BERTicSENTNEG2

This is a negative-polarity classifier for Serbian WordNet synset glosses, fine-tuned from the BERTić model family. It is a component of the S5 sentiment-lexicon construction method described in:

Saša Petalinkar, Ranka M. Stanković, and Milica Ikonić Nešić (2025). Comparative analysis of methods for creating a sentiment lexicon of the Serbian WordNet. The Electronic Library, 43(4), 547–577. Paper and DOI.

Companion code and sentiment lexicons · Paired classifier · Original base model

Model identity and task

Field	Value
Model ID	Tanor/BERTicSENTNEG2
Architecture	ElectraForSequenceClassification
Original base model	classla/bcms-bertic
Input	A Serbian synset gloss, representing one lexical meaning
Target	Negative vs non-negative polarity
Dataset expansion iteration	2 (T2)
Label 0	NON-NEGATIVE
Label 1	NEGATIVE
Derived lexicon family	S5
Paired classifier, same iteration	Tanor/BERTicSENTPOS2

The preparation notebook initializes the BERTić family from classla/bcms-bertic. The linked training script can resume from the corresponding Tanor/BERTicSENT* repository. The metadata identifies the original base model; it does not imply that every run started from that base checkpoint.

A non-negative label is the complement of the target class. It does not by itself mean that the gloss has the opposite polarity. A separate classifier handles that polarity.

Training data and construction method

The paper constructs polarity-labeled synsets from Serbian WordNet, starting from curated positive, negative, and objective seed sets and expanding through semantic relations. The initial sets reported in the paper contain 149 positive, 219 negative, and 19,475 objective synsets. Polarity-preserving relations expand the corresponding set; antonymy contributes to the opposite polarity. The selected datasets are T0, T2, T4, and T6, after zero, two, four, and six expansion iterations.

This checkpoint is associated with T2 and the NEG classification task. Its numerical suffix is a dataset-expansion iteration, not an epoch count or a lexicon identifier.

The training script reads Sysnet from X_train_UPNEG2.csv and the target NEG from y_train_UPNEG2.csv, replaces missing text with an empty string, and creates a stratified validation subset of 10% of that training CSV, with random_state=42. The UP inputs are the non-lemmatized gloss variant in the dataset-generation script.

The paper and code snapshot have different split descriptions: the paper describes an 80/10/10 split for neural models, while the scripts split existing training CSV files and create_sets.py leaves the initial split size at the library default. Exact checkpoint-specific sample assignments are not supplied in the public revision. The train_sets/ files referenced by the scripts are absent from that revision. Reconstructing the experiment requires the relevant lexical resources and saved preparation/split information; the paper's proportions alone do not establish this checkpoint's split.

Use in sentiment lexicon S5

For each iteration, the POS model estimates positive-class probability p_pos and the NEG model estimates negative-class probability p_neg. The lexicon-building code combines them as:

POS = p_pos * (1 - p_neg)
NEG = p_neg * (1 - p_pos)
OBJ = 1 - POS - NEG

It averages each score across the four iteration-specific pairs. A single checkpoint is one component of this construction; its two class probabilities are not the final three lexicon scores.

Iteration	POS classifier	NEG classifier
0	BERTicSENTPOS0	BERTicSENTNEG0
2	BERTicSENTPOS2	BERTicSENTNEG2
4	BERTicSENTPOS4	BERTicSENTNEG4
6	BERTicSENTPOS6	BERTicSENTNEG6

Usage

This example reads one gloss and returns both class probabilities. It pins the model to the weights revision present before the documentation update.

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

MODEL_ID = "Tanor/BERTicSENTNEG2"
WEIGHTS_REVISION = "7a3a40b2f182ee53aad42aee2e8c84dfc24cb7a1"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, revision=WEIGHTS_REVISION)
model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_ID, revision=WEIGHTS_REVISION
)
model.eval()

# An illustrative gloss, not a benchmark item or a claimed prediction.
inputs = tokenizer(
    "koji oseća radost i zadovoljstvo",
    return_tensors="pt", truncation=True, max_length=300,
)
with torch.inference_mode():
    probabilities = model(**inputs).logits.softmax(dim=-1)[0]
scores = {model.config.id2label[i]: float(p) for i, p in enumerate(probabilities)}
print(scores)

The model uses standard PyTorch/Transformers sequence-classification classes, without custom remote code. The example requires PyTorch and Transformers. Where a Trainer record is available below, it includes software versions from the original run. No example prediction or benchmark score was generated for this documentation update.

Evaluation and provenance

Archived test report

The companion repository contains an evaluation report for BERTic, NEG, T2. It is preserved below, with class 0 = NON-NEGATIVE and class 1 = NEGATIVE. Metric values retain the source's rounding; support values are sample counts. The confusion matrix uses that class order.

[[4389   31]
 [  18   58]]

              precision    recall  f1-score   support

           0       1.00      0.99      0.99      4420
           1       0.65      0.76      0.70        76

    accuracy                           0.99      4496
   macro avg       0.82      0.88      0.85      4496
weighted avg       0.99      0.99      0.99      4496

This is an archived experiment artifact, not a new evaluation. The report does not record a model-weight SHA, so its exact correspondence to the currently hosted weights has not been independently re-established. These binary-classification metrics are separate from evaluation of the derived lexicon and from sentiment evaluation of full sentences.

Preserved Trainer record

The following validation summary and training log are retained from the previous model card, without recomputing them. The recorded F1 is a training-validation metric, separate from the archived test report above and from lexicon or sentence-level sentiment evaluation. The source training code uses binary F1 for target label 1 when eval="f1" is selected.

• Loss: 0.0572
• F1: 0.625

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 64
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 32

Training results

Training Loss	Epoch	Step	Validation Loss	F1
No log	0.9895	47	0.0467	0.5263
No log	2.0	95	0.0482	0.6038
No log	2.9895	142	0.0442	0.7143
No log	4.0	190	0.0567	0.6275
No log	4.9895	237	0.0573	0.6
No log	6.0	285	0.0572	0.625

Framework versions

• Transformers 4.40.1
• Pytorch 2.2.2
• Datasets 2.19.0
• Tokenizers 0.19.1

The Trainer record and the linked source script are separate provenance sources. In particular, the recorded optimizer may differ from the script's adafactor setting; the historical record is retained without asserting that the linked script reproduces that exact run.

Settings in the archived training source

These settings describe the linked code revision, not a replacement for the per-run Trainer record.

Setting	Source-code value
Maximum tokenized input length	300
Learning rate	2e-5
Training batch size per device	64
Evaluation batch size per device	16
Gradient accumulation	4 steps
Optimizer	Adafactor
Weight decay	0.01
Validation split seed	42
Evaluation and saving	Each epoch
Early stopping patience	3 evaluation calls
Epoch budget in experiment notebooks	Up to 32

The paper reports early-stopping patience of 3. The linked family script also uses 3.

Intended use and limitations

• Intended for research on polarity of Serbian WordNet meanings and construction or analysis of Serbian sentiment lexicons.
• Training inputs are glosses. Performance on reviews, news, social media, and documents requires separate evaluation; this checkpoint is not documented as a general sentence-sentiment benchmark model.
• Class imbalance is visible in the archived report. Read accuracy and weighted averages alongside target-class precision, recall, F1, and support.
• Labels depend on seed selection and semantic-relation expansion. Polysemy, domain-specific polarity, and propagation errors can affect predictions. English-to-Serbian alignment alone does not establish polarity in Serbian.
• Softmax outputs are model scores; probability calibration is not established by this card.
• The model performs no sense selection for a word in context. Applying the lexicon to text requires a separate choice or aggregation of meanings.

License

This fine-tuned model is distributed under Apache License 2.0 (apache-2.0), matching the license declared for its original base model, classla/bcms-bertic. Read the full LICENSE and NOTICE for attribution and the description of the fine-tuning changes.

The license permits use, modification, and distribution, including commercial use, subject to its terms. When redistributing, include the license, retain applicable attribution and notices, and mark modifications. It does not require every derivative work to use the same license.

Scope and training-resource permissions

The model license covers the model distribution and accompanying documentation within the rights the licensors can grant. External lexical resources, training datasets, and companion code retain their own terms.

An earlier description of Serbian WordNet reports CC BY-NC terms for the downloadable resource (Developing and Maintaining a WordNet: Procedures and Tools, 2014). The terms of the exact SrpWN release used for this fine-tuning, and any separate permission covering it, have not been verified in this documentation update. That historical statement alone does not determine the license of the trained weights. This model-license declaration does not grant rights to redistribute SrpWN or establish that every third-party permission for a particular use has been cleared. The remaining check is to identify the training release and record the applicable permission from its rights holders. See also Creative Commons guidance on AI training.

License declaration history

The apache-2.0 declaration, full license text, and attribution were added on 23 September 2026.

Citation

When using this model family or the resulting lexicon-construction method, cite the paper and record the model ID and revision used.

@article{petalinkar2025sentimentlexicon,
  author = {Petalinkar, Saša and Stanković, Ranka M. and Ikonić Nešić, Milica},
  title = {Comparative analysis of methods for creating a sentiment lexicon of the Serbian WordNet},
  journal = {The Electronic Library},
  year = {2025},
  volume = {43},
  number = {4},
  pages = {547--577},
  doi = {10.1108/EL-08-2024-0253},
  url = {https://doi.org/10.1108/EL-08-2024-0253}
}

Version and documentation sources

• Weights/configuration revision documented here: 7a3a40b2f182ee53aad42aee2e8c84dfc24cb7a1. This pins the hosted checkpoint; it does not prove which weights produced every table in the paper.
• Companion code revision: 833582dbbf561a902fc5b872248db12bf529b3d9.
• BERTić initialization is recorded in Prepare and load models.ipynb; GPT2-Orao and Jerteh-355 initialization is recorded in their training scripts.
• Documentation expanded on 23 September 2026 using the paper, model configuration, repository source, archived evaluation report, and any pre-existing Trainer record. Weights, tokenizer files, and model configuration were not changed by this documentation update.