evelinamorim/bert-lusa-eventype-classifier

Model Description

This model is a fine-tuned BERT model for event type classification in English news text. It identifies events and classifies them into three types:

• State: Events describing states or conditions
• Process: Events describing ongoing processes or activities
• Transition: Events describing changes or transitions

The model uses BIO tagging for sequence labeling.

Training Data

• Dataset: LUSA News Events (English)
• Training documents: 81
• Annotation source: INCEpTION annotation tool
• Event types: State, Process, Transition

Model Performance

Evaluated on test set:

Class	Precision	Recall	F1-Score	support
Event	0	0	0	9
Process	0.39	0.46	0.42	63
State	0.35	0.59	0.44	116
Transition	0.66	0.73	0.69	294
Overall	0.35	0.44	0.39	482

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

# Load model and tokenizer
model_name = "evelinamorim/bert-lusa-eventype-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

# Example text
text = "The company announced a new product launch yesterday."

# Tokenize
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.argmax(outputs.logits, dim=2)

# Decode predictions
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
predicted_labels = [model.config.id2label[p.item()] for p in predictions[0]]

# Print results
for token, label in zip(tokens, predicted_labels):
    if label != "O":
        print(f"{token}: {label}")

Label Schema

• O: Outside any event
• B-State: Beginning of State event
• I-State: Inside State event
• B-Process: Beginning of Process event
• I-Process: Inside Process event
• B-Transition: Beginning of Transition event
• I-Transition: Inside Transition event
• B-Event: Beginning of event without type
• I-Event: Inside of event without type

Training Details

• Base model: bert-base-cased
• Training epochs: 20
• Batch size: 16
• Learning rate: 2e-5
• Loss function: Softer (power=0.5) Weighted Cross-Entropy (to handle class imbalance)
• Optimizer: AdamW
• Framework: Hugging Face Transformers

Limitations

• Trained on news domain - may not generalize to other domains
• Small training dataset (81 documents)
• Performance varies by event type (Transition performs best)
• May produce false positives on ambiguous cases

Citation

If you use this model, please cite:

@misc{lusa_event_classifier,
  author = {Evelin Amorim},
  title = {evelinamorim/bert-lusa-eventype-classifier},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/evelinamorim/bert-lusa-eventype-classifier}}
}

Contact

For questions or issues, please open an issue on the model repository.