DisasterLex: An Expert Concept-to-Schema Knowledge Graph for Geospatial Reasoning in Disaster Analytics

DisasterLex is presented as a knowledge-graph-mediated framework designed to address the challenges of geospatial reasoning in disaster analytics, providing natural-language access to complex, structured tabular data that conventional text-to-SQL and retrieval methods fail to handle.

Problem Statement and Motivation

Disasters require rapid, accurate access to structured geospatial data (hazard exposure, population vulnerability, lifeline-infrastructure readiness) to coordinate response under time-critical decision constraints. However, the disaster domain is inherently complex: it spans dozens of heterogeneous tables with specialised semantics, where sentinel values mark missingness... and the causal relations a domain expert reasons over... must be encoded externally for the LLM to use. Current text-to-SQL systems are limited by assuming a small schema, while standard Retrieval-Augmented Generation (RAG) does not address the schema selection problem that gates structured query answering.

DisasterLex: Concept-to-Schema Linking via Expert Knowledge Graph

DisasterLex introduces an Expert Knowledge Graph (EKG) which serves as a concept-level schema-linking layer. This EKG is an expert-curated graph of domain concepts and typed causal edges mapped to executable database schemas. The system uses this graph to mediate between user vocabulary and column names during the table selection step, thereby reducing the complexity passed to the LLM.

The core mechanism involves:

1. Concept Matching: Using synonym-based concept matching plus 1-hop graph traversal against a curated set of domain concepts (107 concepts in the case study).

2. Schema Reduction: This process reduces the prompt schema context from 150 columns to typically 10–20 per query.

System Architecture and Components

The architecture is built upon three integrated components:

• Unified Knowledge Graph (DDCG): The graph is auto-introspected from a relational database (e.g., a DuckDB build of 36 geospatial tables, 150 columns). It includes DataTable nodes with child DataColumn nodes, and JoinRule nodes encode valid join patterns.

• Concept-Aware Schema Retrieval: This mechanism translates the natural-language query into the necessary subset of database schemas by traversing concept-to-schema edges from activated EKG concepts.

• Four-Stage Orchestration Pipeline: The system employs an operational structure based on the Incident Command System (ICS), rather than a single ReAct loop. This pipeline consists of four distinct stages:

1. Context & Criticality Extraction: Parses the query for its area of interest, hazard or topic, 1–5 criticality level.

2. Operational Domain Classification: Classifies the query into one of operational clusters (e.g., life-safety operations, damage assessment and response) based on domain-specific prompt templates.

3. Causal-Informed Planning: A ReAct agent that retrieves causal edges from the EKG and produces a structured analysis plan.

4. Grounded Execution: A second ReAct agent that executes the plan using a Text-to-SQL tool (concept-aware schema retrieval, LLM-based SQL generation) and a knowledge-graph tool (1-hop edge retrieval).

Contributions and Evaluation

DisasterLex makes three primary contributions:

1. The use of an EKG for concept-level schema linking to address the failure mode of prior text-to-SQL/RAG approaches.

2.The implementation of an ICS-motivated four-stage orchestration pipeline where each component is independently load-bearing.

3. The creation of a four-tier diagnostic benchmark (Table 1), which isolates distinct failure modes: R (routing), K (EKG grounding), M (multi-table SQL composition), and D (data disclosure).

The system was evaluated on a 75-query test set across seven base models. Results show that the full DisasterLex pipeline beats four state-of-the-art baselines... by 1.4× to 2.75×, with absolute scores of 1.65 to 3.56 (of 5.0). Error analysis indicates that baseline failures cluster in routing and multi-table SQL composition, the operations our orchestration explicitly addresses.

Conclusion

The DisastersLex framework demonstrates that the concept-to-schema layer plus orchestration matters more than the choice of retrieval substrate or text-to-SQL agent. The authors note that this pattern may transfer to other domains where expert causal knowledge sits alongside time-sensitive structured data, such as medical triage, supply-chain disruption, or infrastructure incident response.

Based on the analysis of the DisasterLex framework, I have identified several critical areas where its architecture can be generalized and enhanced to improve existing AI systems designed for complex, structured reasoning tasks.

The core strength of DisasterLex—the integration of an Expert Knowledge Graph (EKG) with a four-stage orchestration pipeline—is not merely a solution to the Text-to-SQL problem; it is a blueprint for Domain-Specific Reasoning Systems.

Below are specific improvements, followed by what the enhanced AI system will be capable of.

The following improvements extend DisasterLex from its current domain (Geospatial Disaster Analytics) to general applications requiring structured, causal reasoning over large, heterogeneous datasets.

Current State: The EKG is expert-curated and static relative to the database schema.

Improvement: Implement a Continuous Feedback Loop for Causal Edge Refinement.

• Integrate a validation module that monitors system outputs against real-world outcomes. If the EKG predicts flood occurrence INCREASES hospital operations (as seen in Table 6), but real-time data shows no correlation, this discrepancy is flagged.

• The AI system then initiates a Causal Re-Scout sub-process: it queries external knowledge bases (e.g, academic literature or real-time sensor feeds) to validate the existing EKG edge and either reinforce its weight or suggest a new edge (e.g., power failure INHIBITS hospital operations).

Resulting Improvement: Eliminates static knowledge gaps and allows the AI system to learn from operational reality, not just historical domain expertise.

Current State: The routing is ICS-motivated (Life-Safety, Damage Assessment, Infrastructure Mitigation).

Improvement: Abstract the Operational Domain Classification into a generalized framework for any structured workflow.

• Replace the hardcoded ICS clusters with a Modular Task Decomposition Engine. This engine uses LLMs not just to classify which domain applies (e.g., Logistics vs. Planning), but to identify and execute sub-tasks within that domain, even if they are not explicitly mapped in the EKG.

• The system routes based on Functional Similarity rather than rigid classification, allowing it to handle novel queries that blend concepts from multiple operational clusters (e.g., How does infrastructure failure impact supply chain logistics?).

Resulting Improvement: The AI system is no longer limited to pre-defined operational contexts; it can handle complex, multi-faceted inquiries across different industries (e.g., financial risk assessment, environmental impact analysis).

Current State: The execution stage uses a 3-attempt retry loop on SQL validation or execution errors.

Improvement: Implement Root Cause Schema Diagnostics.

• When an SQL query fails or returns unexpected results (e.g., zero rows when a count is expected), the system does not just retry. It triggers a Schema Diagnostic Agent. This agent analyzes the failed query against the EKG and identifies why it suggests a failure (e.g., "The join between HIFLD-EMERGENC-SHELTER-N and HP FLD 002 requires an intermediate HEX ID lookup table that was not included in the initial concept-to-schema retrieval, suggesting an a 3rd table is missing").

• This failure diagnosis allows for dynamic expansion of the schema context before re-generating the SQL.

Resulting Improvement: The AI system transitions from merely repairing failed queries to actively reconstructing them by understanding the underlying structural deficiencies in its knowledge base.

Current State: The system uses a criticality-gated recommendation flag and a data-availability disclosure.

Improvement: Implement Probabilistic Reasoning Integration.

• Modify the EKG edges to carry not just a confidence weight [0, 1], but also the type of uncertainty (e.g., Observed, Estimated, Modeled).

• The final LLM synthesis stage is forced to incorporate this uncertainty into its output. Instead of stating, There are 6 high-risk hospitals, it states, There are approximately 6 high-risk hospitals (based on the current flood model, plus or minus 10%). This is an estimated figure.

• Resulting Improvement: The AI system provides a quantitative measure of its own reliability for every critical output, making it safer and more trustworthy for high-stakes decision support.

By implementing these enhancements, the improved DisasterLex architecture transitions from a sophisticated QA tool to an Autonomous, Causal Reasoning Agent (CRA) capable of:

1. Real-Time Predictive Synthesis: Instead of simply answering What is the current status?, it can perform proactive analysis: "Given the current rate of riverine flooding (current data), and knowing that this increases runoff, which critical infrastructure nodes are projected to fail within the next 72 hours?"

2. Self-Correcting Knowledge Acquisition: The system will not just rely on its initial curated knowledge base. It will dynamically search, validate, and update its causal understanding of complex phenomena (e.g., how different climate change scenarios affect specific regional infrastructure).

3. Explainable Decision Support: Every recommendation provided by the CRA will be accompanied by a traceable, multi-step causal chain (A causes B which leads to C), allowing human analysts to verify the AI's reasoning against established domain expertise, rather than simply accepting a correct answer.

4. Handling Complex Ambiguity: It can manage queries that require combining data from disparate sources (e.g., integrating population density data with real-time sensor readings) and accurately report the limitations of its knowledge base when those sources conflict or are missing.