Introduction to machine translation

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

Machine translation

Hello translations

Machine translation

Hello translation with Google icon

Course outline

Chapter 1 - Introduction to machine translation
Chapter 2 - Implement a machine translation model (encoder-decoder architecture)
Chapter 3 - Training the model and generating translations
Chapter 4 - Improving the translation model

Dataset (English-French sentence corpus)

English corpus

new jersey is sometimes quiet during autumn , and it is snowy in april .
the united states is usually chilly during july , and it is usually freezing ...
california is usually quiet during march , and it is usually hot in june .

French corpus

new jersey est parfois calme pendant l' automne , et il est neigeux en avril .
les états-unis est généralement froid en juillet , et il gèle habituellement ...
california est généralement calme en mars , et il est généralement chaud en juin .

¹ https://github.com/udacity/deep-learning/tree/master/language-translation/data

Machine translation - Overview

English to French

Machine translation - Overview

Source target language terms

Machine translation - Overview

Machine translation model

Machine translation - Overview

Machine translation model

One-hot encoded vectors

A vector of ones and zeros
Vector length is determined by the size of the vocabulary
Vocabulary - the collection of unique words in the dataset

One hot vectors

One-hot encoded vectors

A mapping containing words and their corresponding indices

word2index = {"I":0, "like": 1, "cats": 2}

Converting words to IDs or indices

words = ["I", "like", "cats"]
word_ids = [word2index[w] for w in words]
print(word_ids)

[0, 1, 2]

One-hot encoded vectors

One-hot encoding without specifying output vector length

onehot_1 = to_categorical(word_ids)
print([(w,ohe.tolist()) for w,ohe in zip(words, onehot_1)])

[('I', [1.0, 0.0, 0.0]), ('like', [0.0, 1.0, 0.0]), ('cats', [0.0, 0.0, 1.0])]

One-hot encoding with specifying output vector length

onehot_2 = to_categorical(word_ids, num_classes=5)
print([(w,ohe.tolist()) for w,ohe in zip(words, onehot_2)])

[('I', [1.0, 0.0, 0.0, 0.0, 0.0]), ('like', [0.0, 1.0, 0.0, 0.0, 0.0]), 
('cats', [0.0, 0.0, 1.0, 0.0, 0.0])]

Let's practice!

Machine Translation with Keras