Elpapudex/Pos-full

u2t01-bert-pos

Training Data

• Dataset: universal-dependencies/universal_dependencies (Config: en_ewt)

• Base Model: bert-base-uncased

• Hyperparameters:

  • head_lr: 1e-3

  • body_lr: 2e-5

  • seed: 42

• Tokenization & Tagging Strategy:

  • 17 UPOS tags

  • First-subword labeling

  • -100 mask for continuation and special tokens

Metrics

{
  "loss": 0.20879080891609192,
  "token_accuracy": 0.9718864323206616,
  "macro_f1": 0.9194535916196505,
  "precision": 0.9664863944129796,
  "recall": 0.9677226794013559,
  "f1": 0.9671041418101243,
  "runtime": 4.9848,
  "samples_per_second": 401.418,
  "steps_per_second": 3.21,
  "epoch": 3.0,
  "method": "full",
  "trainable_params": 108904721,
  "total_train_seconds": 309.63132905960083
}

Headline Metrics: Token Accuracy = 0.9719 | Macro-F1 = 0.9195

Example Usage

from transformers import pipeline

# 1. Load the POS tagger pipeline
pos_tagger = pipeline(
    task="token-classification",
    model="Elpapudex/Pos-full",
    aggregation_strategy="simple"
)

# 2. Input sentence to analyze
text = "The quick brown fox jumps gracefully over the lazy dog."

# 3. Predict POS tags
results = pos_tagger(text)

# 4. Display results
for tag in results:
    print(f"Word: {tag['word']:<12} | Tag: {tag['entity_group']:<9} | Confidence: {tag['score']:.2%}")

Limitations

• English-only.

• Trained on UD English EWT (English Web Treebank).

• Predicts only the 17 Universal POS (UPOS) tags.

• May not generalize well to other domains, genres, or tokenizers.

• Note: The seqeval precision, recall, and F1 values included in the JSON are sequence evaluation metrics and do not represent standard entity-level metrics for POS tagging.

References

• Devlin et al., BERT, arXiv:1810.04805 §5.3

• Hugging Face Transformers Training Documentation

• Hugging Face Universal Dependencies Dataset