What this book covers
Chapter 1, Introduction to Natural Language Processing, embarks us on our journey with a gentle introduction to NLP. In this chapter, we will first look at the reasons we need NLP. Next, we will discuss some of the common subtasks found in NLP. Thereafter, we will discuss the two main eras of NLP—the traditional era and the deep learning era. We will gain an understanding of the characteristics of the traditional era by working through how a language modeling task might have been solved with traditional algorithms. Then, we will discuss the deep learning era, where deep learning algorithms are heavily utilized for NLP. We will also discuss the main families of deep learning algorithms. We will then discuss the fundamentals of one of the most basic deep learning algorithms—a fully connected neural network. We will conclude the chapter with a road map that provides a brief introduction to the coming chapters.
Chapter 2, Understanding TensorFlow, introduces you to the Python TensorFlow library—the primary platform we will implement our solutions on. We will start by writing code to perform a simple calculation in TensorFlow. We will then discuss how things are executed, starting from running the code to getting results. Thereby, we will understand the underlying components of TensorFlow in detail. We will further strengthen our understanding of TensorFlow with a colorful analogy of a restaurant and see how orders are fulfilled. Later, we will discuss more technical details of TensorFlow, such as the data structures and operations (mostly related to neural networks) defined in TensorFlow. Finally, we will implement a fully connected neural network to recognize handwritten digits. This will help us to understand how an end-to-end solution might be implemented with TensorFlow.
Chapter 3, Word2vec – Learning Word Embeddings, begins by discussing how to solve NLP tasks with TensorFlow. In this chapter, we will see how neural networks can be used to learn word vectors or word representations. Word vectors are also known as word embeddings. Word vectors are numerical representations of words that have similar values for similar words and different values for different words. First, we will discuss several traditional approaches to achieving this, which include using a large human-built knowledge base known as WordNet. Then, we will discuss the modern neural network-based approach known as Word2vec, which learns word vectors without any human intervention. We will first understand the mechanics of Word2vec by working through a hands-on example. Then, we will discuss two algorithmic variants for achieving this—the skip-gram and continuous bag-of-words (CBOW) model. We will discuss the conceptual details of the algorithms, as well as how to implement them in TensorFlow.
Chapter 4, Advance Word2vec, takes us on to more advanced topics related to word vectors. First, we will compare skip-gram and CBOW to see whether a winner exists. Next, we will discuss several improvements that can be used to improve the performance of the Word2vec algorithms. Then, we will discuss a more recent and powerful word embedding learning algorithm—the GloVe (global vectors) algorithm. Finally, we will look at word vectors in action, in a document classification task. In that exercise, we will see that word vectors are powerful enough to represent the topic (for example, entertainment and sport) that the document belongs to.
Chapter 5, Sentence Classification with Convolutional Neural Networks, discusses convolution neural networks (CNN)—a family of neural networks that excels at processing spatial data such as images or sentences. First, we will develop a solid high-level understanding of CNNs by discussing how they process data and what sort of operations are involved. Next, we will dive deep into each of the operations involved in the computations of a CNN to understand the underpinning mathematics of a CNN. Finally, we will walk through two exercises. First, we will classify hand written digit images with a CNN. We will see that CNNs are is capable of reaching a very high accuracy quickly for this task. Next, we will explore how CNNs can be used to classify sentences. Particularly, we will ask a CNN to predict whether a sentence is about an object, person, location, and so on.
Chapter 6, Recurrent Neural Networks, is about a powerful family of neural networks that can model sequences of data, known as recurrent neural networks (RNNs). We will first discuss the mathematics behind the RNNs and the update rules that are used to update the RNNs over time during learning. Then, we will discuss section different variants of RNNs and their applications (for example, one-to-one RNNs and one-to-many RNNs). Finally, we will go through an exercise where RNNs are used for a text generation task. In this, we will train the RNN on folk stories and ask the RNN to produce a new story. We will see that RNNs are poor at persisting long-term memory. Finally, we will discuss a more advanced variant of RNNs, which we will call RNN-CF, which is able to persist memory for longer.
Chapter 7, Long Short-Term Memory Networks, allows us to explore more powerful techniques that are able to remember for a longer period of time, having found out that RNNs are poor at retaining long-term memory. We will discuss one such technique in this chapter—Long Short-Term Memory Networks (LSTMs). LSTMs are more powerful and have been shown to outperform other sequential models in many time-series tasks. We will first investigate the underlying mathematics and update the rules of the LSTM, along with a colorful example that illustrates why each computation matters. Then, we will look at how LSTMs can persist memory for longer. Next, we will discuss how we can improve LSTMs prediction capabilities further. Finally, we will discuss several variants of LSTMs that have a more complex structure (LSTMs with peephole connections), as well as a method that tries to simplify the LSTMs gated recurrent units (GRUs).
Chapter 8, Applications of LSTM – Generating Text, extensively evaluates how LSTMs perform in a text generation task. We will qualitatively and quantitatively measure how good the text generated by LSTMs is. We will also conduct comparisons between LSTMs, LSTMs with peephole connections, and GRUs. Finally, we will see how we can bring word embeddings into the model to improve the text generated by LSTMs.
Chapter 9, Applications of LSTM – Image Caption Generation, moves us on to multimodal data (that is, images and text) after coping with textual data. In this chapter, we will investigate how we can automatically generate descriptions for a given image. This involves combining a feed-forward model (that is, a CNN) with a word embedding layer and a sequential model (that is, an LSTM) in a way that forms an end-to-end machine learning pipeline.
Chapter 10, Sequence to Sequence Learning – Neural Machine Translation, is about the implementing neural machine translation (NMT) model. Machine translation is where we translate a sentence/phrase from a source language into a target language. We will first briefly discuss what machine translation is. This will be followed by a section about the history of machine translation. Then, we will discuss the architecture of modern neural machine translation models in detail, including the training and inference procedures. Next, we will look at how to implement an NMT system from scratch. Finally, we will explore ways to improve standard NMT systems.
Chapter 11, Current Trends and Future of Natural Language Processing, the final chapter, focuses on the current and future trends of NLP. We will discuss the latest discoveries related to the systems and tasks we discussed in the previous chapters. This chapter will cover most of the exciting novel innovations, as well as giving you in-depth intuition to implement some of the technologies.
Appendix, Mathematical Foundations and Advanced TensorFlow, will introduce the reader to various mathematical data structures (for example, matrices) and operations (for example, matrix inverse). We will also discuss several important concepts in probability. We will then introduce Keras—a high-level library that uses TensorFlow underneath. Keras makes the implementing of neural networks simpler by hiding some of the details in TensorFlow, which some might find challenging. Concretely, we will see how we can implement a CNN with Keras, to get a feel of how to use Keras. Next, we will discuss how we can use the seq2seq library in TensorFlow to implement a neural machine translation system with much less code that we used in Chapter 11, Current Trends and the Future of Natural Language Processing. Finally, we will walk you through a guide aimed at teaching to use the TensorBoard to visualize word embeddings. TensorBoard is a handy visualization tool that is shipped with TensorFlow. This can be used to visualize and monitor various variables in your TensorFlow client.
