Getting started with Deep Learning for Natural Language Processing

Learn how to build NLP applications with Deep Learning (English Edition)

Sunil Patel

• Editorial: BPB Publications
• Año: 2021
• ISBN:
  9789389898118
• Categoría: Computer Science
• Categoría: Cyber Security

Learn how to redesign NLP applications from scratch. Key Features

• Get familiar with the basics of any Machine Learning or Deep Learning application.
• Understand how does preprocessing work in NLP pipeline.
• Use simple PyTorch snippets to create basic building blocks of the network commonly used in NLP.
• Get familiar with the advanced embedding technique, Generative network, and Audio signal processing techniques.
• 
  Description
  Natural language processing (NLP) is one of the areas where many Machine Learning and Deep Learning techniques are applied. This book covers wide areas, including the fundamentals of Machine Learning, Understanding and optimizing Hyperparameters, Convolution Neural Networks (CNN), and Recurrent Neural Networks (RNN). This book not only covers the classical concept of text processing but also shares the recent advancements. This book will empower users in designing networks with the least computational and time complexity. This book not only covers basics of Natural Language Processing but also helps in deciphering the logic behind advanced concepts/architecture such as Batch Normalization, Position Embedding, DenseNet, Attention Mechanism, Highway Networks, Transformer models and Siamese Networks. This book also covers recent advancements such as ELMo-BiLM, SkipThought, and Bert. This book also covers practical implementation with step by step explanation of deep learning techniques in Topic Modelling, Text Generation, Named Entity Recognition, Text Summarization, and Language Translation. In addition to this, very advanced and open to research topics such as Generative Adversarial Network and Speech Processing are also covered. What you will learn
  
• Learn how to leveraging GPU for Deep Learning
• Learn how to use complex embedding models such as BERT
• Get familiar with the common NLP applications
• Learn how to use GANs in NLP
• Learn how to process Speech data and implementing it in Speech applications
• 
  Who this book is for
  This book is a must-read to everyone who wishes to start the career with Machine learning and Deep Learning. This book is also for those who want to use GPU for developing Deep Learning applications. Table of Contents
  1. Understanding the basics of learning Process
  2. Text Processing Techniques
  3. Representing Language Mathematically
  4. Using RNN for NLP
  5. Applying CNN In NLP Tasks
  6. Accelerating NLP with Advanced Embeddings
  7. Applying Deep Learning to NLP tasks
  8. Application of Complex Architectures in NLP
  9. Understanding Generative Networks
  10. Techniques of Speech Processing
  11. The Road Ahead About the Authors
  Sunil Patel has completed his master's in Information Technology from the Indian Institute of Information technology-Allahabad with a thesis focused on investigating 3D protein-protein interactions with deep learning. Sunil has worked with TCS Innovation Labs, Excelra, and Innoplexus before joining to Nvidia. The main areas of research were using Deep Learning, Natural language processing in Banking, and healthcare domain. Sunil started experimenting with deep learning by implanting the basic layer used in pipelines and then developing complex pipelines for a real-life problem. Apart from this, Sunil has also participated in CASP-2014 in collaboration with SCFBIO-IIT Delhi to efficiently predict possible Protein multimer formation and its impact on diseases using Deep Learning. Currently, Sunil works with Nvidia as Data Scientist – III. LinkedIn Profile: https://www.linkedin.com/in/linus1/

CHAPTER 1

Understanding the Basics of Learning Process

This chapter covers the most basic aspects of machine learning. It will help you in understanding the basic mathematical representation of learning algorithm and teach you how to design a machine learning model from scratch. After constructing this model, we will understand the methods used to gauge the model’s prediction accuracy using different accuracy metrics. Going further, we will coer the bias-variance problem and diagnose such a problem with a technique called learning curves. Once we get our model correct, we need it to generalize well on unknown datasets that can be understood through a chapter on regularization. After perfecting such a model, it must be efficiently deployed to help improve speed and accuracy.

Structure

In this chapter, we will cover the following topics:

• Learning from data
• Error/noise reduction
• Bias-variance reduction
• Learning curves
• Regularization
• Training and inference
• The three learning principles

Objective

Building a simple model for efficient training and gauging the model’s accuracy will be covered in this chapter. It will help you understand the usage of popular software and hardware acceleration for faster training and inference. We’ll end this chapter with three learning principles that are extremely important to machine learning.

Pre-requisites

I have provided some of the examples through code, and the code for this chapter are present in ch1 folder at GitHub repository (https://github.com/bpbpublications/Getting-started-with-Deep-Learning-for-Natural-Language-Processing). Basic know-ledge of the following Python packages is required to understand this chapter:

• NumPy
• Scikit-Learn
• Matplotlib
• Pandas

This chapter has one example that uses PyTorch for demonstration. If you don’t know PyTorch, we will cover it in detail in Chapter 2, Text Pre-Processing Techniques in NLP.

Learning from Data

In this data-centric world, a little improvement in the existing application can potentially help earn millions. We all remember a big prize (the $1,000,000) that Netflix gave to the winner for improving the algorithm’s accuracy by 10.06%. A similar opportunity exists in financial planning, be it Forex forecasting or trade market analysis. Minute improvements in such use cases can provide beneficial results. One must explore the entire logic behind the process to improve something; this is what we call learning from data. Machine learning is an interesting area where one can use historical data to make a system capable of identifying observed patterns in the new data. However, machine learning cannot be applied to all problems; a rule of thumb is considered to decide whether machine learning should be applied to a given problem:

• There must exist a hidden pattern
• We cannot find such a pattern by applying simple mathematical approaches
• There must be historical/relevant data about the task concerned

In this chapter, I will start with a basic perceptron model to give you a taste of the learning model. After building the perceptron model, I will discuss Error/Noise and its detection using bias-variance and learning curves. In the learning curve, we will look at how the complexity of the algorithm helps mitigate a high-bias problem. Then, we will cover regularization techniques to achieve better generalization. All these techniques help get a better model, and then it’s time to deploy such a model efficiently. Later in the chapter, we will cover techniques for faster and better inferencing and then look at the three learning principles that are not directly related to machine learning but help in giving state-of-the-art performance. Before going further, we will briefly discuss the mathematical formalization of any supervised learning problem.

Let’s assume that we are talking about supervised learning paradigm. Supervised learning takes finite pairs of X and Y for learning. X and Y can be of the different types according to learning the goals. In the following table, we can see some examples with the nature of X and Y described for different learning problems:

Data	X	Y
Credit card approval	Vector of numerical, categorical, or ordinal values	Two classes: Accept or Reject
Humpback whale identification	Different images	Two or more classes
Sentiment analysis	Text classification	Two or more classes
Wake word identification	Speech classification	Two or more classes
Producing an abstract representation	Detailed text	Summarized text
Producing an abstract representation	Full-length video	Summarized video
Text translation	Text in one language	Translated text in another language
Speech translation	Speech in one language	Translated speech in another language
High-resolution images	Low-resolution image	High-resolution image

Table 1.1

Here, Y is the label for X. Each X and Y is paired, as shown in the following equation:

(X, Y) = ((x_i, y_i),…,(x_n, y_n))

Where x_i,…..,x_n are individual data points in X, and y_i,…..,y_n are individual data points in Y. The main task of our hypothesis function f is to apply it over x_i to predict ŷ_i = f(x_i). The goal is to predict (ŷ_i) so that it is the same as or near the original label (y_i) and the Error (E) between the predicted label and the original one(|Ŷ – Y|) tends to become 0. The function is (X, Y) = (x_i, y_i, …, x_n, x_n). The overall procedure to learn any function can be summarized as in the following flow diagram:

Figure 1.1: Supervise learning paradigm with all major components involved in training and evaluation.

As shown in the preceding figure, any of the hypotheses h is used for making predictions. Here, hypothesis can be anything like Linear, Logistic, Polynomial, SVM, RF, or Neural Network. Hypothesis h is also called function f in general terms.

There are also other types of learning techniques, like unsupervised learning and reinforcement learning. An unsupervised learning paradigm has no label attached to the data; it only has x_i,…..,x_n ∈ X, and there is no Y label. In fact, unsupervised methodologies are gaining popularity nowadays and are responsible for pushing the state-of-the-art model in the field of vision and NLP even higher. Popular models like Bert and Megatron are examples of unsupervised models. On the other hand, reinforcement learning is a technique where an agent tries to maximize the immediate or cumulative reward by learning/adapting to a given environment. We will learn and apply unsupervised learning to NLP problems in the upcoming chapters.

Implementing the Perceptron Model

Well, this chapter is a little out of sync, but we will discuss the perceptron model. Perceptron with no activation function is a linear model. The perceptron algorithm was invented by Frank Rosenblatt in 1957 at the Cornell Aeronautical Laboratory. To date, this is the most important and widely used model in the course of the machine learning.

Let’s take all the features, that is, x_i,…..,x_n, and try to derive a hypothesis h that maps y ⇒ ŷ., where ŷ is the predicted label and y is the original label. Our hypothesis h can be one of the simplest possible perceptron models with a linear activation function. The preliminary hypothesis can be thought of as in the following equation:

Where w_i are learnable weights, w_i changes as per the feedback sign...