File size: 9,446 Bytes
248ac9b eb087ab 05139f2 c0e2471 248ac9b 05139f2 c0e2471 05139f2 d3f4e36 05139f2 c0e2471 05139f2 c0e2471 05139f2 0da0849 05139f2 c0e2471 05139f2 c0e2471 05139f2 c0e2471 05139f2 c0e2471 05139f2 22d7b25 a101ab9 22d7b25 da955ed 9d591bb da955ed a101ab9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 |
---
license: cc-by-sa-4.0
language:
- hr
- sl
- bs
- sr
datasets:
- classla/xlm-r-bertic-data
---
# XLM-R-SloBertić
This model was produced by pre-training [XLM-Roberta-large](https://huggingface.co/xlm-roberta-large) 48k steps on South Slavic languages using [XLM-R-BERTić dataset](https://huggingface.co/datasets/classla/xlm-r-bertic-data)
# Benchmarking
Three tasks were chosen for model evaluation:
* Named Entity Recognition (NER)
* Sentiment regression
* COPA (Choice of plausible alternatives)
In all cases, this model was finetuned for specific downstream tasks.
## NER
Mean F1 scores were used to evaluate performance. Datasets used: [hr500k](https://huggingface.co/datasets/classla/hr500k), [ReLDI-sr](https://huggingface.co/datasets/classla/reldi_sr), [ReLDI-hr](https://huggingface.co/datasets/classla/reldi_hr), and [SETimes.SR](https://huggingface.co/datasets/classla/setimes_sr).
| system | dataset | F1 score |
|:-----------------------------------------------------------------------|:--------|---------:|
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | hr500k | 0.927 |
| [BERTić](https://huggingface.co/classla/bcms-bertic) | hr500k | 0.925 |
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | hr500k | 0.923 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | hr500k | 0.919 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | hr500k | 0.918 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | hr500k | 0.903 |
| system | dataset | F1 score |
|:-----------------------------------------------------------------------|:---------|---------:|
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | ReLDI-hr | 0.812 |
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | ReLDI-hr | 0.809 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | ReLDI-hr | 0.794 |
| [BERTić](https://huggingface.co/classla/bcms-bertic) | ReLDI-hr | 0.792 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | ReLDI-hr | 0.791 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | ReLDI-hr | 0.763 |
| system | dataset | F1 score |
|:-----------------------------------------------------------------------|:-----------|---------:|
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | SETimes.SR | 0.949 |
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | SETimes.SR | 0.940 |
| [BERTić](https://huggingface.co/classla/bcms-bertic) | SETimes.SR | 0.936 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | SETimes.SR | 0.933 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | SETimes.SR | 0.922 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | SETimes.SR | 0.914 |
| system | dataset | F1 score |
|:-----------------------------------------------------------------------|:---------|---------:|
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | ReLDI-sr | 0.841 |
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | ReLDI-sr | 0.824 |
| [BERTić](https://huggingface.co/classla/bcms-bertic) | ReLDI-sr | 0.798 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | ReLDI-sr | 0.774 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | ReLDI-sr | 0.751 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | ReLDI-sr | 0.734 |
## Sentiment regression
[ParlaSent dataset](https://huggingface.co/datasets/classla/ParlaSent) was used to evaluate sentiment regression for Bosnian, Croatian, and Serbian languages.
The procedure is explained in greater detail in the dedicated [benchmarking repository](https://github.com/clarinsi/benchich/tree/main/sentiment).
| system | train | test | r^2 |
|:-----------------------------------------------------------------------|:--------------------|:-------------------------|------:|
| [xlm-r-parlasent](https://huggingface.co/classla/xlm-r-parlasent) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.615 |
| [BERTić](https://huggingface.co/classla/bcms-bertic) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.612 |
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.607 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.605 |
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.601 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.537 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | 0.500 |
| dummy (mean) | ParlaSent_BCS.jsonl | ParlaSent_BCS_test.jsonl | -0.12 |
## COPA
| system | dataset | Accuracy score |
|:-----------------------------------------------------------------------|:--------|---------------:|
| [BERTić](https://huggingface.co/classla/bcms-bertic) | Copa-SR | 0.689 |
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | Copa-SR | 0.665 |
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | Copa-SR | 0.637 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | Copa-SR | 0.607 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | Copa-SR | 0.573 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | Copa-SR | 0.570 |
| system | dataset | Accuracy score |
|:-----------------------------------------------------------------------|:--------|---------------:|
| [BERTić](https://huggingface.co/classla/bcms-bertic) | Copa-HR | 0.669 |
| [**XLM-R-SloBERTić**](https://huggingface.co/classla/xlm-r-slobertic) | Copa-HR | 0.628 |
| [XLM-R-BERTić](https://huggingface.co/classla/xlm-r-bertic) | Copa-HR | 0.635 |
| [crosloengual-bert](https://huggingface.co/EMBEDDIA/crosloengual-bert) | Copa-HR | 0.669 |
| [XLM-Roberta-Base](https://huggingface.co/xlm-roberta-base) | Copa-HR | 0.585 |
| [XLM-Roberta-Large](https://huggingface.co/xlm-roberta-large) | Copa-HR | 0.571 |
# Citation
<!---
The following paper has been submitted for review:
```
@misc{ljubesic2024language,
author = "Ljube\v{s}i\'{c}, Nikola and Suchomel, Vit and Rupnik, Peter and Kuzman, Taja and van Noord, Rik",
title = "Language Models on a Diet: Cost-Efficient Development of Encoders for Closely-Related Languages via Additional Pretraining",
howpublished = "Submitted for review",
year = "2024",
}
```
--->
Please cite the following paper:
```
@article{Ljubešić_Suchomel_Rupnik_Kuzman_van Noord_2024,
title={Language Models on a Diet: Cost-Efficient Development of Encoders for Closely-Related Languages via Additional Pretraining},
url={http://arxiv.org/abs/2404.05428},
DOI={10.48550/arXiv.2404.05428},
abstractNote={The world of language models is going through turbulent times, better and ever larger models are coming out at an unprecedented speed. However, we argue that, especially for the scientific community, encoder models of up to 1 billion parameters are still very much needed, their primary usage being in enriching large collections of data with metadata necessary for downstream research. We investigate the best way to ensure the existence of such encoder models on the set of very closely related languages - Croatian, Serbian, Bosnian and Montenegrin, by setting up a diverse benchmark for these languages, and comparing the trained-from-scratch models with the new models constructed via additional pretraining of existing multilingual models. We show that comparable performance to dedicated from-scratch models can be obtained by additionally pretraining available multilingual models even with a limited amount of computation. We also show that neighboring languages, in our case Slovenian, can be included in the additional pretraining with little to no loss in the performance of the final model.},
note={arXiv:2404.05428 [cs]},
number={arXiv:2404.05428},
publisher={arXiv},
author={Ljubešić, Nikola and Suchomel, Vít and Rupnik, Peter and Kuzman, Taja and van Noord, Rik},
year={2024},
month=apr
}
```
|