This is a SetFit model that can be used for Text Classification of actions in asynchronous conversation. This particular model detects if a comment includes a challenge or not. The configuration of the model is that the model is based on only one annotator's annotations (annotator A2). Metric evaluations are based on conservative ground truth (see paper). This SetFit model uses TurkuNLP/bert-base-finnish-cased-v1 as the Sentence Transformer embedding model (using word embeddings). A LogisticRegression instance is used for classification. The model has been trained using an efficient few-shot learning technique that involves:

Fine-tuning a Sentence Transformer with contrastive learning.
Training a classification head with features from the fine-tuned Sentence Transformer.

Model Details

Model Description

Model Type: SetFit
Sentence Transformer body: TurkuNLP/bert-base-finnish-cased-v1
Classification head: a LogisticRegression instance
Maximum Sequence Length: 512 tokens
Number of Classes: 2 classes (action present yes/no))

Model Sources

Repository: GitHub
Paper: Paakki, H., Toivanen, P. and Kajava K. (2025). Implicit and Indirect: Detecting Face-threatening and Paired Actions in Asynchronous Online Conversations. Northern European Journal of Language Technology (NEJLT), 11(1), pp. 58-83.

Model Labels

Label

Examples

'Etunimi Sukunimi herra senkun aloittaa keräyksen♥️'
'Etunimi Sukunimi venäjän syy hintojen nousu vai syytätkö sodastakin Suomen hallitusta ? 😖'
'Etunimi Sukunimi Olikhaan se virve'

'Etunimi Sukunimi Ei tarkoita vaihtoehdottomuutta. Vaan että jos hakee, niin kannattaa valmistautua kaikenlaiseen, "roskinpia" rajaloukkauksia myöten. Valmistautuminen ei tarkoita, etteikö silti voisi hakea, vaan sitä että voi taklata vaikutusyritykset mahdollisimman hyvin.'
'Etunimi Etunimi Sukunimi, mikä on vastaus sössöttäjille, mahtava kuulla vastauksesi??'
'Etunimi Sukunimi eli he ilmiselvästi tekee töitä palkatta? Miksi sit puhutaan sairaanhoitajien bonareista ja palkankorotuksista? Meillä on paljon isompia ongelmia käsissämme kuin tämä pandemia joka tilastojenkin puolesta on rokottamattomien pandemia.. mielenterveys ongelmat huippulukemissa ja leikkauksia siirretään samalla kuin kansa lihavoituu entisestään! Näitäkään ei saada hoidettu pyhällä hengellä ja lainanotto ei ole ratkaisu'

Evaluation

Metrics

Label	Metric
all	0.76

Uses

Direct Use for Inference

First install the SetFit library:

pip install setfit

Then you can load this model and run inference.

python

from setfit import SetFitModel

# Download from the 🤗 Hub
model = SetFitModel.from_pretrained("Finnish-actions/SetFit-FinBERT1-A2-challenge")
# Run inference
preds = model("Etunimi Sukunimi minä myös 🙋‍♀️")

Downstream Use

NB. This model has been trained on data coming from Finnish language asynchronous conversations under crisis related news on Facebook. This specific model has been trained to detect whether a comment includes a question or not. It reflects only one of our annotators' label interpretations, so the best use of our models (see our paper) would be to combine a set of models we provide on our Huggingface (Finnish-actions), and use a model ensemble to provide label predictions. It needs to be noted also that the model may not be well applicable outside of its empirical context, so in downstream applications, one should always conduct an evaluation of the model applicability using manually annotated data from that specific context (see our paper for annotation instructions).

Out-of-Scope Use

Please use this model only for action detection and analysis. Uses of this model and the involved data for generative purposes (e.g. NLG) is prohibited.

Bias, Risks and Limitations

Note that the model may produce errors. Due to the size of the training dataset, model may not generalize very well even for other novel topics within the same context. Note that model predictions should not be regarded as final judgments e.g. for online moderation purposes, but each case should also be regarded individually if using model predictions to support moderation. Also, the annotations only reflect three (though experienced) annotators' interpretations, so there might be perspectives on data intepretation that have not been taken into account here. If model is used to support moderation on social media, we recommend that final judgments should always be left for human moderators.

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	1	20.3115	213

Label	Training Sample Count
0	685
1	278

Training Hyperparameters

batch_size: (16, 16)
num_epochs: (4, 4)
max_steps: -1
sampling_strategy: oversampling
num_iterations: 6
body_learning_rate: (2e-05, 1e-05)
head_learning_rate: 0.01
loss: CosineSimilarityLoss
distance_metric: cosine_distance
margin: 0.25
end_to_end: False
use_amp: False
warmup_proportion: 0.1
l2_weight: 0.01
seed: 42
evaluation_strategy: epoch
eval_max_steps: -1
load_best_model_at_end: False

Framework Versions

Python: 3.11.9
SetFit: 1.1.3
Sentence Transformers: 3.2.0
Transformers: 4.44.0
PyTorch: 2.4.0+cu124
Datasets: 2.21.0
Tokenizers: 0.19.1

Citation

If you use this model, please cite the following work:

BibTeX

bibtex

@article{paakki-implicit-indirect,
    doi = {https://doi.org/10.3384/nejlt.2000-1533.2025.5980},
    url = {https://nejlt.ep.liu.se/article/view/5980},
    author = {Paakki, Henna and Toivanen, Pihla and Kajava, Kaisla},
    title = {Implicit and Indirect: Detecting Face-threatening and Paired Actions in Asynchronous Online Conversations},
    publisher = {Northern European Journal of Language Technology (NEJLT)},
    volume= {11},
    number= {1},
    year = {2025}
}