Natural Language Interfaces to Databases

This book presents a comprehensive overview of Natural Language Interfaces to Databases (NLIDBs), an indispensable tool in the ever-expanding realm of data-driven exploration and decision making. After first demonstrating the importance of the field using an interactive ChatGPT session, the book explores the remarkable progress and general challenges faced with real-world deployment of NLIDBs. It goes on to provide readers with a holistic understanding of the intricate anatomy, essential components, and mechanisms underlying NLIDBs and how to build them. Key concepts in representing, querying, and processing structured data as well as approaches for optimizing user queries are established for the reader before their application in NLIDBs is explored. The book discusses text to data through early relevant work on semantic parsing and meaning representation before turning to cutting-edge advancements in how NLIDBs are empowered to comprehend and interpret human languages. Various evaluation methodologies, metrics, datasets and benchmarks that play a pivotal role in assessing the effectiveness of mapping natural language queries to formal queries in a database and the overall performance of a system are explored. The book then covers data to text, where formal representations of structured data are transformed into coherent and contextually relevant human-readable narratives. It closes with an exploration of the challenges and opportunities related to interactivity and its corresponding techniques for each dimension, such as instances of conversational NLIDBs and multi-modal NLIDBs where user input is beyond natural language. This book provides a balanced mixture of theoretical insights, practical knowledge, and real-world applications that will be an invaluable resource for researchers, practitioners, and students eager to explore the fundamental concepts of NLIDBs. Foreword by the Series Editor Preface Acknowledgements Contents 1 Overview [DELETE] 1.1 A Session with ChatGPT 1.2 NLIDBs in the Wild 1.2.1 Language Support 1.2.2 Assisting Query Input 1.2.3 Expressivity 1.2.4 Customizability 1.3 What's Ahead 2 Building an NLIDB: The Basics 2.1 Example Database 2.2 Anatomy of NLIDB 2.3 Building an NLIDB 2.3.1 Query Understanding 2.3.2 Query Translation 2.3.3 External Knowledge 2.3.4 Interaction Generation 2.3.5 Result Generation 2.4 Summary 3 Data and Query Model 3.1 Conceptual Models 3.1.1 Entity-Relationship Model 3.1.2 UML 3.1.3 Ontology 3.2 Relational Model 3.2.1 Relational Query Languages 3.2.2 First-Order Logic 3.2.3 Relational Algebra 3.2.4 SQL 3.3 Graph Model 3.3.1 SPARQL 3.4 Storage and Indexing 3.5 Query Evaluation and Optimization 3.5.1 Query Parsing and Plan Generation 3.5.2 Query Optimizer 3.5.3 Evaluation Engine 3.6 Summary 3.7 Further Reading 4 Text to Data 4.1 Introduction 4.1.1 Natural Language Understanding and Natural Language Generation 4.1.2 Historical Overview 4.1.3 Semantic Parsing 4.2 Meaning Representations 4.2.1 First-Order Logic 4.2.2 Lambda Calculus 4.2.3 Abstract Meaning Representation 4.2.4 Word Embeddings 4.2.5 Semantic Compositionality 4.2.6 Knowledge Graphs and RDF 4.2.7 Syntactic Representations 4.2.8 Combinatory Categorial Grammar 4.3 Converting Sentences to Structured Form 4.3.1 Information Extraction 4.3.2 GeoQuery 4.3.3 Semantic Parsing 4.3.4 Semi-supervised Semantic Parsing 4.4 Neural Semantic Parsing 4.4.1 Sequence-to-Sequence Methods 4.4.2 Applications of Neural Semantic Parsing 4.5 Text-to-SQL 4.5.1 WikiSQL 4.5.2 Dataset Splits 4.5.3 Spider 4.5.4 Extensions to Spider 4.5.5 Selective Recent Papers 4.6 Summary 4.6.1 Further Reading 5 Evaluation 5.1 Methodology Overview 5.2 Datasets and Benchmarks 5.2.1 ATIS 5.2.2 GeoQuery 5.2.3 Scholar 5.2.4 Academic 5.2.5 Advising 5.2.6 IMDB and Yelp 5.2.7 Fiben 5.2.8 WikiSQL 5.2.9 SPIDER 5.2.10 BIRD 5.2.11 Benchmarks Statistics and Query Composition 5.2.12 Other Benchmarks 5.3 Reference-Based Evaluation 5.3.1 Generating a Reference 5.3.2 Candidate Answer Evaluation Based on a Reference 5.3.3 Performance Metrics 5.3.4 Semantic Equivalence 5.4 Human-Centric Evaluation 5.4.1 Expressing and Performing the Tasks 5.4.2 Quality of Generated Queries 5.4.3 Performance in Downstream Tasks 5.5 Other Performance Metrics 5.5.1 Resource Consumption 5.5.2 Query Hardness 5.5.3 Robustness to Noise and Ambiguity 5.5.4 Adaptability to Unseen Databases and Questions 5.6 Top Performing Models 5.6.1 The Models 5.6.2 Large Language Models 5.6.3 Constraining the Decoder 5.7 Further Reading 6 Data-to-Text 6.1 Introduction 6.1.1 Traditional Generation 6.1.2 Data-to-Text Generation 6.1.3 Abstract Meaning Representation (AMR) for Text Generation 6.1.4 Neural Generation 6.2 Domain Specific Table-to-Text 6.2.1 SRST 6.2.2 e2e 6.2.3 WebNLG 6.2.4 WikiBio 6.2.5 RotoWire 6.2.6 WikiTableT 6.3 Domain Independent Table-to-Text 6.3.1 ToTTo 6.3.2 DART 6.3.3 FeTaQA 6.3.4 TabFact 6.3.5 LogicNLG 6.3.6 Logic2Text 6.3.7 GEM 6.4 Pretraining for Tables 6.4.1 TURL 6.4.2 TUTA 6.4.3 TAPAS 6.4.4 TaBERT 6.4.5 Grappa 6.4.6 Tabbie 6.4.7 Other Recent Papers 6.5 Summary 7 Interactivity 7.1 Disambiguation 7.1.1 Ambiguity 7.1.2 Spell Correction 7.1.3 Interactive Disambiguation 7.2 Query Suggestion 7.2.1 Auto-Completion 7.2.2 Query Suggestions Beyond Auto-Completion 7.3 Automatic Data Insights 7.3.1 Categorization 7.3.2 Visualization Recommendation 7.4 Explanation 7.5 Conversational Natural Language Interfaces to Databases 7.5.1 Discourse Structure 7.5.2 Discourse Transition 7.5.3 Unidirectional Conversation 7.5.4 Bidirectional Conversation 7.6 Multi-modal Conversational NLIDB 7.6.1 Conversational Transition Modeling 7.6.2 Conversation via Query Suggestion 7.6.3 Discussions 7.7 Summary 7.8 Further Reading Correction to: Natural Language Interfaces to Databases Correction to: Y. Li et al., Natural Language Interfaces to Databases, Synthesis Lectures on Data Management, https://doi.org/10.1007/978-3-031-45043-3 Index