Fon and evaluating on Fon, (b) training on all languages and evaluating on Fon. We evaluate the multi-task learning model on NER and POS tasks, and compare it with baselines (models finetuned and evaluated on single tasks)

How to get started

• Run the training: sbatch run.sh

This command will:

• Set up the environement
• Install required libraries: pip install -r requirements.txt -q
• Move to the code folder: cd code
• Run the training & evaluate: python run_train.py

NER Results

Model	Task	Pretraining/Finetuning Dataset	Pretraining/Finetuning Language(s)	Evaluation Dataset	Metric	Metric's Value
AfroLM-Large	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	80.48
AfriBERTa-Large	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	79.90
XLMR-Base	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	81.90
XLMR-Large	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	81.60
AfroXLMR-Base	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	82.30
AfroXLMR-Large	Single Task	MasakhaNER 2.0	All	FON NER	F1-Score	82.70
:---:	:---:	:---:	:---:	:---:	:---:
MTL Sum (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	All	FON NER	F1-Score	79.87
MTL Weighted (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	All	FON NER	F1-Score	81.92
MTL Weighted (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	Fon Data	FON NER	F1-Score	64.43

POS Results

Model	Task	Pretraining/Finetuning Dataset	Pretraining/Finetuning Language(s)	Evaluation Dataset	Metric	Metric's Value
AfroLM-Large	Single Task	MasakhaPOS	All	FON POS	Accuracy	82.40
AfriBERTa-Large	Single Task	MasakhaPOS	All	FON POS	Accuracy	88.40
XLMR-Base	Single Task	MasakhaPOS	All	FON POS	Accuracy	90.10
XLMR-Large	Single Task	MasakhaPOS	All	FON POS	Accuracy	90.20
AfroXLMR-Base	Single Task	MasakhaPOS	All	FON POS	Accuracy	90.10
AfroXLMR-Large	Single Task	MasakhaPOS	All	FON POS	Accuracy	90.40
:---:	:---:	:---:	:---:	:---:	:---:
MTL Sum (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	All	FON POS	Accuracy	82.45
MTL Weighted (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	All	FON POS	Accuracy	89.20
MTL Weighted (ours)	Multi-Task	MasakhaNER 2.0 & MasakhaPOS	Fon Data	FON POS	Accuracy	80.85

Importance of Merging Representation Type

Merging Type	Models	Task	Metric	Metric's Value
Multiplicative	MTL Weighted (multi-task; ours; *)	NER	F1-Score	81.92
Multiplicative	MTL Weighted (multi-task; ours; +)	NER	F1-Score	64.43
:---:	:---:	:---:	:---:	:---:
Multiplicative	MTL Weighted (multi-task; ours; *)	POS	Accuracy	89.20
Multiplicative & MTL Weighted (multi-task; ours; +)	POS	Accuracy	80.85
:---:	:---:	:---:	:---:	:---:
Additive	MTL Weighted (multi-task; ours; *)	NER	F1-Score	78.91
Additive	MTL Weighted (multi-task; ours; +)	NER	F1-Score	60.93
:---:	:---:	:---:	:---:	:---:
Additive	MTL Weighted (multi-task; ours; *)	POS	Accuracy	86.99
Additive	MTL Weighted (multi-task; ours; +)	POS	Accuracy	78.25

Model End-Points

• multitask_model_fon_False_multiplicative.bin is the MTL Fon Model which has been pre-trained on all MasakhaNER 2.0 and MasakhaPOS datasets, and merging representations in a multiplicative way.

• multitask_model_fon_True_multiplicative.bin is the MTL Fon Model which has been pre-trained only on Fon data from the MasakhaNER 2.0 and MasakhaPOS datasets, and merging representations in a multiplicative way.

How to run inference when you have the model

To run inference with the model(s), you can use the testing block defined in our MultitaskFON class.

TODO

• leverage the impact of the dynamic weighted average loss