XLMRobertaGreekNews

Model Description

This model is finetuned version of XLM RoBERTa

Dataset

The model finetuned on the GreekNews-20k dataset.

Results

Perfomance on the GreekNews-20k dataset :

Performance on the elNER dataset :

To use this model

pip install transformers==4.46.1, torch

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("katrjohn/XLMRobertaGreekNews", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

Example usage

import torch

# Classification label dictionary (reverse)
classification_label_dict_reverse = {
    0: "Αυτοκίνητο", 1: "Επιχειρήσεις και βιομηχανία", 2: "Έγκλημα και δικαιοσύνη",
    3: "Ειδήσεις για καταστροφές και έκτακτες ανάγκες", 4: "Οικονομικά και χρηματοοικονομικά", 5: "Εκπαίδευση",
    6: "Ψυχαγωγία και πολιτισμός", 7: "Περιβάλλον και κλίμα", 8: "Οικογένεια και σχέσεις",
    9: "Μόδα", 10: "Τρόφιμα και ποτά", 11: "Υγεία και ιατρική", 12: "Μεταφορές και υποδομές",
    13: "Ψυχική υγεία και ευεξία", 14: "Πολιτική και κυβέρνηση", 15: "Θρησκεία",
    16: "Αθλητισμός", 17: "Ταξίδια και αναψυχή", 18: "Τεχνολογία και επιστήμη"
}

ner_label_set = ["PAD", "O",
    "B-ORG", "I-ORG", "B-PERSON", "I-PERSON", "B-CARDINAL", "I-CARDINAL",
    "B-GPE", "I-GPE", "B-DATE", "I-DATE", "B-ORDINAL", "I-ORDINAL",
    "B-PERCENT", "I-PERCENT", "B-LOC", "I-LOC", "B-NORP", "I-NORP",
    "B-MONEY", "I-MONEY", "B-TIME", "I-TIME", "B-EVENT", "I-EVENT",
    "B-PRODUCT", "I-PRODUCT", "B-FAC", "I-FAC", "B-QUANTITY", "I-QUANTITY"
]
tag2idx = {t:i for i,t in enumerate(ner_label_set)}
idx2tag = {i:t for t,i in tag2idx.items()}

sentence = "Ο Κυριάκος Μητσοτάκης επισκέφθηκε τη Θεσσαλονίκη για τα εγκαίνια της ΔΕΘ."
inputs = tokenizer(sentence, return_tensors="pt")

with torch.no_grad():
    classification_logits, ner_logits = model(**inputs)

# Classification
classification_probs = torch.softmax(classification_logits, dim=-1)
predicted_class = torch.argmax(classification_probs, dim=-1).item()
predicted_class_label = classification_label_dict_reverse.get(predicted_class, "Unknown")

print(f"Predicted class index: {predicted_class}")
print(f"Predicted class label: {predicted_class_label}")

# NER
ner_predictions = torch.argmax(ner_logits, dim=-1).squeeze().tolist()
tokens = tokenizer.convert_ids_to_tokens(inputs['input_ids'].squeeze())

for token, pred_idx in zip(tokens, ner_predictions):
    tag = idx2tag.get(pred_idx, "O")
    print(f"{token}: {tag}")

Output:

Predicted class index: 6
Predicted class label: Ψυχαγωγία και πολιτισμός
<s>: O
▁Ο: O
▁Κυρι: B-PERSON
άκος: B-PERSON
▁Μητσοτάκη: I-PERSON
ς: O
▁επι: O
σκέ: O
φ: O
θηκε: O
▁τη: O
▁Θεσσαλονίκη: B-GPE
▁για: O
▁τα: O
▁εγκ: O
αί: O
νια: O
▁της: O
▁Δ: B-EVENT
ΕΘ: I-EVENT
.: O
</s>: O

Author

This model has been released along side with the article: Named Entity Recognition and News Article Classification: A Lightweight Approach.

To use this model please cite the following:

@ARTICLE{11148234,
  author={Katranis, Ioannis and Troussas, Christos and Krouska, Akrivi and Mylonas, Phivos and Sgouropoulou, Cleo},
  journal={IEEE Access}, 
  title={Named Entity Recognition and News Article Classification: A Lightweight Approach}, 
  year={2025},
  volume={13},
  number={},
  pages={155031-155046},
  keywords={Accuracy;Transformers;Pipelines;Named entity recognition;Computational modeling;Vocabulary;Tagging;Real-time systems;Benchmark testing;Training;Distilled transformer;edge-deployable model;multiclass news-topic classification;named entity recognition},
  doi={10.1109/ACCESS.2025.3605709}}