bcms-bertic-parlasent-bcs-ter

Ternary text classification model based on classla/bcms-bertic and fine-tuned on the BCS Political Sentiment dataset (sentence-level data).

This classifier classifies text into only three categories: Negative, Neutral, and Positive. For the binary classifier (Negative, Other) check this model.

For details on the dataset and the finetuning procedure, please see this paper.

Fine-tuning hyperparameters

Fine-tuning was performed with simpletransformers. Beforehand a brief sweep for the optimal number of epochs was performed and the presumed best value was 9. Other arguments were kept default.

model_args = {
        "num_train_epochs": 9
        }

Performance

The same pipeline was run with two other transformer models and fasttext for comparison. Macro F1 scores were recorded for each of the 6 fine-tuning sessions and post festum analyzed.

model	average macro F1
bcms-bertic-parlasent-bcs-ter	0.7941 ± 0.0101 **
EMBEDDIA/crosloengual-bert	0.7709 ± 0.0113
xlm-roberta-base	0.7184 ± 0.0139
fasttext + CLARIN.si embeddings	0.6312 ± 0.0043

Two best performing models have been compared with the Mann-Whitney U test to calculate p-values (** denotes p<0.01).

Use example with simpletransformers==0.63.7

from simpletransformers.classification import ClassificationModel

model = ClassificationModel("electra", "classla/bcms-bertic-parlasent-bcs-ter")

predictions, logits = model.predict([
    "Vi niste normalni",
    "Đački autobusi moraju da voze svaki dan",
    "Ovo je najbolji zakon na svetu",
     ]
    )

predictions
# Output: array([0, 1, 2])

[model.config.id2label[i] for i in predictions]
# Output: ['Negative', 'Neutral', 'Positive']

Citation

If you use the model, please cite the following paper on which the original model is based:

@inproceedings{ljubesic-lauc-2021-bertic,
    title = "{BERT}i{\'c} - The Transformer Language Model for {B}osnian, {C}roatian, {M}ontenegrin and {S}erbian",
    author = "Ljube{\v{s}}i{\'c}, Nikola  and Lauc, Davor",
    booktitle = "Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing",
    month = apr,
    year = "2021",
    address = "Kiyv, Ukraine",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.bsnlp-1.5",
    pages = "37--42",
}

and the paper describing the dataset and methods for the current finetuning:

@misc{https://doi.org/10.48550/arxiv.2206.00929,
  doi = {10.48550/ARXIV.2206.00929},
  
  url = {https://arxiv.org/abs/2206.00929},
  
  author = {Mochtak, Michal and Rupnik, Peter and Ljubešič, Nikola},
  
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {The ParlaSent-BCS dataset of sentiment-annotated parliamentary debates from Bosnia-Herzegovina, Croatia, and Serbia},
  
  publisher = {arXiv},
  
  year = {2022},
  
  copyright = {Creative Commons Attribution Share Alike 4.0 International}
}