def predict(sentence):
n_test = 1
test_vec_seq = []
test_seq = []
for i in range(n_test):
test_text = sentence
test_seq.append(test_text)
test_doc = nlp(test_text)
test_vec = []
for word in test_doc:
test_vec.append(word.vector)
test_vec_seq.append(test_vec)
test_vec_seq = pad_vec_sequences(test_vec_seq)
prediction = model.predict(test_vec_seq)
label_predictions = np.zeros(prediction.shape)
for i in range(n_test):
m = max(prediction[i])
print(m)
p = np.where(prediction[i] > 0.55 * m)
q = np.where(prediction[i] == m) and np.where(prediction[i] > 0.34)
label_predictions[i][p] = 1
label_predictions[i][q] = 2
#print (p,q)
for i in range(n_test):
for x in range(len(label_predictions[i])):
if label_predictions[i][x] == 2:
return sentence, labels[x]
if len(set(label_predictions[i])) == 1:
return sentence, "NO INTENT"
return sentence, "NO INTENT"
def make_data_set():
ds = Dataset()
print("Datasets loaded.")
X_all = pad_vec_sequences(ds.X_all_vec_seq)
Y_all = ds.Y_all
#print (X_all.shape)
x_train, x_test, y_train, y_test = model_selection.train_test_split(X_all,Y_all,test_size=0.2)
y_train = pad_class_sequence(y_train, nb_classes)
y_test = pad_class_sequence(y_test, nb_classes)
y_test = np.array(y_test)
x_train = np.asarray(x_train)
x_train.ravel()
y_train = np.asarray(y_train)
y_train.ravel()
return x_train,y_train,x_test,y_test
test_ent_seq = [] #list of lists of entities in each test query
test_seq = [] #list of all test queries
for i in range(n_test):
print("\nEnter Query", i + 1, " to be classified:")
test_text = input()
test_seq.append(test_text)
#vectorize text.
test_doc = nlp(test_text)
test_vec = []
for word in test_doc:
test_vec.append(word.vector)
test_vec_seq.append(test_vec)
test_ent_seq.append(test_doc.ents)
#convert all the sentences into matrices of equal size.
test_vec_seq = pad_vec_sequences(test_vec_seq)
#get predictions
prediction = model.predict(test_vec_seq)
label_predictions = np.zeros(prediction.shape)
for i in range(n_test):
m = max(prediction[i])
p = np.where(prediction[i] > 0.55 * m) # p collects possible sub intents
q = np.where(prediction[i] == m) #q collects intent
label_predictions[i][p] = 1
label_predictions[i][q] = 2
'''
label_predictions[i] is now a list of intent predictions for Query i, where:
One of the following integer values is saved per intent label:
from keras.preprocessing import sequence
from keras.models import Model
from keras.layers import Dense, Dropout, Embedding, Input, merge
from keras.layers.recurrent import LSTM, GRU
from keras.layers import Concatenate
from keras.utils import np_utils, generic_utils
from keras import optimizers, metrics
maxlen = 50 #sentences with length > maxlen will be ignored
hidden_dim = 32
nb_classes = len(labels)
ds = Dataset()
print("Datasets loaded.")
X_all = pad_vec_sequences(ds.X_all_vec_seq)
Y_all = ds.Y_all
#print (X_all.shape)
x_train, x_test, y_train, y_test = model_selection.train_test_split(
X_all, Y_all, test_size=0.2)
y_train = pad_class_sequence(y_train, nb_classes)
y_test = pad_class_sequence(y_test, nb_classes)
#THE MODEL
sequence = Input(shape=(maxlen, 384), dtype='float32', name='input')
#forwards lstm
forwards = LSTM(hidden_dim,
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
dropout=0.1,