gamaly/bert-vessel-ner

BERT-NER Vessel Detection Model

Model Description

This model is a fine-tuned version of dslim/bert-base-NER for detecting vessels (ships) and organizations in maritime news articles and documents.

Key Features

• Vessel Detection: Identifies ship names in text (mapped to MISC slot)
• Organization Detection: Identifies maritime organizations, ship owners, operators, and related entities (uses ORG slot)
• Multilingual Support: Trained on English, Spanish, Chinese, French, Portuguese, Russian, and other languages
• Preserves Base Model: Maintains original PER, LOC, and other entity detection capabilities

Model Architecture

• Base Model: dslim/bert-base-NER
• Task: Token Classification (Named Entity Recognition)
• Labels: O, B-PER, I-PER, B-ORG, I-ORG, B-LOC, I-LOC, B-MISC, I-MISC
  • VESSEL entities are mapped to MISC slot (B-MISC, I-MISC)
  • Organization entities use ORG slot (B-ORG, I-ORG)

Model Performance

Evaluation Metrics

• Precision: 1.0000
• Recall: 1.0000
• F1 Score: 1.0000

Note: Metrics reported on validation set. Real-world performance may vary.

Example Predictions

Text	Detected Entities
"The fishing vessel Hai Feng 718 was detained by authorities."	Hai Feng 718 (VESSEL: 1.00)
"Coast guard seized the trawler Thunder near disputed waters."	Thunder (VESSEL: 1.00)
"Pacific Seafood Inc. announced quarterly earnings today."	Pacific Seafood Inc (ORG: 0.95)
"The vessel Thunder owned by Pacific Seafood Inc. was seized."	Thunder (VESSEL: 1.00), Pacific Seafood Inc (ORG: 0.98)

Training Details

Training Data

• Total Examples: ~60,000 synthetic multilingual examples
• VESSEL Examples: ~20,000
• ORG Examples: ~40,000 (ship owners, operators, brands, retailers, importers, fishmeal plants, etc.)
• Languages: English, Spanish, Chinese, French, Portuguese, Russian, and others
• Source: Synthetically generated from maritime entity databases

Training Procedure

• Base Model: dslim/bert-base-NER
• Training Epochs: 3
• Batch Size: 32
• Learning Rate: 2e-5
• Max Sequence Length: 128 tokens
• Optimizer: AdamW with weight decay 0.01
• Mixed Precision: FP16 enabled

Training Configuration

TrainingArguments(
    num_train_epochs=3,
    per_device_train_batch_size=32,
    learning_rate=2e-5,
    weight_decay=0.01,
    max_length=128,
    fp16=True
)

How to Use

Direct Use

from transformers import pipeline

# Load the model
ner = pipeline("ner", model="your-username/bert-vessel-ner", aggregation_strategy="simple")

# Example 1: Vessel detection
text = "The fishing vessel Hai Feng 718 was detained by authorities."
entities = ner(text)
for entity in entities:
    print(f"{entity['word']} -> {entity['entity_group']} ({entity['score']:.2f})")
# Output: Hai Feng 718 -> MISC (1.00)  # MISC = VESSEL

# Example 2: Mixed entities
text = "The vessel Thunder owned by Pacific Seafood Inc. was seized."
entities = ner(text)
for entity in entities:
    print(f"{entity['word']} -> {entity['entity_group']} ({entity['score']:.2f})")
# Output: 
# Thunder -> MISC (1.00)  # VESSEL
# Pacific Seafood Inc -> ORG (0.98)  # Organization

Advanced Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

# Load model and tokenizer
model = AutoModelForTokenClassification.from_pretrained("your-username/bert-vessel-ner")
tokenizer = AutoTokenizer.from_pretrained("your-username/bert-vessel-ner")

# Tokenize and predict
text = "The vessel Thunder was seized."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_ids = torch.argmax(predictions, dim=-1)

# Decode predictions
tokens = tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])
for token, pred_id in zip(tokens, predicted_ids[0]):
    if token not in ['[CLS]', '[SEP]', '[PAD]']:
        label = model.config.id2label[pred_id.item()]
        print(f"{token}: {label}")

Post-Processing

Note: The model outputs VESSEL entities as MISC labels. You may want to rename them for clarity:

entities = ner(text)
for entity in entities:
    # Rename MISC to VESSEL for clarity
    if entity['entity_group'] == 'MISC':
        entity['entity_group'] = 'VESSEL'
    print(f"{entity['word']} -> {entity['entity_group']}")

Limitations and Bias

Known Limitations

1. False Positives: May occasionally classify organization names as vessels if they resemble ship names (e.g., "Pacific Seafood Inc."). Use a higher threshold (0.98+) to reduce false positives.

2. Multilingual Performance: While trained on multiple languages, performance may vary by language. Best results on English, Spanish, and Chinese.

3. Domain Specificity: Trained primarily on maritime crime and enforcement contexts. Performance may vary in other domains (e.g., commercial shipping, recreational boating).

4. Synthetic Data: Model was trained on synthetically generated data. Real-world performance may differ from validation metrics.

Recommendations

• Threshold Tuning: Adjust the confidence threshold based on your use case:
  • High precision (fewer false positives): Use threshold ≥ 0.98
  • High recall (catch more vessels): Use threshold ≥ 0.90
• Post-Processing: Consider adding rules to filter obvious false positives (e.g., entities containing "Inc.", "Co.", "Ltd.")
• Domain Adaptation: For best results in specific domains, consider fine-tuning on domain-specific data

Training Data Sources

The model was trained on synthetically generated data from:

• Maritime vessel databases
• Ship owner and operator registries
• Brand and retailer information
• Fishmeal plant and processor databases

All training data was synthetically generated using large language models (Gemini 2.5 Flash-Lite) to create realistic maritime news contexts.

Evaluation

Test Set

• Size: ~2,280 examples (10% of total data)
• Distribution: Balanced across languages and entity types
• Metrics: Precision, Recall, F1 Score

Performance by Entity Type

Entity Type	Precision	Recall	F1
VESSEL (MISC)	1.0000	1.0000	1.0000
ORG	1.0000	1.0000	1.0000

Note: Metrics on validation set. Real-world performance may vary.

Environmental Impact

• Hardware: GPU (CUDA)
• Training Time: ~3 minutes per epoch (total ~10 minutes)
• Carbon Emissions: Minimal (short training duration)

Citation

If you use this model, please cite:

@misc{bert-vessel-ner,
  title={BERT-NER Vessel Detection Model},
  author={Your Name},
  year={2025},
  howpublished={\url{https://huggingface.co/your-username/bert-vessel-ner}}
}

Model Card Contact

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

License

This model is licensed under Apache 2.0, same as the base model dslim/bert-base-NER.