stefan-it/hmbench-letemps-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-2

Fine-tuned Flair Model on LeTemps French NER Dataset (HIPE-2022)

This Flair model was fine-tuned on the LeTemps French NER Dataset using hmBERT Tiny as backbone LM.

The LeTemps dataset consists of NE-annotated historical French newspaper articles from mid-19C to mid 20C.

The following NEs were annotated: loc, org and pers.

Results

We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:

• Batch Sizes: [4, 8]
• Learning Rates: [5e-05, 3e-05]

And report micro F1-score on development set:

Configuration	Seed 1	Seed 2	Seed 3	Seed 4	Seed 5	Average
bs4-e10-lr5e-05	0.5297	0.5073	0.5106	0.5111	0.5282	0.5174 ± 0.0107
bs4-e10-lr3e-05	0.5012	0.4703	0.5019	0.4857	0.5072	0.4933 ± 0.0151
bs8-e10-lr5e-05	0.5027	0.4727	0.5021	0.4912	0.4733	0.4884 ± 0.0148
bs8-e10-lr3e-05	0.489	0.4226	0.4656	0.4744	0.4511	0.4605 ± 0.0253

The training log and TensorBoard logs (not available for hmBERT Base model) are also uploaded to the model hub.

More information about fine-tuning can be found here.

Acknowledgements

We thank Luisa März, Katharina Schmid and Erion Çano for their fruitful discussions about Historic Language Models.

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC). Many Thanks for providing access to the TPUs ❤️