How to use a trained model for inference on new input phrases and evaluate the model skill.Tutorial Overview
This tutorial is divided into 4 parts; they are:
- German to English Translation Dataset
- Preparing the Text Data
- Train Neural Translation Model
- Evaluate Neural Translation Model
Python Environment
This tutorial assumes you have a Python 3 SciPy environment installed.
You must have Keras (2.0 or higher) installed with either the TensorFlow or Theano backend.
The tutorial also assumes you have NumPy and Matplotlib installed.
If you need help with your environment, see this post:
A GPU is not require for thus tutorial, nevertheless, you can access GPUs cheaply on Amazon Web Services.
German to English Translation Dataset
In this tutorial, we will use a dataset of German to English terms used as the basis for flashcards for language learning.
The dataset is available from the ManyThings.org website, with examples drawn from the Tatoeba Project. The dataset is comprised of German phrases and their English counterparts and is intended to be used with the Anki flashcard software.
The page provides a list of many language pairs, and I encourage you to explore other languages:
Note, the original dataset has changed which if used directly will break this tutorial and result in an error:
| ValueError: too many values to unpack (expected 2) |
As such you can download the original dataset in the correct format directly from here:
Download the dataset file to your current working directory.
You will have a file called deu.txt that contains 152,820 pairs of English to German phases, one pair per line with a tab separating the language.
For example, the first 5 lines of the file look as follows:
| Hi. Hallo! Hi. Grüß Gott! Run! Lauf! Wow! Potzdonner! Wow! Donnerwetter! |
We will frame the prediction problem as given a sequence of words in German as input, translate or predict the sequence of words in English.
The model we will develop will be suitable for some beginner German phrases.
Preparing the Text Data
The next step is to prepare the text data ready for modeling.
If you are new to cleaning text data, see this post:
Take a look at the raw data and note what you see that we might need to handle in a data cleaning operation.
For example, here are some observations I note from reviewing the raw data:
- There is punctuation.
- The text contains uppercase and lowercase.
- There are special characters in the German.
- There are duplicate phrases in English with different translations in German.
- The file is ordered by sentence length with very long sentences toward the end of the file.
Did you note anything else that could be important?
Let me know in the comments below.
A good text cleaning procedure may handle some or all of these observations.
Data preparation is divided into two subsections:
- Clean Text
- Split Text
1. Clean Text
First, we must load the data in a way that preserves the Unicode German characters. The function below called load_doc() will load the file as a blob of text.
| # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, mode='rt', encoding='utf-8') # read all text text = file.read() # close the file file.close() return text |
Each line contains a single pair of phrases, first English and then German, separated by a tab character.
We must split the loaded text by line and then by phrase. The function to_pairs() below will split the loaded text.
| # split a loaded document into sentences def to_pairs(doc): lines = doc.strip().split('\n') pairs = [line.split('\t') for line in lines] return pairs |
We are now ready to clean each sentence. The specific cleaning operations we will perform are as follows:
- Remove all non-printable characters.
- Remove all punctuation characters.
- Normalize all Unicode characters to ASCII (e.g. Latin characters).
- Normalize the case to lowercase.
- Remove any remaining tokens that are not alphabetic.
We will perform these operations on each phrase for each pair in the loaded dataset.
The clean_pairs() function below implements these operations.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 | # clean a list of lines def clean_pairs(lines): cleaned = list() # prepare regex for char filtering re_print = re.compile('[^%s]' % re.escape(string.printable)) # prepare translation table for removing punctuation table = str.maketrans('', '', string.punctuation) for pair in lines: clean_pair = list() for line in pair: # normalize unicode characters line = normalize('NFD', line).encode('ascii', 'ignore') line = line.decode('UTF-8') # tokenize on white space line = line.split() # convert to lowercase line = [word.lower() for word in line] # remove punctuation from each token line = [word.translate(table) for word in line] # remove non-printable chars form each token line = [re_print.sub('', w) for w in line] # remove tokens with numbers in them line = [word for word in line if word.isalpha()] # store as string clean_pair.append(' '.join(line)) cleaned.append(clean_pair) return array(cleaned) |
Finally, now that the data has been cleaned, we can save the list of phrase pairs to a file ready for use.
The function save_clean_data() uses the pickle API to save the list of clean text to file.
Pulling all of this together, the complete example is listed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 | import string import re from pickle import dump from unicodedata import normalize from numpy import array # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, mode='rt', encoding='utf-8') # read all text text = file.read() # close the file file.close() return text # split a loaded document into sentences def to_pairs(doc): lines = doc.strip().split('\n') pairs = [line.split('\t') for line in lines] return pairs # clean a list of lines def clean_pairs(lines): cleaned = list() # prepare regex for char filtering re_print = re.compile('[^%s]' % re.escape(string.printable)) # prepare translation table for removing punctuation table = str.maketrans('', '', string.punctuation) for pair in lines: clean_pair = list() for line in pair: # normalize unicode characters line = normalize('NFD', line).encode('ascii', 'ignore') line = line.decode('UTF-8') # tokenize on white space line = line.split() # convert to lowercase line = [word.lower() for word in line] # remove punctuation from each token line = [word.translate(table) for word in line] # remove non-printable chars form each token line = [re_print.sub('', w) for w in line] # remove tokens with numbers in them line = [word for word in line if word.isalpha()] # store as string clean_pair.append(' '.join(line)) cleaned.append(clean_pair) return array(cleaned) # save a list of clean sentences to file def save_clean_data(sentences, filename): dump(sentences, open(filename, 'wb')) print('Saved: %s' % filename) # load dataset filename = 'deu.txt' doc = load_doc(filename) # split into english-german pairs pairs = to_pairs(doc) # clean sentences clean_pairs = clean_pairs(pairs) # save clean pairs to file save_clean_data(clean_pairs, 'english-german.pkl') # spot check for i in range(100): print('[%s] => [%s]' % (clean_pairs[i,0], clean_pairs[i,1])) |
Running the example creates a new file in the current working directory with the cleaned text called english-german.pkl.
Some examples of the clean text are printed for us to evaluate at the end of the run to confirm that the clean operations were performed as expected.
| [hi] => [hallo] [hi] => [gru gott] [run] => [lauf] [wow] => [potzdonner] [wow] => [donnerwetter] [fire] => [feuer] [help] => [hilfe] [help] => [zu hulf] [stop] => [stopp] [wait] => [warte] ... |
2. Split Text
The clean data contains a little over 150,000 phrase pairs and some of the pairs toward the end of the file are very long.
This is a good number of examples for developing a small translation model. The complexity of the model increases with the number of examples, length of phrases, and size of the vocabulary.
Although we have a good dataset for modeling translation, we will simplify the problem slightly to dramatically reduce the size of the model required, and in turn the training time required to fit the model.
You can explore developing a model on the fuller dataset as an extension; I would love to hear how you do.
We will simplify the problem by reducing the dataset to the first 10,000 examples in the file; these will be the shortest phrases in the dataset.
Further, we will then stake the first 9,000 of those as examples for training and the remaining 1,000 examples to test the fit model.
Below is the complete example of loading the clean data, splitting it, and saving the split portions of data to new files.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 | from pickle import load from pickle import dump from numpy.random import rand from numpy.random import shuffle # load a clean dataset def load_clean_sentences(filename): return load(open(filename, 'rb')) # save a list of clean sentences to file def save_clean_data(sentences, filename): dump(sentences, open(filename, 'wb')) print('Saved: %s' % filename) # load dataset raw_dataset = load_clean_sentences('english-german.pkl') # reduce dataset size n_sentences = 10000 dataset = raw_dataset[:n_sentences, :] # random shuffle shuffle(dataset) # split into train/test train, test = dataset[:9000], dataset[9000:] # save save_clean_data(dataset, 'english-german-both.pkl') save_clean_data(train, 'english-german-train.pkl') save_clean_data(test, 'english-german-test.pkl') |
Running the example creates three new files: the english-german-both.pkl that contains all of the train and test examples that we can use to define the parameters of the problem, such as max phrase lengths and the vocabulary, and the english-german-train.pkl and english-german-test.pkl files for the train and test dataset.
We are now ready to start developing our translation model.
Train Neural Translation Model
In this section, we will develop the neural translation model.
If you are new to neural translation models, see the post:
This involves both loading and preparing the clean text data ready for modeling and defining and training the model on the prepared data.
Let’s start off by loading the datasets so that we can prepare the data. The function below named load_clean_sentences() can be used to load the train, test, and both datasets in turn.
| # load a clean dataset def load_clean_sentences(filename): return load(open(filename, 'rb')) # load datasets dataset = load_clean_sentences('english-german-both.pkl') train = load_clean_sentences('english-german-train.pkl') test = load_clean_sentences('english-german-test.pkl') |
We will use the “both” or combination of the train and test datasets to define the maximum length and vocabulary of the problem.
This is for simplicity. Alternately, we could define these properties from the training dataset alone and truncate examples in the test set that are too long or have words that are out of the vocabulary.
We can use the Keras Tokenize class to map words to integers, as needed for modeling. We will use separate tokenizer for the English sequences and the German sequences. The function below-named create_tokenizer() will train a tokenizer on a list of phrases.
| # fit a tokenizer def create_tokenizer(lines): tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer |
Similarly, the function named max_length() below will find the length of the longest sequence in a list of phrases.
| # max sentence length def max_length(lines): return max(len(line.split()) for line in lines) |
We can call these functions with the combined dataset to prepare tokenizers, vocabulary sizes, and maximum lengths for both the English and German phrases.
| # prepare english tokenizer eng_tokenizer = create_tokenizer(dataset[:, 0]) eng_vocab_size = len(eng_tokenizer.word_index) + 1 eng_length = max_length(dataset[:, 0]) print('English Vocabulary Size: %d' % eng_vocab_size) print('English Max Length: %d' % (eng_length)) # prepare german tokenizer ger_tokenizer = create_tokenizer(dataset[:, 1]) ger_vocab_size = len(ger_tokenizer.word_index) + 1 ger_length = max_length(dataset[:, 1]) print('German Vocabulary Size: %d' % ger_vocab_size) print('German Max Length: %d' % (ger_length)) |
We are now ready to prepare the training dataset.
Each input and output sequence must be encoded to integers and padded to the maximum phrase length. This is because we will use a word embedding for the input sequences and one hot encode the output sequences The function below named encode_sequences() will perform these operations and return the result.
| # encode and pad sequences def encode_sequences(tokenizer, length, lines): # integer encode sequences X = tokenizer.texts_to_sequences(lines) # pad sequences with 0 values X = pad_sequences(X, maxlen=length, padding='post') return X |
The output sequence needs to be one-hot encoded. This is because the model will predict the probability of each word in the vocabulary as output.
The function encode_output() below will one-hot encode English output sequences.
| # one hot encode target sequence def encode_output(sequences, vocab_size): ylist = list() for sequence in sequences: encoded = to_categorical(sequence, num_classes=vocab_size) ylist.append(encoded) y = array(ylist) y = y.reshape(sequences.shape[0], sequences.shape[1], vocab_size) return y |
We can make use of these two functions and prepare both the train and test dataset ready for training the model.
| # prepare training data trainX = encode_sequences(ger_tokenizer, ger_length, train[:, 1]) trainY = encode_sequences(eng_tokenizer, eng_length, train[:, 0]) trainY = encode_output(trainY, eng_vocab_size) # prepare validation data testX = encode_sequences(ger_tokenizer, ger_length, test[:, 1]) testY = encode_sequences(eng_tokenizer, eng_length, test[:, 0]) testY = encode_output(testY, eng_vocab_size) |
We are now ready to define the model.
We will use an encoder-decoder LSTM model on this problem. In this architecture, the input sequence is encoded by a front-end model called the encoder then decoded word by word by a backend model called the decoder.
The function define_model() below defines the model and takes a number of arguments used to configure the model, such as the size of the input and output vocabularies, the maximum length of input and output phrases, and the number of memory units used to configure the model.
The model is trained using the efficient Adam approach to stochastic gradient descent and minimizes the categorical loss function because we have framed the prediction problem as multi-class classification.
The model configuration was not optimized for this problem, meaning that there is plenty of opportunity for you to tune it and lift the skill of the translations. I would love to see what you can come up with.
For more advice on configuring neural machine translation models, see the post:
| # define NMT model def define_model(src_vocab, tar_vocab, src_timesteps, tar_timesteps, n_units): model = Sequential() model.add(Embedding(src_vocab, n_units, input_length=src_timesteps, mask_zero=True)) model.add(LSTM(n_units)) model.add(RepeatVector(tar_timesteps)) model.add(LSTM(n_units, return_sequences=True)) model.add(TimeDistributed(Dense(tar_vocab, activation='softmax'))) return model # define model model = define_model(ger_vocab_size, eng_vocab_size, ger_length, eng_length, 256) model.compile(optimizer='adam', loss='categorical_crossentropy') # summarize defined model print(model.summary()) plot_model(model, to_file='model.png', show_shapes=True) |
Finally, we can train the model.
We train the model for 30 epochs and a batch size of 64 examples.
We use checkpointing to ensure that each time the model skill on the test set improves, the model is saved to file.
| # fit model filename = 'model.h5' checkpoint = ModelCheckpoint(filename, monitor='val_loss', verbose=1, save_best_only=True, mode='min') model.fit(trainX, trainY, epochs=30, batch_size=64, validation_data=(testX, testY), callbacks=[checkpoint], verbose=2) |
We can tie all of this together and fit the neural translation model.
The complete working example is listed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 | from pickle import load from numpy import array from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.utils import to_categorical from keras.utils.vis_utils import plot_model from keras.models import Sequential from keras.layers import LSTM from keras.layers import Dense from keras.layers import Embedding from keras.layers import RepeatVector from keras.layers import TimeDistributed from keras.callbacks import ModelCheckpoint # load a clean dataset def load_clean_sentences(filename): return load(open(filename, 'rb')) # fit a tokenizer def create_tokenizer(lines): tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer # max sentence length def max_length(lines): return max(len(line.split()) for line in lines) # encode and pad sequences def encode_sequences(tokenizer, length, lines): # integer encode sequences X = tokenizer.texts_to_sequences(lines) # pad sequences with 0 values X = pad_sequences(X, maxlen=length, padding='post') return X # one hot encode target sequence def encode_output(sequences, vocab_size): ylist = list() for sequence in sequences: encoded = to_categorical(sequence, num_classes=vocab_size) ylist.append(encoded) y = array(ylist) y = y.reshape(sequences.shape[0], sequences.shape[1], vocab_size) return y # define NMT model def define_model(src_vocab, tar_vocab, src_timesteps, tar_timesteps, n_units): model = Sequential() model.add(Embedding(src_vocab, n_units, input_length=src_timesteps, mask_zero=True)) model.add(LSTM(n_units)) model.add(RepeatVector(tar_timesteps)) model.add(LSTM(n_units, return_sequences=True)) model.add(TimeDistributed(Dense(tar_vocab, activation='softmax'))) return model # load datasets dataset = load_clean_sentences('english-german-both.pkl') train = load_clean_sentences('english-german-train.pkl') test = load_clean_sentences('english-german-test.pkl') # prepare english tokenizer eng_tokenizer = create_tokenizer(dataset[:, 0]) eng_vocab_size = len(eng_tokenizer.word_index) + 1 eng_length = max_length(dataset[:, 0]) print('English Vocabulary Size: %d' % eng_vocab_size) print('English Max Length: %d' % (eng_length)) # prepare german tokenizer ger_tokenizer = create_tokenizer(dataset[:, 1]) ger_vocab_size = len(ger_tokenizer.word_index) + 1 ger_length = max_length(dataset[:, 1]) print('German Vocabulary Size: %d' % ger_vocab_size) print('German Max Length: %d' % (ger_length)) # prepare training data trainX = encode_sequences(ger_tokenizer, ger_length, train[:, 1]) trainY = encode_sequences(eng_tokenizer, eng_length, train[:, 0]) trainY = encode_output(trainY, eng_vocab_size) # prepare validation data testX = encode_sequences(ger_tokenizer, ger_length, test[:, 1]) testY = encode_sequences(eng_tokenizer, eng_length, test[:, 0]) testY = encode_output(testY, eng_vocab_size) # define model model = define_model(ger_vocab_size, eng_vocab_size, ger_length, eng_length, 256) model.compile(optimizer='adam', loss='categorical_crossentropy') # summarize defined model print(model.summary()) plot_model(model, to_file='model.png', show_shapes=True) # fit model filename = 'model.h5' checkpoint = ModelCheckpoint(filename, monitor='val_loss', verbose=1, save_best_only=True, mode='min') model.fit(trainX, trainY, epochs=30, batch_size=64, validation_data=(testX, testY), callbacks=[checkpoint], verbose=2) |
Running the example first prints a summary of the parameters of the dataset such as vocabulary size and maximum phrase lengths.
| English Vocabulary Size: 2404 English Max Length: 5 German Vocabulary Size: 3856 German Max Length: 10 |
Next, a summary of the defined model is printed, allowing us to confirm the model configuration.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 | _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= embedding_1 (Embedding) (None, 10, 256) 987136 _________________________________________________________________ lstm_1 (LSTM) (None, 256) 525312 _________________________________________________________________ repeat_vector_1 (RepeatVecto (None, 5, 256) 0 _________________________________________________________________ lstm_2 (LSTM) (None, 5, 256) 525312 _________________________________________________________________ time_distributed_1 (TimeDist (None, 5, 2404) 617828 ================================================================= Total params: 2,655,588 Trainable params: 2,655,588 Non-trainable params: 0 _________________________________________________________________ |
A plot of the model is also created providing another perspective on the model configuration.
Plot of Model Graph for NMT
Next, the model is trained.
Each epoch takes about 30 seconds on modern CPU hardware; no GPU is required.
Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.
During the run, the model will be saved to the file model.h5, ready for inference in the next step.
| ... Epoch 26/30 Epoch 00025: val_loss improved from 2.20048 to 2.19976, saving model to model.h5 17s - loss: 0.7114 - val_loss: 2.1998 Epoch 27/30 Epoch 00026: val_loss improved from 2.19976 to 2.18255, saving model to model.h5 17s - loss: 0.6532 - val_loss: 2.1826 Epoch 28/30 Epoch 00027: val_loss did not improve 17s - loss: 0.5970 - val_loss: 2.1970 Epoch 29/30 Epoch 00028: val_loss improved from 2.18255 to 2.17872, saving model to model.h5 17s - loss: 0.5474 - val_loss: 2.1787 Epoch 30/30 Epoch 00029: val_loss did not improve 17s - loss: 0.5023 - val_loss: 2.1823 |
Evaluate Neural Translation Model
We will evaluate the model on the train and the test dataset.
The model should perform very well on the train dataset and ideally have been generalized to perform well on the test dataset.
Ideally, we would use a separate validation dataset to help with model selection during training instead of the test set. You can try this as an extension.
The clean datasets must be loaded and prepared as before.
| ... # load datasets dataset = load_clean_sentences('english-german-both.pkl') train = load_clean_sentences('english-german-train.pkl') test = load_clean_sentences('english-german-test.pkl') # prepare english tokenizer eng_tokenizer = create_tokenizer(dataset[:, 0]) eng_vocab_size = len(eng_tokenizer.word_index) + 1 eng_length = max_length(dataset[:, 0]) # prepare german tokenizer ger_tokenizer = create_tokenizer(dataset[:, 1]) ger_vocab_size = len(ger_tokenizer.word_index) + 1 ger_length = max_length(dataset[:, 1]) # prepare data trainX = encode_sequences(ger_tokenizer, ger_length, train[:, 1]) testX = encode_sequences(ger_tokenizer, ger_length, test[:, 1]) |
Next, the best model saved during training must be loaded.
| # load model model = load_model('model.h5') |
Evaluation involves two steps: first generating a translated output sequence, and then repeating this process for many input examples and summarizing the skill of the model across multiple cases.
Starting with inference, the model can predict the entire output sequence in a one-shot manner.
| translation = model.predict(source, verbose=0) |
This will be a sequence of integers that we can enumerate and lookup in the tokenizer to map back to words.
The function below, named word_for_id(), will perform this reverse mapping.
| # map an integer to a word def word_for_id(integer, tokenizer): for word, index in tokenizer.word_index.items(): if index == integer: return word return None |
We can perform this mapping for each integer in the translation and return the result as a string of words.
The function predict_sequence() below performs this operation for a single encoded source phrase.
| # generate target given source sequence def predict_sequence(model, tokenizer, source): prediction = model.predict(source, verbose=0)[0] integers = [argmax(vector) for vector in prediction] target = list() for i in integers: word = word_for_id(i, tokenizer) if word is None: break target.append(word) return ' '.join(target) |
Next, we can repeat this for each source phrase in a dataset and compare the predicted result to the expected target phrase in English.
We can print some of these comparisons to screen to get an idea of how the model performs in practice.
We will also calculate the BLEU scores to get a quantitative idea of how well the model has performed.
You can learn more about the BLEU score here:
The evaluate_model() function below implements this, calling the above predict_sequence() function for each phrase in a provided dataset.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 | # evaluate the skill of the model def evaluate_model(model, tokenizer, sources, raw_dataset): actual, predicted = list(), list() for i, source in enumerate(sources): # translate encoded source text source = source.reshape((1, source.shape[0])) translation = predict_sequence(model, eng_tokenizer, source) raw_target, raw_src = raw_dataset[i] if i < 10: print('src=[%s], target=[%s], predicted=[%s]' % (raw_src, raw_target, translation)) actual.append([raw_target.split()]) predicted.append(translation.split()) # calculate BLEU score print('BLEU-1: %f' % corpus_bleu(actual, predicted, weights=(1.0, 0, 0, 0))) print('BLEU-2: %f' % corpus_bleu(actual, predicted, weights=(0.5, 0.5, 0, 0))) print('BLEU-3: %f' % corpus_bleu(actual, predicted, weights=(0.3, 0.3, 0.3, 0))) print('BLEU-4: %f' % corpus_bleu(actual, predicted, weights=(0.25, 0.25, 0.25, 0.25))) |
We can tie all of this together and evaluate the loaded model on both the training and test datasets.
The complete code listing is provided below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 | from pickle import load from numpy import array from numpy import argmax from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.models import load_model from nltk.translate.bleu_score import corpus_bleu # load a clean dataset def load_clean_sentences(filename): return load(open(filename, 'rb')) # fit a tokenizer def create_tokenizer(lines): tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer # max sentence length def max_length(lines): return max(len(line.split()) for line in lines) # encode and pad sequences def encode_sequences(tokenizer, length, lines): # integer encode sequences X = tokenizer.texts_to_sequences(lines) # pad sequences with 0 values X = pad_sequences(X, maxlen=length, padding='post') return X # map an integer to a word def word_for_id(integer, tokenizer): for word, index in tokenizer.word_index.items(): if index == integer: return word return None # generate target given source sequence def predict_sequence(model, tokenizer, source): prediction = model.predict(source, verbose=0)[0] integers = [argmax(vector) for vector in prediction] target = list() for i in integers: word = word_for_id(i, tokenizer) if word is None: break target.append(word) return ' '.join(target) # evaluate the skill of the model def evaluate_model(model, tokenizer, sources, raw_dataset): actual, predicted = list(), list() for i, source in enumerate(sources): # translate encoded source text source = source.reshape((1, source.shape[0])) translation = predict_sequence(model, eng_tokenizer, source) raw_target, raw_src = raw_dataset[i] if i < 10: print('src=[%s], target=[%s], predicted=[%s]' % (raw_src, raw_target, translation)) actual.append([raw_target.split()]) predicted.append(translation.split()) # calculate BLEU score print('BLEU-1: %f' % corpus_bleu(actual, predicted, weights=(1.0, 0, 0, 0))) print('BLEU-2: %f' % corpus_bleu(actual, predicted, weights=(0.5, 0.5, 0, 0))) print('BLEU-3: %f' % corpus_bleu(actual, predicted, weights=(0.3, 0.3, 0.3, 0))) print('BLEU-4: %f' % corpus_bleu(actual, predicted, weights=(0.25, 0.25, 0.25, 0.25))) # load datasets dataset = load_clean_sentences('english-german-both.pkl') train = load_clean_sentences('english-german-train.pkl') test = load_clean_sentences('english-german-test.pkl') # prepare english tokenizer eng_tokenizer = create_tokenizer(dataset[:, 0]) eng_vocab_size = len(eng_tokenizer.word_index) + 1 eng_length = max_length(dataset[:, 0]) # prepare german tokenizer ger_tokenizer = create_tokenizer(dataset[:, 1]) ger_vocab_size = len(ger_tokenizer.word_index) + 1 ger_length = max_length(dataset[:, 1]) # prepare data trainX = encode_sequences(ger_tokenizer, ger_length, train[:, 1]) testX = encode_sequences(ger_tokenizer, ger_length, test[:, 1]) # load model model = load_model('model.h5') # test on some training sequences print('train') evaluate_model(model, eng_tokenizer, trainX, train) # test on some test sequences print('test') evaluate_model(model, eng_tokenizer, testX, test) |
Running the example first prints examples of source text, expected and predicted translations, as well as scores for the training dataset, followed by the test dataset.
Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.
Looking at the results for the test dataset first, we can see that the translations are readable and mostly correct.
For example: “ich bin brillentrager” was correctly translated to “i wear glasses“.
We can also see that the translations were not perfect, with “hab ich nicht recht” translated to “am i fat” instead of the expected “am i wrong“.
We can also see the BLEU-4 score of about 0.45, which provides an upper bound on what we might expect from this model.
| src=[er ist ein blodmann], target=[hes a jerk], predicted=[hes a jerk] src=[ich bin brillentrager], target=[i wear glasses], predicted=[i wear glasses] src=[tom hat mich aufgezogen], target=[tom raised me], predicted=[tom tricked me] src=[ich zahle auf tom], target=[i count on tom], predicted=[ill call tom tom] src=[ich kann rauch sehen], target=[i can see smoke], predicted=[i can help you] src=[tom fuhlte sich einsam], target=[tom felt lonely], predicted=[tom felt uneasy] src=[hab ich nicht recht], target=[am i wrong], predicted=[am i fat] src=[gestatten sie mir zu gehen], target=[allow me to go], predicted=[do me to go] src=[du hast mir gefehlt], target=[i missed you], predicted=[i missed you] src=[es ist zu spat], target=[it is too late], predicted=[its too late] BLEU-1: 0.844852 BLEU-2: 0.779819 BLEU-3: 0.699516 BLEU-4: 0.452614 |
Looking at the results on the test set, do see readable translations, which is not an easy task.
For example, we see “tom erblasste” correctly translated to “tom turned pale“.
We also see some poor translations and a good case that the model could suffer from further tuning, such as “ich brauche erste hilfe” translated as “i need them you” instead of the expected “i need first aid“.
A BLEU-4 score of about 0.153 was achieved, providing a baseline skill to improve upon with further improvements to the model.
| src=[mein hund hat es gefressen], target=[my dog ate it], predicted=[my dog is tom] src=[ich hore das telefon], target=[i hear the phone], predicted=[i want this this] src=[ich fuhlte mich hintergangen], target=[i felt betrayed], predicted=[i didnt] src=[wer scherzt], target=[whos joking], predicted=[whos is] src=[wir furchten uns], target=[were afraid], predicted=[we are] src=[reden sie weiter], target=[keep talking], predicted=[keep them] src=[was fur ein spa], target=[what fun], predicted=[what an fun] src=[ich bin auch siebzehn], target=[im too], predicted=[im so expert] src=[ich bin dein vater], target=[im your father], predicted=[im your your] src=[ich brauche erste hilfe], target=[i need first aid], predicted=[i need them you] BLEU-1: 0.499623 BLEU-2: 0.365875 BLEU-3: 0.295824 BLEU-4: 0.153535 |
Extensions
This section lists some ideas for extending the tutorial that you may wish to explore.
- Data Cleaning. Different data cleaning operations could be performed on the data, such as not removing punctuation or normalizing case, or perhaps removing duplicate English phrases.
- Vocabulary. The vocabulary could be refined, perhaps removing words used less than 5 or 10 times in the dataset and replaced with “unk“.
- More Data. The dataset used to fit the model could be expanded to 50,000, 100,000 phrases, or more.
- Input Order. The order of input phrases could be reversed, which has been reported to lift skill, or a Bidirectional input layer could be used.
- Layers. The encoder and/or the decoder models could be expanded with additional layers and trained for more epochs, providing more representational capacity for the model.
- Units. The number of memory units in the encoder and decoder could be increased, providing more representational capacity for the model.
- Regularization. The model could use regularization, such as weight or activation regularization, or the use of dropout on the LSTM layers.
- Pre-Trained Word Vectors. Pre-trained word vectors could be used in the model.
- Recursive Model. A recursive formulation of the model could be used where the next word in the output sequence could be conditional on the input sequence and the output sequence generated so far.
Further Reading
This section provides more resources on the topic if you are looking to go deeper.
Summary
In this tutorial, you discovered how to develop a neural machine translation system for translating German phrases to English.
Specifically, you learned:
- How to clean and prepare data ready to train a neural machine translation system.
- How to develop an encoder-decoder model for machine translation.
- How to use a trained model for inference on new input phrases and evaluate the model skill.
Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.
Note: This post is an excerpt chapter from: “Deep Learning for Natural Language Processing“. Take a look, if you want more step-by-step tutorials on getting the most out of deep learning methods when working with text data.
![](https://blogger.googleusercontent.com/img/proxy/AVvXsEgiYV_qJof_iG3Vfme-VO0SYYPzs1aybB5LTgE0xw-lYZ-B80MJMw9p2cefwChHUmES3FgIa_Zr3uk3tonjksLNddOSOTJjWre5kw84WPYA-nR_9fs3ldFzFiBAO2-0-cl-lRVmFBj5YJIlCoKkZqRMjyU7g00NqvKpPg=s0-d-e1-ft"")
No comments:
Post a Comment