Natural Language Processing has many interesting applications and Sequence to Sequence modelling is one of those interesting applications. It has major applications in question-answering systems and language translation systems. Sequence-to-Sequence (Seq2Seq) modelling is about training the models that can convert sequences from one domain to sequences of another domain, for example, English to French. This Seq2Seq modelling is performed by the LSTM encoder and decoder. We can guess this process from the below illustration.
(Image Source: blog.keras.io)
In this article, we will implement deep learning in the Sequence-to-Sequence (Seq2Seq) modelling for language translation. This approach will be applied to convert the short English sentences into the corresponding French sentences. The LSTM encoder and decoder are used to process the sequence to sequence modelling in this task.
In this experiment, the data set is taken from Kaggle that is publically available as French-English Bilingual Pairs. This dataset contains pairs of French sentences as well as their English translation.
Implementation of Sequence-to-Sequence (Seq2Seq) Modelling
First of all, we will import the required libraries. This program was executed in Google Colab with hosted runtime. If you are working on your local system, make sure to install TensorFlow before executing this program.
#Importing library import numpy as np from keras.models import Model from keras.layers import Input, LSTM, Dense from keras.utils import * from keras.initializers import * import tensorflow as tf import time, random
After importing the libraries, we will specify the value for the hyperparameters including the batch size for training, latent dimensionality for the encoding space and the number of samples to train on.
#Hyperparameters batch_size = 64 latent_dim = 256 num_samples = 10000
The below lines of codes will perform the data vectorization where we will read the file containing the English and corresponding French sentences. In the vectorization process, the collection of text documents is converted into feature vectors.
#Vectorize the data. input_texts =  target_texts =  input_chars = set() target_chars = set() with open('fra.txt', 'r', encoding='utf-8') as f: lines = f.read().split('\n') for line in lines[: min(num_samples, len(lines) - 1)]: input_text, target_text = line.split('\t') target_text = '\t' + target_text + '\n' input_texts.append(input_text) target_texts.append(target_text) for char in input_text: if char not in input_chars: input_chars.add(char) for char in target_text: if char not in target_chars: target_chars.add(char) input_chars = sorted(list(input_chars)) target_chars = sorted(list(target_chars)) num_encoder_tokens = len(input_chars) num_decoder_tokens = len(target_chars) max_encoder_seq_length = max([len(txt) for txt in input_texts]) max_decoder_seq_length = max([len(txt) for txt in target_texts]) #Print size print('Number of samples:', len(input_texts)) print('Number of unique input tokens:', num_encoder_tokens) print('Number of unique output tokens:', num_decoder_tokens) print('Max sequence length for inputs:', max_encoder_seq_length) print('Max sequence length for outputs:', max_decoder_seq_length)
After getting the data set with all features, we will define the input data encoder and decoder and the target data for the decoder.
#Define data for encoder and decoder input_token_id = dict([(char, i) for i, char in enumerate(input_chars)]) target_token_id = dict([(char, i) for i, char in enumerate(target_chars)]) encoder_in_data = np.zeros((len(input_texts), max_encoder_seq_length, num_encoder_tokens), dtype='float32') decoder_in_data = np.zeros((len(input_texts), max_decoder_seq_length, num_decoder_tokens), dtype='float32') decoder_target_data = np.zeros((len(input_texts), max_decoder_seq_length, num_decoder_tokens), dtype='float32') for i, (input_text, target_text) in enumerate(zip(input_texts, target_texts)): for t, char in enumerate(input_text): encoder_in_data[i, t, input_token_id[char]] = 1. for t, char in enumerate(target_text): decoder_in_data[i, t, target_token_id[char]] = 1. if t > 0: decoder_target_data[i, t - 1, target_token_id[char]] = 1.
Below lines of codes will define the input sequence for the encoder defined above and process this sequence. After that, an initial state will be set up for the decoder using ‘encoder_states’.
#Define and process the input sequence encoder_inputs = Input(shape=(None, num_encoder_tokens)) encoder = LSTM(latent_dim, return_state=True) encoder_outputs, state_h, state_c = encoder(encoder_inputs) #We discard `encoder_outputs` and only keep the states. encoder_states = [state_h, state_c] #Using `encoder_states` set up the decoder as initial state. decoder_inputs = Input(shape=(None, num_decoder_tokens)) decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True) decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states) decoder_dense = Dense(num_decoder_tokens, activation='softmax') decoder_outputs = decoder_dense(decoder_outputs)
The below line of code will define the final model that will turn `encoder_in_data` & `decoder_in_data` into `decoder_target_data`.
#Final model model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
After defining the final model, we will check it by its summary, data shape and a visualization.
#Model Summary model.summary() #Model data Shape print("encoder_in_data shape:",encoder_in_data.shape) print("decoder_in_data shape:",decoder_in_data.shape) print("decoder_target_data shape:",decoder_target_data.shape) #Visuaize the model plot_model(model,show_shapes=True)
Once we are ready with the final model, we will compile and train the model. Here, the model will be trained in 50 epochs only. For more accuracy, you can perform this for more number of epochs.
#Compiling and training the model model.compile(optimizer=Adam(lr=0.01, beta_1=0.9, beta_2=0.999, decay=0.001), loss='categorical_crossentropy') model.fit([encoder_in_data, decoder_in_data], decoder_target_data, batch_size = batch_size, epochs=50, validation_split=0.2)
After successful training, we will define the sample model using the parameters of the trained model to test language translation.
#Define sampling models encoder_model = Model(encoder_inputs, encoder_states) decoder_state_input_h = Input(shape=(latent_dim,)) decoder_state_input_c = Input(shape=(latent_dim,)) decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c] decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs) decoder_states = [state_h, state_c] decoder_outputs = decoder_dense(decoder_outputs) decoder_model = Model([decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)
The below part of codes will define the decode sequence for the text that we will pass to the model as the input sequence. This could be understood as the module for translating the input language into the target language. In this part, the input sequence is encoded into the state vectors. The state vector and the target sequence is passed to the decoder and it produces the prediction for the next character. Using these predictions, the next character is sampled and it is appended to the target sequence. This process is repeated to generate until the end of the sequence.
reverse_input_char_index = dict((i, char) for char, i in input_token_id.items()) reverse_target_char_index = dict((i, char) for char, i in target_token_id.items()) #Define Decode Sequence def decode_sequence(input_seq): #Encode the input as state vectors. states_value = encoder_model.predict(input_seq) #Generate empty target sequence of length 1. target_seq = np.zeros((1, 1, num_decoder_tokens)) #Get the first character of target sequence with the start character. target_seq[0, 0, target_token_id['\t']] = 1. #Sampling loop for a batch of sequences #(to simplify, here we assume a batch of size 1). stop_condition = False decoded_sentence = '' while not stop_condition: output_tokens, h, c = decoder_model.predict([target_seq] + states_value) #Sample a token sampled_token_index = np.argmax(output_tokens[0, -1, :]) sampled_char = reverse_target_char_index[sampled_token_index] decoded_sentence += sampled_char #Exit condition: either hit max length #or find stop character. if (sampled_char == '\n' or len(decoded_sentence) > max_decoder_seq_length): stop_condition = True #Update the target sequence (of length 1). target_seq = np.zeros((1, 1, num_decoder_tokens)) target_seq[0, 0, sampled_token_index] = 1. #Update states states_value = [h, c] return decoded_sentence
Finally, we will check our model to decode the input sequence into the target sequence, i.e., translate the English sentences into the French sentences.
for seq_index in range(10): input_seq = encoder_in_data[seq_index: seq_index + 1] decoded_sentence = decode_sequence(input_seq) print('-') print('Input sentence:', input_texts[seq_index]) print('Decoded sentence:', decoded_sentence)
As we can see in the above output, our model has correctly converted the 10 English sentences to the equivalent French sentences. We checked our model to convert 10 input sequences. You can check this on more sequences, say 100 and if you find few incorrect conversions, then tune the hyperparameters and train the model for more number of epochs.
- Ilya Sutskever et al, ‘Sequence to Sequence Learning with Neural Networks’, arXiv.org.
- K Cho et al, ‘Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation’, arXiv.org.
- Keras tutorial on ‘Sequence to sequence example in Keras (character-level)’.
- Keras tutorial on ‘A ten-minute introduction to sequence-to-sequence learning in Keras’
Join Our Discord Server. Be part of an engaging online community. Join Here.
Subscribe to our NewsletterGet the latest updates and relevant offers by sharing your email.
Vaibhav Kumar has experience in the field of Data Science and Machine Learning, including research and development. He holds a PhD degree in which he has worked in the area of Deep Learning for Stock Market Prediction. He has published/presented more than 15 research papers in international journals and conferences. He has an interest in writing articles related to data science, machine learning and artificial intelligence.