Book Image

Getting Started with Google BERT

By : Sudharsan Ravichandiran
Book Image

Getting Started with Google BERT

By: Sudharsan Ravichandiran

Overview of this book

BERT (bidirectional encoder representations from transformer) has revolutionized the world of natural language processing (NLP) with promising results. This book is an introductory guide that will help you get to grips with Google's BERT architecture. With a detailed explanation of the transformer architecture, this book will help you understand how the transformer’s encoder and decoder work. You’ll explore the BERT architecture by learning how the BERT model is pre-trained and how to use pre-trained BERT for downstream tasks by fine-tuning it for NLP tasks such as sentiment analysis and text summarization with the Hugging Face transformers library. As you advance, you’ll learn about different variants of BERT such as ALBERT, RoBERTa, and ELECTRA, and look at SpanBERT, which is used for NLP tasks like question answering. You'll also cover simpler and faster BERT variants based on knowledge distillation such as DistilBERT and TinyBERT. The book takes you through MBERT, XLM, and XLM-R in detail and then introduces you to sentence-BERT, which is used for obtaining sentence representation. Finally, you'll discover domain-specific BERT models such as BioBERT and ClinicalBERT, and discover an interesting variant called VideoBERT. By the end of this BERT book, you’ll be well-versed with using BERT and its variants for performing practical NLP tasks.
Table of Contents (15 chapters)
1
Section 1 - Starting Off with BERT
5
Section 2 - Exploring BERT Variants
8
Section 3 - Applications of BERT

Summary

We started off the chapter by understanding how Sentence-BERT works. We learned that in Sentence-BERT, we use mean or max pooling for computing the sentence representation. We also learned that Sentence-BERT is basically a pre-trained BERT model that is fine-tuned for computing sentence representation. For fine-tuning the pre-trained BERT model, Sentence-BERT uses a Siamese and triplet network architecture, which makes the fine-tuning faster and helps in obtaining accurate sentence embeddings.

Later, we learned how to use the sentence-transformers library. We learned how to compute sentence representation and also how to compute the semantic similarity between a sentence pair using sentence-transformers. Following this, we learned how to make monolingual embeddings multilingual using knowledge distillation. We learned how to make the student (XLM-R) generate multilingual embeddings the same as how the teacher (Sentence-BERT) generates the monolingual embedding.

Next, we explored...