book

Deep Learning for Natural Language Processing

by Stephan Raaijmakers

November 2022

Beginner to intermediate

296 pages

8h 27m

English

Manning Publications

Read now

Unlock full access

prefaceacknowledgmentsabout this bookWho should read this bookHow this book is organized: A road mapAbout the codeliveBook discussion forumabout the authorabout the cover illustration
1.1 A selection of machine learning methods for NLP1.1.1 The perceptron1.1.2 Support vector machines1.1.3 Memory-based learning1.2 Deep learning1.3 Vector representations of language1.3.1 Representational vectors1.3.2 Operational vectors1.4 Vector sanitization1.4.1 The hashing trick1.4.2 Vector normalizationSummary
2.1 Basic architectures of deep learning2.1.1 Deep multilayer perceptrons2.1.2 Two basic operators: Spatial and temporal2.2 Deep learning and NLP: A new paradigmSummary
3.1 Embeddings3.1.1 Embedding by direct computation: Representational embeddings3.1.2 Learning to embed: Procedural embeddings3.2 From words to vectors: Word2Vec3.3 From documents to vectors: Doc2VecSummary

4.1 The problem4.2 The data4.2.1 Authorship attribution and verification data4.3 Data representation4.3.1 Segmenting documents4.3.2 Word-level information4.3.3 Subword-level information4.4 Models for measuring similarity4.4.1 Authorship attribution4.4.2 Verifying authorshipSummary
5.1 Memory and language5.1.1 The problem: Question Answering5.2 Data and data processing5.3 Question Answering with sequential models5.3.1 RNNs for Question Answering5.3.2 LSTMs for Question Answering5.3.3 End-to-end memory networks for Question AnsweringSummary
6.1 Memory networks for sequential NLP6.2 Data and data processing6.2.1 PP-attachment data6.2.2 Dutch diminutive data6.2.3 Spanish part-of-speech data6.3 Strongly supervised memory networks: Experiments and results6.3.1 PP-attachment6.3.2 Dutch diminutives6.3.3 Spanish part-of-speech tagging6.4 Semi-supervised memory networks6.4.1 Semi-supervised memory networks: Experiments and resultsSummary
7.1 Neural attention7.2 Data7.3 Static attention: MLP7.4 Temporal attention: LSTM7.5 Experiments7.5.1 MLP7.5.2 LSTMSummary
8.1 Introduction to multitask learning8.2 Multitask learning8.3 Multitask learning for consumer reviews: Yelp and Amazon8.3.1 Data handling8.3.2 Hard parameter sharing8.3.3 Soft parameter sharing8.3.4 Mixed parameter sharing8.4 Multitask learning for Reuters topic classification8.4.1 Data handling8.4.2 Hard parameter sharing8.4.3 Soft parameter sharing8.4.4 Mixed parameter sharing8.5 Multitask learning for part-of-speech tagging and named-entity recognition8.5.1 Data handling8.5.2 Hard parameter sharing8.5.3 Soft parameter sharing8.5.4 Mixed parameter sharingSummary
9.1 BERT up close: Transformers9.2 Transformer encoders9.2.1 Positional encoding9.3 Transformer decoders9.4 BERT: Masked language modeling9.4.1 Training BERT9.4.2 Fine-tuning BERT9.4.3 Beyond BERTSummary
10.1 Introduction: Working with BERT in practice10.2 A BERT layer10.3 Training BERT on your data10.4 Fine-tuning BERT10.5 Inspecting BERT10.5.1 Homonyms in BERT10.6 Applying BERTSummary

Content preview from Deep Learning for Natural Language Processing

9 Transformers

This chapter covers

Understanding the inner workings of Transformers
Deriving word embeddings with BERT
Comparing BERT and Word2Vec
Working with XLNet

In late 2018, researchers from Google published a paper introducing a deep learning technique that would soon become a major breakthrough: Bidirectional Encoder Representations from Transformers, or BERT (Devlin et al. 2018). BERT aims to derive word embeddings from raw textual data just like Word2Vec, but does it in a much more clever and powerful manner: it takes into account both the left and right contexts when learning vector representations for words (figure 9.1). In contrast, Word2Vec uses a single piece of context. But this is not the only difference. BERT is grounded in ...