The Traveleh app analysis

Building the model is like below 

def build_smallBERT_CNN_classifier_model():

    text_input = tf.keras.layers.Input(shape=(), dtype=tf.string, name='text')

    preprocessing_layer = hub.KerasLayer(bert_preprocessor, name='preprocessing')

    encoder_inputs = preprocessing_layer(text_input)

    encoder = hub.KerasLayer(bert_encoder_model, trainable=True, name='BERT_encoder')

    outputs = encoder(encoder_inputs)

    net = sequence_output = outputs["sequence_output"]

    net = tf.keras.layers.Conv1D(32, 2, activation='relu')(net)

    net = tf.keras.layers.Conv1D(64, 2, activation='relu')(net)

    net = tf.keras.layers.MaxPooling1D(pool_size=2)(net)

    

    net = tf.keras.layers.GlobalMaxPool1D()(net)

    

    net = tf.keras.layers.Dense(512, activation="relu")(net)

    net = tf.keras.layers.Dropout(0.1)(net)

    

    net = tf.keras.layers.Dense(num_classes, activation="softmax", name='classifier')(net)

    

    return tf.keras.Model(text_input, net)

Compiling is like below 

loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False)

metrics = tf.metrics.CategoricalAccuracy()

epochs = 15

optimizer=tf.keras.optimizers.Adam(1e-5)

# build the model

classifier_smallBERT_model_cnn = build_smallBERT_CNN_classifier_model()

# compile the model

classifier_smallBERT_model_cnn.compile(optimizer=optimizer,

                         loss=loss,

                         metrics=metrics)

Optional: Enhance with Callbacks:

class MyCallback(tf.keras.callbacks.Callback):

    def on_epoch_end(self, epoch, logs=None):

        if logs['val_categorical_accuracy'] >= 0.94:

            print("\nValidation accuracy reached 94%!")

            self.model.stop_training = True

Training the Model:

callback = MyCallback()

history = classifier_smallBERT_model_cnn.fit(x=X_train, y=y_train_one_hot,

                                             validation_data=(X_val, y_val_one_hot),

                                             batch_size=32,

                                             epochs=epochs,

                                             callbacks=[callback])

Testing the model

intent_mapping = df.groupby('intent_encoding')['intent'].first().to_dict()

user_utterance = "what should i prepare before i go"

user_utterance = np.array([user_utterance])

# Make the prediction

predictions = loaded_model.predict(user_utterance)

predicted_intent_index = np.argmax(predictions)

# Map the predicted intent index to intent label and response

predicted_intent_label = intent_mapping[str(predicted_intent_index)]

predicted_response = df.loc[df['intent'] == predicted_intent_label, 'response'].iloc[0]

print("Predicted Intent:", predicted_intent_label)

print("VELA:", predicted_response)

In this project, the spaCy Custom NER model to accurately identify and classify that.

def offsetter(lbl, doc, matchitem):

    one = len(str(doc[0:matchitem[1]]))

    subdoc = doc[matchitem[1]:matchitem[2]]

    two = one + len(str(subdoc))

    

    if one != 0:

        one += 1

        two += 1

    return (one, two, lbl)

def spacify_row(document, label, entity_keywords):

    matcher = PhraseMatcher(nlp.vocab)

    for i in entity_keywords:

        matcher.add(label, None, nlp(i))

    nlp_document = nlp(document)

    matches = matcher(nlp_document)

    

    entity_list = [offsetter(label, nlp_document, match) for match in matches]

    

    return (document, {'entities': entit

spacify_row('I love to eat seafood', 'FOOD_PREFERENCE', 

            entity_keywords = entity_phrases.get('FOOD_PREFERENCE'))

References:

https://medium.com/@reynxzz/building-a-smart-chatbot-with-intent-classification-and-named-entity-recognition-travelah-a-case-ee905f23fd9d