bilstm-crf-fasttext-charcnn NER CoNLL-2002 (bs=32)
BiLSTM + CRF+CharCNN con embeddings FastText español (cc.es.300).
Entrenado sobre CoNLL-2002 español. Modelo CUSTOM con trust_remote_code=True.
Métricas en test (CoNLL-2002)
| Metric | Valor |
|---|---|
| F1 | 0.8040 |
| Precision | 0.8130 |
| Recall | 0.7952 |
Uso
# Requiere: pip install pytorch-crf
from transformers import AutoModelForTokenClassification, AutoConfig
import json
config = AutoConfig.from_pretrained("cvalenciaunivalle/bilstm-crf-fasttext-charcnn-conll-bs32", trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained("cvalenciaunivalle/bilstm-crf-fasttext-charcnn-conll-bs32", trust_remote_code=True)
# Cargar vocab (incluye word2idx, char2idx, id2tag)
from huggingface_hub import hf_hub_download
vocab_path = hf_hub_download("cvalenciaunivalle/bilstm-crf-fasttext-charcnn-conll-bs32", "vocab.json")
with open(vocab_path) as f:
vocab = json.load(f)
# Predecir
oraciones_tokenizadas = [["Juan", "vive", "en", "Bogotá", "."]]
tags = model.predict(oraciones_tokenizadas, vocab)
print(tags) # [['B-PER', 'O', 'O', 'B-LOC', 'O']]
Detalles
- Arquitectura: BiLSTM hidden=200, dropout=0.3, CRF de salida
- Word embeddings: 300d (FastText)
- CharCNN: 30d char emb -> Conv1d(filters=50, kernel=3) -> max-pool
- Batch size: 32, épocas: 10
- Compañeros: Laura Valero, Cristian Valencia
- Taller PLN — Maestría Univalle