generation_script.py

#This was generated using the train_spacy.txt and valida_spacy.txt files. 
# We have replace all the lbels by B-PROFESION or I-PROFESION 
# Validation and training set half of the sentences with labels and the other ones without them.

from collections import defaultdict
import random

def contiene_b(frase):
    return any(label.startswith("B-") for _, label in frase)

def procesar_training_set_balanceado(archivo_entrada, archivo_salida, archivo_ids):
    frases = defaultdict(list)

    with open(archivo_entrada, encoding="utf-8") as f:
        for linea in f:
            if linea.strip():
                partes = linea.strip().split()
                if len(partes) == 5:
                    token, doc_id, _, _, label = partes
                    frases[doc_id].append((token, label))

    con_b_ids = [id_ for id_, frase in frases.items() if contiene_b(frase)]
    sin_b_ids = [id_ for id_, frase in frases.items() if not contiene_b(frase)]

    # Queremos aproximadamente la mitad con B
    n = min(len(con_b_ids), len(sin_b_ids))
    selected_ids = con_b_ids[:n] + sin_b_ids[:n]
    random.shuffle(selected_ids)

    # Guardar los IDs seleccionados
    with open(archivo_ids, "w") as f_ids:
        for id_ in selected_ids:
            f_ids.write(f"{id_}\n")

    # Guardar el archivo con formato CoNLL
    with open(archivo_salida, "w", encoding="utf-8") as out:
        for id_ in selected_ids:
            for token, label in frases[id_]:
                out.write(f"{token} {label}\n")
            out.write("\n")

def procesar_dev_test_balanceado(archivo_entrada, archivo_salida_dev, archivo_salida_test, archivo_ids_dev, archivo_ids_test):
    frases = defaultdict(list)

    with open(archivo_entrada, encoding="utf-8") as f:
        for linea in f:
            if linea.strip():
                partes = linea.strip().split()
                if len(partes) == 5:
                    token, doc_id, _, _, label = partes
                    frases[doc_id].append((token, label))

    # Separar documentos con y sin B
    con_b_ids = [id_ for id_, frase in frases.items() if contiene_b(frase)]
    sin_b_ids = [id_ for id_, frase in frases.items() if not contiene_b(frase)]

    # Balancear la división
    random.shuffle(con_b_ids)
    mitad_b = len(con_b_ids) // 2
    dev_ids_b = con_b_ids[:mitad_b]
    test_ids_b = con_b_ids[mitad_b:]

    random.shuffle(sin_b_ids)
    mitad_sin = len(sin_b_ids) // 2
    dev_ids_sin = sin_b_ids[:mitad_sin]
    test_ids_sin = sin_b_ids[mitad_sin:]

    dev_ids = dev_ids_b + dev_ids_sin
    test_ids = test_ids_b + test_ids_sin
    random.shuffle(dev_ids)
    random.shuffle(test_ids)

    # Guardar los IDs
    with open(archivo_ids_dev, "w") as f_dev, open(archivo_ids_test, "w") as f_test:
        for id_ in dev_ids:
            f_dev.write(f"{id_}\n")
        for id_ in test_ids:
            f_test.write(f"{id_}\n")

    # Escribir los archivos
    def escribir(ids, archivo_salida):
        with open(archivo_salida, "w", encoding="utf-8") as out:
            for id_ in ids:
                for token, label in frases[id_]:
                    out.write(f"{token} {label}\n")
                out.write("\n")

    escribir(dev_ids, archivo_salida_dev)
    escribir(test_ids, archivo_salida_test)

# 🛠️ Uso
procesar_training_set_balanceado("train_spacy.txt", "train_conll.txt","train_ids.txt")
procesar_dev_test_balanceado(
    "valid_spacy.txt", 
    "dev_conll.txt", 
    "test_conll.txt", 
    "dev_ids.txt", 
    "test_ids.txt"
)


from datasets import load_dataset, ClassLabel, DatasetDict

from datasets import load_dataset, Dataset, DatasetDict
from collections import defaultdict

def normalizar_etiqueta(label):
    if label.startswith("B-"):
        return "B-PROFESION"
    elif label.startswith("I-"):
        return "I-PROFESION"
    return label

def cargar_y_preparar_conll(paths):
    def parse_conll_dataset(file_path):
        raw = load_dataset("text", data_files=file_path)["train"]

        tokens = []
        ner_tags = []
        current_tokens = []
        current_tags = []

        for example in raw:
            line = example["text"]
            if not line.strip():
                if current_tokens:
                    tokens.append(current_tokens)
                    ner_tags.append(current_tags)
                    current_tokens = []
                    current_tags = []
            else:
                token, tag = line.strip().split()
                current_tokens.append(token)
                current_tags.append(normalizar_etiqueta(tag))

        return {"tokens": tokens, "ner_tags": ner_tags}

    # Cargar y procesar cada división
    parsed = {split: parse_conll_dataset(path) for split, path in paths.items()}

    # Generar label2id
    all_labels = set(tag for split_data in parsed.values() for seq in split_data["ner_tags"] for tag in seq)
    label_list = sorted(all_labels)
    label2id = {label: i for i, label in enumerate(label_list)}
    id2label = {i: label for label, i in label2id.items()}

    def tag_ids(ner_tags):
        return [[label2id[tag] for tag in seq] for seq in ner_tags]

    dataset = DatasetDict({
        split: Dataset.from_dict({
            "tokens": parsed_data["tokens"],
            "ner_tags": tag_ids(parsed_data["ner_tags"])
        })
        for split, parsed_data in parsed.items()
    })

    return dataset, label2id, id2label

paths = {
    "train": "train_conll.txt",
    "validation": "dev_conll.txt",
    "test": "test_conll.txt"
}

dataset, label2id, id2label = cargar_y_preparar_conll(paths)


from datasets import Features, Sequence, ClassLabel, Value

# Asumimos que label_list ya está definido (por ejemplo: ['B-PROFESION', 'I-PROFESION', 'O'])
ner_feature = Sequence(ClassLabel(names=list(label2id.keys())))
features = Features({
    "tokens": Sequence(Value("string")),
    "ner_tags": ner_feature
})

# Aplicar a cada split del dataset
for split in dataset:
    dataset[split] = dataset[split].cast(features)