natural language understanding

This paper addresses the limitations of traditional pre-trained language models (PLMs) in text-to-SQL tasks under the large language model (LLM) era—namely, poor generalization, high generation error rates, and prohibitive adaptation costs. We systematically survey LLM-driven natural language-to-SQL generation techniques. We propose the first structured, knowledge-graph-inspired survey framework and formally characterize the paradigm shift from PLM fine-tuning to emerging approaches: prompt engineering, retrieval-augmented generation (RAG), database-schema-aware encoding, multi-step reasoning, and in-context learning. We comprehensively catalog mainstream benchmarks, evaluation metrics, and technical challenges, with particular emphasis on critical open issues including scalability and robustness. Our work provides researchers with a clear evolutionary trajectory and practitioners with a reusable technology roadmap and concrete directions for future advancement.

Large language models (LLMs) suffer from unpredictability, opacity, and limited controllability. Method: This paper introduces “representation engineering”—a novel paradigm that identifies and edits semantic concept directions (e.g., honesty, harmfulness) in high-level representation spaces via contrastive input probing, enabling interpretable and intervention-based behavioral control. Contribution/Results: We formally define the paradigm’s objectives, scope, and methodology, rigorously distinguishing it from mechanistic interpretability, prompt engineering, and fine-tuning. We propose a unified framework integrating contrastive analysis, concept-level representation editing, high-dimensional causal intervention, and interpretability evaluation. This framework supports controllable, safe, and dynamically adaptive LLM governance, reveals critical challenges—including performance degradation and controllability collapse—and charts a technical pathway toward predictable, secure, and personalized LLMs.

Existing text-to-SQL methods exhibit significant performance variance in cross-database generalization, primarily due to the lack of systematic alignment between domain semantics embedded in natural language queries and structural patterns in database schemas, compounded by inefficient, non-generalizable manual prompt engineering for domain knowledge injection. Method: We propose a structured-domain-knowledge-based multi-database text-to-SQL framework that explicitly models domain knowledge as retrievable, structured statements; employs lightweight substring matching for database-adaptive retrieval; and seamlessly integrates retrieved knowledge into the LLM’s reasoning pipeline—eliminating reliance on handcrafted prompts. Contribution/Results: Evaluated across 11 real-world databases and 5 open-source and commercial LLMs, our approach achieves substantial gains in SQL execution accuracy over strong baselines. It is the first to enable plug-and-play cross-database transfer of domain knowledge, markedly improving model robustness in understanding semantic correspondences between domain vocabulary and schema elements.

This work addresses the limited effectiveness of traditional retrieval-augmented generation (RAG) approaches in querying structured, non-textual data—particularly numerical and highly structured datasets—using natural language. To overcome this, the authors propose an open-source solution that leverages a synthetic question-answering pair generation pipeline to create diverse training data capturing both user intent and data semantics. They fine-tune the DeepSeek-R1-Distill-8B model using 4-bit quantized low-rank adaptation (QLoRA) to generate executable query statements. Experiments on a public service accessibility dataset from the Durangaldea region in Spain demonstrate high accuracy across monolingual, multilingual, and unseen-location scenarios, confirming that small, domain-specific models can effectively and generalizably support natural language access to structured data without relying on large, closed-source models.

This work addresses the challenge of compiling natural language queries into backend query languages in document-centric, hybrid, and heterogeneous data environments, where semantic intent is often ambiguous or incomplete. The authors propose the NLIQ framework, which introduces a “goal sufficiency” criterion to classify queries according to their semantic determinacy. It emphasizes that when intermediate goals must be dynamically constructed, intermediate representations should serve as core semantic objects rather than mere syntactic intermediaries. Through conceptual analysis, case modeling, and formal categorization, the study establishes a unified query paradigm that integrates goal recognition, intermediate representation design, and heterogeneous execution. This framework provides a theoretical foundation for natural language querying in complex data settings and opens new research directions in semantic goal construction, heterogeneous compilation, and answer generation.

This work addresses the limitations imposed by the inherent constraints of natural language expressions on the reasoning capabilities of large language models in complex tasks. It proposes that designing structured linguistic representations can effectively construct and activate internal cognitive schemata within these models, thereby enhancing their intelligent performance. For the first time, the study formalizes linguistic representation design as a critical pathway for extending the capabilities of large language models, integrating insights from linguistics and symbolic representation theory. Through controlled experiments, it systematically analyzes how different representational forms influence internal model activations and outputs. The findings demonstrate that optimizing linguistic representations—without altering model parameters or scale—can significantly improve task performance, thereby validating the efficacy and potential of this approach.

This paper systematically surveys large language model (LLM)-driven Text-to-SQL techniques to lower the barrier for non-experts accessing relational databases. Methodologically, it employs bibliometric analysis, taxonomic comparison of approaches, meta-analysis of benchmark datasets, and critical cross-benchmark evaluation. It establishes, for the first time, a multidimensional methodology framework that clarifies paradigmatic evolution—from prompt engineering and supervised fine-tuning to reasoning augmentation—and identifies persistent evaluation blind spots. The study distills four major technical branches and five core challenges, proposing a unified conceptual evaluation framework. Furthermore, it introduces a structured knowledge graph and 12 scalable research directions, offering both theoretical guidance and a practical roadmap for the community. (136 words)

This work proposes a novel approach to address the challenge novice modelers often face in ensuring semantic alignment between domain models and textual specifications during early software engineering phases. The method first employs natural language processing to preprocess specification texts and generates human-authored natural language descriptions for each model element. It then leverages a large language model (LLM) to compare these descriptions against the original specifications, automatically classifying their alignment status as aligned, misaligned, or uncertain, while providing interpretable evidence for each judgment. By uniquely integrating LLM capabilities with human-crafted model descriptions, the approach achieves high-precision semantic alignment verification, demonstrating near-perfect precision (≈100%) and 78% recall across multiple domain datasets. Individual element analysis requires between 18 seconds and one minute, indicating strong potential for integration into modeling tools.

Automatically aligning unstructured text—such as clinical case notes—with diverse, heterogeneous database schemas remains challenging due to rigid schema-binding assumptions in existing approaches. Method: We propose a model-agnostic text structuring method grounded in an attribute grammar metamodel (MetaG), which constructs semantically enriched parse trees. Through iterative tree rewriting and extraction, the method jointly generates executable semantic schemas and their corresponding instances—without requiring prior knowledge of target data models. Contribution/Results: Our key innovation lies in decoupling text understanding from database schema design, enabling seamless adaptation to heterogeneous storage backends—including relational, graph, and document databases. Empirical evaluation on clinical texts demonstrates accurate generation of cross-model-compatible schemas and instances, significantly overcoming limitations of conventional schema-bound paradigms. The approach achieves robust schema generalization and instance fidelity across disparate target models, establishing a foundation for flexible, database-agnostic text-to-structure translation.

Text-to-structured generation (e.g., tables, knowledge graphs, charts) for agent-centric AI is a foundational infrastructure enabling context-aware retrieval and autonomous reasoning, yet suffers from fragmented methodologies, scarce standardized datasets, and inconsistent evaluation protocols. Method: We conduct a systematic literature review integrating techniques from NLP, information extraction, knowledge representation, and machine learning to establish the first holistic analytical framework—comprising task taxonomy, benchmark dataset inventory, and unified evaluation metrics. Contribution/Results: We introduce the first general-purpose evaluation framework for structured output generation, explicitly identifying methodological limitations and core challenges (e.g., fidelity, composability, and reasoning-aware assessment). We comprehensively map research gaps and affirm the centrality of this direction in next-generation AI systems, providing both theoretical grounding and practical guidance for future algorithmic development and empirical validation.

Current text-to-SQL system evaluations rely on a single static database, which fails to capture model robustness across diverse data instances and may introduce significant bias. This work proposes SynSQL, a novel framework that leverages large language models to directly generate semantically consistent and schema-aligned relational test data from natural language questions. SynSQL formulates database construction as a structured generation task governed by semantic and relational constraints, comprising three stages: schema selection, question-guided data synthesis, and constraint-aware iterative refinement. Experiments on Spider, BIRD, and Spider 2.0 demonstrate that databases generated by SynSQL reduce the performance of ten state-of-the-art models by 3–14%, effectively uncovering errors masked by static evaluation and substantially enhancing assessment reliability and stress-testing capability.

To address the overlooked problem of database intent identification in text-to-SQL tasks under multi-database settings, this paper proposes the first end-to-end joint modeling framework that explicitly incorporates a database selection stage into the conventional text-to-SQL pipeline, enabling collaborative optimization of db_id prediction and SQL generation. Methodologically, the approach integrates large language model (LLM)-based semantic understanding, a fine-tuned RoBERTa encoder, structured prompt engineering, and a critic-agent-driven SQL correction mechanism. Crucially, it innovatively leverages LLMs to extract implicit database clues from natural language queries, thereby enhancing the robustness of database intent recognition. Evaluated on standard multi-database benchmarks, the framework achieves significant improvements over existing state-of-the-art models in both database selection accuracy and SQL execution accuracy.