query understanding

To address the challenges of ambiguous intent definitions, high annotation costs, and data scarcity in low-resource language intent classification, this paper proposes a zero-shot retrieval-based intent identification method. Instead of relying on explicit intent labels, the approach models each intent as a set of historically annotated queries and performs intent inference via dense retrieval—matching an input query to its most semantically similar labeled queries in a shared latent space. The method employs a multilingual semantic encoder to produce cross-lingually aligned query embeddings, enhances semantic consistency through contrastive learning, and accelerates retrieval using approximate nearest neighbor (ANN) search. Evaluated on eight low-resource languages, it achieves an average F1 score of 72.4%, outperforming zero-shot fine-tuning baselines by 18.6 percentage points; remarkably, it attains near fully supervised performance using only 10% of the labeled data. This work is the first to reformulate intent classification as an annotation-free, cross-lingual similarity search task.

To address the challenge non-expert users face in efficiently querying and analyzing large tabular data, this paper proposes an end-to-end framework that translates natural language queries into executable query plans using large language models (LLMs). Unlike conventional SQL generation approaches, our method employs iterative semantic parsing to map natural language into heterogeneous operation sequences—including statistical and machine learning primitives (e.g., PCA, anomaly detection)—and executes them externally to the database, thereby circumventing LLM context-length limitations and the overhead of full-data loading. Experiments on standard benchmarks and large-scale scientific tabular datasets demonstrate substantial improvements in task completion rate and execution efficiency for complex analytical tasks. The framework supports flexible, scalable data analysis beyond the expressive capacity of SQL, offering a novel pathway for the NL2Data analysis paradigm.

This paper addresses the failure of conventional equivalence-based query rewriting when original data tables are inaccessible due to access control policies, privacy constraints, or prohibitive retrieval costs. To overcome this, we propose INQURE, a semantic intent-preserving query rewriting framework. Unlike traditional approaches relying on syntactic equivalence and query plan optimization, INQURE introduces, for the first time, large language model (LLM)-driven intent understanding and cross-table reconstruction—enabling semantically consistent rewriting across structurally heterogeneous and non-aligned schemas. The system incorporates pre-filtering of candidate tables, pruning heuristics, and learned ranking to form an end-to-end rewriting pipeline. Evaluated on a benchmark spanning 900+ real-world database schemas, INQURE demonstrates superior rewriting quality and practical utility. A user study further confirms its effective trade-off between execution feasibility and fidelity of analytical insights.

In criminal investigations, manual analysis of instant messaging data (e.g., WhatsApp) incurs high labor costs and suffers from low efficiency in evidence triage. To address this, we propose a knowledge graph–enhanced multimodal NLP framework that unifies semantic representations across message text, speech (via ASR), and metadata through end-to-end named entity recognition, automatic speech recognition, and dynamic knowledge graph construction. The framework enables auditable graph queries grounded in original evidentiary artifacts, interactive visual exploration, and semantics-aware search. A prototype system implementing this approach has been deployed in real-world investigations, demonstrating significant improvements in identifying key persons, events, and relational patterns—validated by frontline prosecutors. This work advances explainable, verifiable AI for digital forensics, bridging the gap between interpretable reasoning and operational legal requirements.

To address the challenges of ambiguous user queries and insufficient intent feedback in exploratory search—leading to iterative trial-and-error—we propose, for the first time, the generative query intent description task. Our method jointly models relevant and irrelevant documents to automatically generate precise, interpretable intent descriptions. Technically, we introduce a novel dual-space modeling paradigm: semantic separation of relevant and irrelevant documents in a representation space, and explicit suppression of irrelevant information in a disentangled space. We design a semantic projection encoder, a semantic disentanglement decoder, and an attention-guided generation framework. Experiments on benchmark datasets demonstrate that our approach significantly outperforms existing intent classification, clustering, and query summarization methods. Attention visualization confirms its effectiveness in filtering out distracting topics, yielding more accurate and interpretable intent descriptions.

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.

Traditional SQL struggles with semantic matching and unstructured text analysis in natural language queries, while existing LLM-augmented approaches require users to manually construct complex pipelines. This work proposes SEMA-SQL, the first system to seamlessly integrate the semantic capabilities of large language models (LLMs) into a declarative query language via user-defined functions. By introducing Hybrid Relational Algebra (HRA), SEMA-SQL unifies relational operations with semantic operators, enabling automatic translation of natural language into efficient hybrid queries. The system leverages in-context learning for query generation, cost-based optimization, and intelligent batching of LLM invocations, significantly enhancing both expressiveness and execution efficiency on standard and extended benchmarks. Notably, it reduces average LLM calls for semantic join operations by 93%.

Resource-constrained edge devices face challenges in accurately understanding user intent from UI interaction traces, while simultaneously ensuring privacy preservation and real-time responsiveness. Method: This paper proposes a two-stage decomposed architecture: (1) generating structured sequential summaries of interaction behaviors, followed by (2) lightweight intent inference based on these summaries. The approach integrates context-aggregated enhancement and task-adaptive fine-tuning to strengthen semantic modeling capabilities of small models. Contribution/Results: Experimental results demonstrate that, under identical privacy guarantees and low-latency constraints, the proposed method achieves higher intent recognition accuracy than state-of-the-art large multimodal language models. It establishes an efficient, privacy-aware, and real-time interaction understanding paradigm for on-device intelligent agents.

Natural language queries in table-based question answering (TQA) exhibit inherent ambiguity, yet existing research fails to systematically distinguish among three distinct query types—unambiguous, ambiguously cooperative, and non-cooperative—leading to biased evaluation. Method: This paper introduces the novel paradigm of “collaborative ambiguity,” reframing ambiguity as an interactive feature reflecting shared human-AI responsibility rather than a flaw, and proposes a query classification framework grounded in Grice’s Cooperative Principle. Through theoretical modeling and empirical analysis, we conduct systematic annotation and evaluation across 15 mainstream TQA benchmarks. Contribution/Results: Our study is the first to reveal pervasive query-type mixing across these datasets and demonstrates that current evaluation protocols conflate system execution accuracy with semantic understanding capability. The findings establish a theoretical foundation and practical methodology for developing interpretable, robust TQA interfaces and type-aware evaluation standards.

This work addresses the challenge of accurately interpreting complex user intents expressed in natural language within traditional e-commerce search systems. We propose a semantic search framework that integrates large language models (LLMs) to jointly optimize an embedding model and a structured constraint generation module, enabling precise intent extraction from conversational queries and facilitating synergistic semantic retrieval with attribute-based filtering. To mitigate the scarcity of labeled data, we leverage LLMs to generate high-quality synthetic data for embedding fine-tuning and devise an end-to-end training strategy. Experimental results on real-world e-commerce datasets demonstrate that our approach significantly outperforms existing baselines, achieving notable improvements in both retrieval precision and recall.