Schema-Aware Reference as Prompt (RAP)

• The paper introduces RAP, which dynamically retrieves and injects schema-aware references into model prompts to bridge the semantic gap between natural language and structured outputs.
• It leverages techniques such as schema paraphrasing, learnable soft embeddings, and context-aware retrieval to enhance performance across multi-domain tasks.
• Empirical results demonstrate significant improvements, including 3–5 F1 gains in event extraction and up to 82.6% execution accuracy in text-to-SQL tasks.

Schema-Aware Reference as Prompt (RAP) is a methodological paradigm in which structured schema elements—sometimes paired with annotated instances or latent representations—are dynamically retrieved and injected as prompt material for language and vision models. This approach exploits explicit schema knowledge and task-annotated references to bridge the semantic gap between unstructured natural language and pre-defined output schemas, enabling models to generalize more robustly across tasks, domains, and low-resource settings. RAP encompasses diverse instantiations ranging from retrieval-augmented generation in event extraction and knowledge graph construction to learnable schema-conditioned prompting in multi-task learning, context-aware schema matching, dialogue state tracking, and even structured visual metaphor transfer.

1. Formal Definition and Core Mechanics

At its core, RAP operates by selecting a subset of schema-driven reference material—such as instance-annotated examples, subgraphs of type definitions, or learnable schema-specific embeddings—and concatenating or conditioning this material as a prompt within the input to a language or vision model. The general process consists of:

• Schema Specification: Let $\mathcal{G}$ be a schema, typically a labeled graph or compositional structure where nodes denote entity types, event roles, columns, slots, or visual components.
• Reference Pool Construction: A datastore $\mathcal{D}$ is constructed comprising entries $(c_i, y_i, p_i)$, where $c_i$ is a text (or image) context, $y_i$ is a structured label, and $p_i \subseteq \mathrm{Nodes}(\mathcal{G})$ is a set of schema pointers (Yao et al., 2022, Liang et al., 13 May 2025).
• Retrieval Policy: At inference, a scoring or retrieval function $$r: (\mathrm{query}, \mathcal{D}) \to \{\mathcal{Z}\}$$ selects the top-$k$ references based on embedding similarity, keyword overlap, or task-specific measures.
• Prompt Construction: Retrieved schema-aware references, optionally paraphrased or further filtered by relevance or difficulty metrics, are interleaved or serialized with the query input to produce a dynamic prompt $\mathcal{P}$.
• Model Conditioning: The model—whether sequence-to-sequence, encoder-decoder, or multi-modal—consumes the prompt, leveraging both explicit schema context and analogical cues for downstream prediction or generation.

RAP is model-agnostic; it can be instantiated as retrieval-augmented prompting (Yao et al., 2022, Liang et al., 13 May 2025), soft schema embeddings (“unified schema prompt” (Zhong et al., 2022)), GNN-encoded schema graphs (Su et al., 2023), or hierarchical tree- and group-based schemas for evidence packing (Chen et al., 28 Jan 2026).

2. RAP in Retrieval-Augmented Structured Prediction

In knowledge graph construction and event extraction, RAP enables models to condition on both schema definitions and annotated examples by retrieving and concatenating relevant references at inference time (Yao et al., 2022). The methodology encompasses:

• Reference Store: Human-annotated or weakly-supervised instances are aligned with schema nodes and indexed for retrieval.
• Retrieval: For input $X$, retrieve references based on BM25 or dense similarity (e.g., via text embeddings).
• Prompt Assembly: The model prompt includes definitions of event/relation types, argument roles, triggers, and exemplar sentences. See formulas:

$$p(Y|X) \approx \sum_{Z\in \mathrm{top}\text{-}k\, r(X;\mathcal{D})} p_\eta(Z|X)\, p_\theta(Y|X,Z)$$

• Model Integration: Prompts are concatenated with inputs for both generative (e.g., BART) and classification-based architectures (e.g., PRGC), with standard cross-entropy losses for structured prediction.

Experimental evidence demonstrates substantial performance gains for RAP-augmented models in low-resource regimes, with improvements of 3–5+ F1 points in both event and relation extraction compared to non-RAP baselines (Yao et al., 2022).

3. Schema Paraphrasing and Retrieval-Augmentation

RAP is extended in high-cardinality event extraction by introducing schema paraphrasing and dense retrieval (Liang et al., 13 May 2025):

• Schema Paraphrasing: Each canonical schema is paraphrased via frozen LLMs, producing a diverse pool of schema references that bridge terminology mismatches.
• Retrieval-Augmentation: At inference, only the top-$k$ schema paraphrases—scored by embedding similarity with the input—are selected, allowing scaling to hundreds of schemas within context windows.
• Prompt Formatting: The prompt alternates between “Schemas: [references]” and “Text: [query]”, efficiently guiding schema-constrained extraction.
• Model Training: Both standard generative and schema-conditioned losses are used to encourage schema adherence.

In the Multi-Dimensional Schema-aware Event Extraction (MD-SEE) benchmark, RAP achieves significant gains in Recall@10 (0.78 with BGE-M3) and E2E-F1 (0.68 with Llama3.1-8B), outperforming other retrieval and generation strategies (Liang et al., 13 May 2025).

4. Unified Schema Prompting and Learnable Schema Embeddings

RAP can be realized via learnable, schema-specific soft embeddings integrated as prompts in pre-trained LLMs, as in Unified Schema Prompt (SchemaPro) (Zhong et al., 2022):

• Schema Decomposition: For task $A$, the schema is $C_A = \{c_1, c_2, ..., c_n\}$ where $c_i = (\text{key}_i, \text{value}_i)$.
• Prompt Parameterization:
  • Key Prompts $\{K_t\}$: Learnable matrices for each component type.
  • Value Prompts: Format prompts $F_f$, task prompts $T_A$, and output prompts $O_o$ for task-attribute fields.
• Prompt Sequence Construction: Schema-tokenized inputs are constructed as $X = [c_1; c_2; ...; c_n]$, with learned prompt vectors prepended for format, task, and output.
• Pretraining and Adaptation: Pretrained with joint optimization over both model parameters and prompt embeddings. For unseen tasks, adaptation tunes only the task- (and optionally type-) specific prompts.
• Empirical Performance: SchemaPro achieves marked improvements over natural-language prompts in zero-shot (49.15% vs. 40.86%) and few-shot (56.96% vs. 50.32%) settings on 16 unseen tasks, with ablations confirming the need for each prompt type (Zhong et al., 2022).

5. RAP in Structured Schema Selection, Filtering, and Context Packing

RAP is leveraged for scalable schema selection and context-efficient prompting in large schema spaces for tasks such as text-to-SQL and schema matching.

• Schema Pruning and Hardness Prompting (RH-SQL):
  • Refined Schema Extraction: A relevance-ranking network selects up to four tables and five columns per query, significantly reducing prompt length.
  • Difficulty Classification: A lightweight classifier predicts SQL hardness (Easy–Extra-hard), encoding the result as a special token in the prompt.
  • Combined Prompt: The seq2seq model receives a sequence: hardness token, query, refined schema (as “table.column” tokens) (Yi et al., 2024).
  • Performance: RH-SQL achieves 82.6% execution accuracy on Spider (with NatSQL), and halves training time and storage versus alternatives.
• Context-Aware Schema Matching (ConStruM):
  • Context Tree: Multi-level tree structures represent schema hierarchies, enabling fine-grained evidence selection under prompt budget constraints.
  • Groupwise Differentiation via Hypergraphs: Highly similar columns are grouped, and “differentiation cues” are generated to disambiguate candidates.
  • Prompt Generation: For each matching instance, context packs for source and target columns, together with group summaries and difference cues, are injected into the LLM prompt.
  • Effectiveness: ConStruM achieves Top-1 accuracy up to 0.935 in context-stress settings, substantially outperforming unstructured baselines (Chen et al., 28 Jan 2026).

6. RAP for Schema-Driven Prompting in Multi-Domain Tasks and Vision

• Dialogue State Tracking (SHEGO): Domain schema is encoded as a graph; slot relationships are embedded using GNNs, yielding continuous “graph prompts” concatenated with standard input (Su et al., 2023). Joint tuning of GNN and prompt tokens allows parameter-efficient and domain-adaptive prompting, improving Joint Goal Accuracy by several points on standard benchmarks.
• Visual Metaphor Transfer: Schema Grammar G, a formal grammar over visual elements (subject, carrier, generic relations, style attributes), is extracted and transferred as a prompt to guide image generation, supporting cross-domain logic in Visual Metaphor Transfer tasks (Xu et al., 1 Feb 2026). Specialized agents automate schema construction, transfer via relational invariants, prompt encoding, and hierarchical critique.

7. Ablation Studies, Analysis, and Limitations

• Empirical Evidence: Ablations consistently show that injecting schema-aware references or schema-conditioned embeddings improves generalization—especially in low-resource regimes and cross-domain transfer (Yao et al., 2022, Zhong et al., 2022, Liang et al., 13 May 2025, Su et al., 2023). Key trends include:
  • Performance degrades notably with the removal of schema or instance cues from prompts.
  • Compositionality and modular schema-prompting induce strong zero- and few-shot performance (Zhong et al., 2022).
• Limitations: RAP methods incur computational cost in retrieval and prompt assembly. Weak supervision may introduce noise. Schema evolution may require costly updates to reference pools. Excessive prompt length or over-retrieval can add noise and degrade performance.
• Generalization: RAP is model-agnostic and extensible to new tasks, including semantic parsing, code generation, multi-modal reasoning, and structured vision tasks, provided the target output or process can be decomposed schemaically and referenced at inference (Yao et al., 2022, Liang et al., 13 May 2025, Xu et al., 1 Feb 2026).

8. Summary Table: RAP Applications and Methodologies

Domain/Task	Reference Mechanism	Key Model or Paper
Event Extraction	Instance/schema reference retrieval	(Liang et al., 13 May 2025, Yao et al., 2022)
Knowledge Graph Constr.	Schema-augmented reference prompt	(Yao et al., 2022)
Multi-Task NLP	Learnable schema soft embeddings	(Zhong et al., 2022)
Dialogue State Tracking	GNN-encoded schema graph prompts	(Su et al., 2023)
Text-to-SQL	Schema pruning + difficulty prompt	(Yi et al., 2024)
Schema Matching	Context tree + group cues in prompt	(Chen et al., 28 Jan 2026)
Visual Metaphor Transfer	Grammar-extracted schema prompt	(Xu et al., 1 Feb 2026)

These developments collectively establish RAP as a versatile, schema-grounded prompting framework with proven advantages in data efficiency, task generalization, and context-sensitive model performance across textual and visual modalities.