pollitoconpapass/intent_classification_model

1---2language:3- es4metrics:5- accuracy6library_name: keras7tags:8- code9---10# Model to detect Chat Intention11 12This model was trained for academic purposes to detect the intention of the user while chatting with a bot13 14## Classes15As part of a college project about a HealthCare Org Chatbot the classes are: 16- 0: Normal Conversation17- 1: Patient Information18- 2: Administrative Questions19 20IMPORTANT: The model was trained with Spanish Sentences21 22## Accuracy23We ended up with a 0.85 percent of accuracy. 24 25```sh26Classification Report:27                          precision    recall  f1-score   support28 29     Normal conversation       0.87      0.82      0.85        4030     Patient information       0.83      0.85      0.84        4031Administrative questions       0.85      0.88      0.86        4032 33                accuracy                           0.85       12034               macro avg       0.85      0.85      0.85       12035            weighted avg       0.85      0.85      0.85       12036 37```38 39 40## How to use it? 41Use the following script: 42```py43 44import json45import numpy as np46import tensorflow as tf47from huggingface_hub import hf_hub_download48from tensorflow.keras.preprocessing.text import tokenizer_from_json49from tensorflow.keras.preprocessing.sequence import pad_sequences50 51repo_id = "pollitoconpapass/intent_classification_model"52tokenizer_path = hf_hub_download(repo_id=repo_id, filename="tokenizer.json")53 54# with open(tokenizer_path, 'r', encoding='utf-8') as f:55#     loaded_tokenizer_config = json.load(f)56#     loaded_tokenizer = tokenizer_from_json(loaded_tokenizer_config)57 58with open(tokenizer_path, 'r', encoding='utf-8') as f:59    loaded_tokenizer_config = json.load(f)60    loaded_max_len = loaded_tokenizer_config['config']['max_len']61 62    del loaded_tokenizer_config['config']['max_len']63    loaded_tokenizer = tokenizer_from_json(json.dumps(loaded_tokenizer_config))64 65model_file_path = hf_hub_download(repo_id=repo_id, filename="intent_classification_model.keras")66loaded_model = tf.keras.models.load_model(model_file_path)67 68INTENT_MAP = {69    0: "Normal conversation",70    1: "Patient information",71    2: "Administrative questions"72}73 74def predict_single_sentence(sentence, max_len) -> tuple[str, float]:75    # Preprocess the whole sentence76    sequence = loaded_tokenizer.texts_to_sequences([sentence])77    # Use the loaded_max_len for padding78    padded_sequence = pad_sequences(sequence, maxlen=loaded_max_len, padding='post')79 80    prediction = loaded_model.predict(padded_sequence, verbose=0)[0] # -> get 1st prediction81 82    # Prediction + confidence83    predicted_class = np.argmax(prediction)84    confidence = prediction[predicted_class] * 10085 86    intent = INTENT_MAP[predicted_class]87    return intent, confidence88 89 90sentence = "Holaaaa"91intent, confidence = predict_single_sentence(sentence, 10)92print(f"Intent: {intent} (Confidence: {confidence:.2f}%)")93```