andgonzalez/bert-uncased-slot-filling

Model Card for BERT Slot Filling Model

This BERT-based model is designed for slot filling tasks in natural language sentences, ideal for extracting specific information in applications like chatbots and virtual assistants.

For example: input: Transfer $500 from checking to student savings output: transfer [$500:B-amount] from [checking:B-account-from] to [student:B-account-to] [savings:I-account-to]

The model was trained with dataset https://github.com/SunLemuria/JointBERT-Tensorflow1/blob/master/data/

Please make yourself familiar with BERT: BERT for Joint Intent Classification and Slot Filling https://arxiv.org/pdf/1902.10909.pdf

Model Details

Tag	Definition
B-account-from	Start of the source account in a transaction.
I-account-from	Complement of the source account in a transaction.
B-account-to	Start of the target account in a transaction.
I-account-to	Complement of the target account in a transaction.
B-bill_type	Start of the type of bill or service.
I-bill_type	Complement of the type of bill or service.
B-transaction-from	Start of the origin of a transaction or fraud.
I-transaction-from	Complement of the origin of a transaction or fraud.
B-transaction-to	Start of the destination or end of a transaction or fraud.
I-transaction-to	Complement of the destination of a transaction or fraud.
B-amount	Start of a specified amount of money.
I-amount	Complement of a specified amount of money.
B-timeRange	Start of a specific time range or date.
I-timeRange	Complement of a specific time range or date.

Model Description

• Developed by: Andy González
• Model type: Token Classification
• Language(s) (NLP): English
• Finetuned from model [optional]: bert-base-uncased

How to Get Started with the Model

!pip install torch transformers

import os
import requests
import torch
from transformers import BertForTokenClassification, BertTokenizerFast

# URL y archivo para los slots
slots_url = 'https://huggingface.co/andgonzalez/bert-uncased-slot-filling/raw/main/slots.txt'
slots_file = 'slots.txt'
device = "cpu"

# Descargar y guardar los slots si no existen
if not os.path.exists(slots_file):
    response = requests.get(slots_url)
    response.raise_for_status()
    with open(slots_file, 'w') as file:
        file.write(response.text)

# Leer los slots
with open(slots_file, 'r') as file:
    slot_labels = file.read().splitlines()

# Cargar el tokenizador y el modelo
tokenizer = BertTokenizerFast.from_pretrained('andgonzalez/bert-uncased-slot-filling')
model = BertForTokenClassification.from_pretrained('andgonzalez/bert-uncased-slot-filling')

# Ejemplo 
sentence = "Transfer $500 from checking to student savings"

inputs = tokenizer(sentence, truncation=True, padding='max_length', max_length=20, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    model.eval()
    outputs = model(**inputs)

# Procesar los logits para obtener predicciones
logits = outputs.logits
predictions = torch.argmax(logits, dim=2).squeeze().cpu().numpy()
words = tokenizer.convert_ids_to_tokens(inputs["input_ids"].squeeze().cpu().numpy())

# Inicializar skip_next
skip_next = False

# Formatear la oracion
formatted_sentence = []
for i, (word, pred) in enumerate(zip(words, predictions)):
    if word not in ['[PAD]', '[SEP]', '[CLS]']:
        label = slot_labels[pred]

        if word == "$" and i + 1 < len(words) and words[i + 1].replace("##", "").isdigit():
            next_word = words[i + 1].replace("##", "")
            combined_word = word + next_word
            formatted_word = f'[{combined_word}:{label}]'
            formatted_sentence.append(formatted_word)
            skip_next = True
        elif skip_next:
            skip_next = False
            continue
        elif not word.startswith("##"):
            if label != 'O':
                formatted_word = f'[{word}:{label}]'
            else:
                formatted_word = word
            formatted_sentence.append(formatted_word)

formatted_sentence = ' '.join(formatted_sentence)
print(formatted_sentence)

Training Details

Metrics

• Metrics used: Precision, Recall, F1-Score

Results

• Epoch 1: Loss: 1.3253, Precision: 0.5862, Recall: 0.5758, F1-Score: 0.5633
• Epoch 2: Loss: 0.3507, Precision: 0.7491, Recall: 0.7476, F1-Score: 0.7374
• Epoch 3: Loss: 0.2156, Precision: 0.8180, Recall: 0.8138, F1-Score: 0.8007
• Epoch 4: Loss: 0.1593, Precision: 0.8252, Recall: 0.8274, F1-Score: 0.8173
• Epoch 5: Loss: 0.1236, Precision: 0.8613, Recall: 0.8549, F1-Score: 0.8466
• Epoch 6: Loss: 0.0961, Precision: 0.8839, Recall: 0.8810, F1-Score: 0.8786
• Epoch 7: Loss: 0.0787, Precision: 0.8795, Recall: 0.8917, F1-Score: 0.8808
• Epoch 8: Loss: 0.0644, Precision: 0.8956, Recall: 0.8958, F1-Score: 0.8911
• Epoch 9: Loss: 0.0542, Precision: 0.8889, Recall: 0.9012, F1-Score: 0.8913
• Epoch 10: Loss: 0.0468, Precision: 0.8980, Recall: 0.9007, F1-Score: 0.8935
• Best Model: Epoch 8, Test Loss: 0.1588

Plots