With the advancement of computer technology over the past five decades, natural language processing (NLP) has developed into an important interdisciplinary research field which covers a variety of areas, such as artificial intelligence (AI), human-computer interaction, information processing, quantitative linguistics, computational linguistics, corpus linguistics, machine translation (MT), and so forth. Research on NLP modeling has generally experienced three major phases, namely, rationalism, empiricism, and deep learning. In particular, deep learning has demonstrated robustness in the development of NLP since early 2010 and has breathed new life into both AI engineering and language studies involving quantitative methods. In this respect, Deng and Liu’s book entitled Deep Learning in Natural Language Processing appears to be a timely and interesting publication because it reports on the state of the art in deep learning and introduces a number of studies pertaining to independent application areas of NLP, such as conversational language understanding (CLU), lexical analysis and parsing, and MT. The potential readership of the book is a broad spectrum, which may include NLP researchers, language engineers, quantitative and computational language researchers, translation educators, computer programmers, and anyone who shows an interest in quantitative studies.

The book under review is composed of 11 chapters, each devoted to a specific topic in relation to NLP applications. The book starts with an introduction to the fundamental knowledge of NLP from the angle of AI. Based on the introduction, Deng and Liu provide a holistic overview of NLP development in terms of rationalism, empiricism and deep learning (Chapter 1). They emphasize a paradigm shift in which deep learning has become the mainstream research direction, applying neural networks to NLP. A number of potential key elements supporting this paradigm shift are identified, including distributed representations of linguistic entities, embedding-driven semantic generalization, long-span deep sequence modeling of natural language, hierarchical networks representing low-to-high linguistic levels, and deep learning methods for NLP tasks. With regard to the aforementioned elements, Deng and Liu clarify five future research directions: construction of neural-symbolic integration frameworks; exploration of more advanced memory models; provision of new algorithms for unsupervised and generative deep learning; multimodal and multitask learning; and meta-learning. The studies in the following chapters primarily follow these research directions.

Focusing on deep learning in CLU, Tur et al.’s study (Chapter 2) illustrates a number of practical goal-oriented CLU frameworks and models which are designed for three core tasks: specifically, domain detection, intent determination, and slot filling. These systems are evaluated from a holistic perspective, ranging from Deng and Yu’s deep convex net to Liu and Lane’s joint slot filling and intent determination model (Lane et al. Reference Lane, Howard and Hapke2018). Tur et al. summarize two dimensions concerning deep learning CLU: end-to-end learning, and efficient encoders implemented without recursive neural network unrolling, and they suggest that future studies in this area should concentrate on transfer learning, unsupervised learning, and reinforcement learning.

Celikyilmaz et al.’s study (Chapter 3) focuses on spoken and text-based dialog systems and demonstrates a survey on the application of neural models to dialog modeling. A variety of open-source software applications (e.g. OpenDial, ParlAI, SimpleDS) and datasets (e.g. the Carnegie Mellon Communicator Corpus, the Dialog State Tracking Challenge Dataset, and Facebook’s Dialog Datasets) are introduced for dialog modeling. In addition, a number of commonly used metrics, such as BLEU and METEOR, are discussed and analyzed with regard to the evaluation of dialog model performance.

Chapter 4 is concerned with deep learning in lexical analysis and parsing and is particularly useful for studies involving word segmentation, part-of-speech tagging and syntactic parsing. Che and Zhang first make a clear distinction between two types of traditional methods in structured prediction: graph-based methods and transition-based methods. Then they explain how deep learning and neural networks can be incorporated into these two methods and exemplify the advantages of neural network models over the traditional statistical models. Finally, they emphasize the integral role of deep learning in future research in this area.

Following an introduction to knowledge graphs (e.g. Freebase and Wikidata), Liu and Han’s study (Chapter 5) looks primarily at the application of three types of deep learning-based techniques to knowledge graphs: knowledge representation learning techniques; neural relation extraction techniques; and deep learning-based entity linking techniques. Similarly, Liu and Zhang’s study (Chapter 6) illustrates how deep learning-based methods can improve MT. Two categories of methods are identified: component-wise deep learning for MT and end-to-end deep learning for MT. The former aims to increase MT constituents’ capacity, while the latter aims to construct, with neural networks, intrinsic links between the source language and the target language on the basis of the encoder-decoder framework.

In Chapter 7, Liu and Feng focus on the deep learning application for question answering: specifically, deep learning in question answering over knowledge bases and deep learning in machine comprehension. In their study, neural networks are employed to comprehend and translate questions into structured queries or semantic representations and to build an end-to-end paradigm which can make the question answering system more straightforward.

Taking account of sentiment analysis in NLP, Tang and Zhang believe that deep learning methods can be employed in many aspects as robust computational models for identifying semantic representations (Chapter 8). Their study mainly covers sentiment-specific word embeddings (e.g. Collobert & Weston’s context-based model), sentence-level sentiment classification (e.g. the use of recurrent and recursive neural networks), document-level sentiment classification (e.g. Tang et al.’s neural network architecture), and fine-grained sentiment analysis (i.e. opinion mining, targeted sentiment analysis, aspect-level sentiment analysis, stance detection, and sarcasm detection).

Zhao and Li’s study (Chapter 9) reports on the latest developments of deep learning models regarding social computing and demonstrates how to fit deep learning techniques (e.g. shallow embedding-based and deep neural network-based methods) into a variety of diverse and flexible NLP tasks, such as user-generated content, social connections, and recommendation.

Chapter 10, by He and Deng, presents the use of deep learning technologies in natural language generation from images, a “cross realm” of computer vision and NLP. They discuss two major schemes: namely, the end-to-end framework and the compositional framework, and explain, with examples, the deep learning techniques to generate natural language captions which are used to demonstrate image contents.

The final chapter wraps up the book from the task-centric and representation-generic perspectives and outlines a number of major recent advances in deep learning for NLP and research frontiers in terms of compositionality for generalization, unsupervised learning, reinforcement learning, meta-learning and interpretability.

With respect to the contents, the book under review exhibits a variety of academic merits by addressing a number of important issues associated with deep learning. The emerging paradigm shift in NLP studies indicates that deep learning has become a robust methodology, playing an integral part in modeling various NLP tasks. Compared with other similar publications such as Neural Network Methods for Natural Language Processing (Goldberg Reference Goldberg2017), the book concentrates more on the integration of sophisticated neural networks and natural language models and elucidates for readers the recent advances in both deep learning and NLP by providing a holistic synthesis across different chapters. Given that the book is not intended to provide novel algorithms or models/frameworks, the contributors of each chapter simply choose a particular topic in NLP and comment on the achievements of previous studies in an attempt to confirm the validity of the deep learning methodology. In addition, the book is sequenced in a logical manner, providing useful examples and case studies to help readers achieve a better understanding. A total of more than 200 references at the end of each chapter also provides readers with a collection of valuable resources pertinent to NLP, with which readers could become self-motivated to explore, in breadth and depth, more potential applications of deep learning technology. This would mean that novice NLP learners who do not have sufficient knowledge about model construction would find this book extremely helpful and would have a clear way to familiarize themselves with fundamental NLP knowledge. Furthermore, since the fast-growing deep learning trend is regarded as a paradigmatic revolution, clarifying the interplay between deep learning and NLP appears to help build valid intrinsic links for natural language modeling. Such a forward step in methodological innovation would naturally motivate researchers to carry out more studies conjoining computer science and linguistics, helping expedite the development of AI.

Overall, in terms of what this book can offer NLP researchers, language and AI engineers, quantitative and computational linguistics researchers, and translation software designers, it appears that the aforementioned theoretical and methodological innovations fill an important gap in the literature. Such a holistic and thought-provoking edition would be greatly helpful for providing new algorithms, building new models, and exploring new applications. In this sense, the book is a great contribution to the research fields of NLP, quantitative linguistics and computational linguistics, and therefore, it deserves our wholehearted recommendation.