Neural network model architecture and inference

model = Sequential([
    Dense(256, input_shape=(len(train_x[0]),), activation='relu'),
    Dropout(0.5),
    Dense(128, activation='relu'),
    Dropout(0.4),
    Dense(64, activation='relu'),
    Dropout(0.3),
    Dense(len(train_y[0]), activation='softmax'),
])

sgd = SGD(learning_rate=0.005, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
model.fit(train_x, train_y, epochs=300, batch_size=8, verbose=1)

def _remove_accents(text: str) -> str:
    return ''.join(
        c for c in unicodedata.normalize('NFD', text)
        if unicodedata.category(c) != 'Mn'
    )

def _normalize(text: str) -> str:
    text = text.lower().strip()
    text = _remove_accents(text)
    text = re.sub(r'[^\w\s]', '', text)
    text = re.sub(r'\s+', ' ', text)
    return text

def clean_up_sentence(sentence: str) -> list[str]:
    tokens = nltk.word_tokenize(_normalize(sentence))
    return [lemmatizer.lemmatize(w) for w in tokens]

def bag_of_words(sentence: str) -> np.ndarray:
    tokens = clean_up_sentence(sentence)
    bag = [0] * len(words)
    for token in tokens:
        for i, word in enumerate(words):
            if word == token:
                bag[i] = 1
    return np.array(bag)

_THRESHOLD_HIGH = 0.60
_THRESHOLD_LOW  = 0.35