Extensive empirical research has quantified the effects of rainfall on traffic operations. Tsapakis et al. [6] demonstrated that the impact of rainfall on macroscopic urban travel times is intensity-dependent, with empirical and meta-review studies confirming speed reductions of 2%–17% and capacity reductions of 4%–32% under varying rainfall intensities on freeways and urban roads [7–9]. Pregnolato et al. [3] developed a depth-disruption function relating flood depth to vehicle speed, and Ni et al. [4] demonstrated that waterlogging amplifies traffic impacts by 1.3–2.0 times compared to rainfall alone. Despite this extensive body of quantitative knowledge, these findings have rarely been systematically integrated into automated scenario generation pipelines for disaster preparedness. Moreover, data scarcity remains a pervasive challenge across traffic and disaster-response applications, particularly when conventional data collection becomes difficult or unsafe under extreme weather conditions.

Large language models (LLMs) such as GPT-3 [10] and Llama 2 [11] have demonstrated strong text generation capabilities. However, as comprehensively reviewed by Moradi et al. [12], LLM-generated outputs in specialized domains frequently suffer from hallucination (producing fluent but factually inaccurate content [5]), and current alignment techniques remain insufficient for domain-critical applications such as disaster management. Small language models (sLLMs), typically under 10 billion parameters, offer a practical alternative through domain-specific fine-tuning, enabling local deployment without cloud dependency. Recent work on instruction-tuned models such as Flan-T5 [13] and Qwen2.5 [14] has shown that efficient fine-tuning techniques like QLoRA [15] can achieve competitive performance at substantially reduced computational costs. Shen et al. [16] further demonstrated that layer type is a more significant determinant than layer depth when optimizing fine-tuning in quantized LLMs, providing practical guidance for parameter-efficient strategies in resource-constrained settings. However, fine-tuning requires sufficient domain-specific training data, which remains a limiting factor for specialized applications.

Knowledge Graphs have emerged as a promising approach for structured data augmentation and grounded text generation. Lewis et al. [17] proposed Retrieval-Augmented Generation (RAG), an approach that integrates parametric language model memory with non-parametric retrieved knowledge to ground language model outputs and reduce factual errors in knowledge-intensive tasks. Unlike RAG, which retrieves external knowledge at inference time and requires runtime access to a graph store, the present approach distills KG knowledge into model parameters via fine-tuning. This design enables offline operation during disaster events when network connectivity may be compromised. In domain-specific applications, Liu et al. [18] demonstrated that combining template-generated synthetic data with fine-tuned LLMs significantly improves domain-specific information extraction under data scarcity, and Chen et al. [19] applied knowledge graph-guided LLM reasoning to emergency decision-making, demonstrating its applicability in disaster-response contexts. While these studies demonstrate the potential of KG-driven text generation, none have applied this approach to urban flood-traffic scenario synthesis or systematically evaluated its effectiveness for fine-tuning sLLMs across different model scales.

The study area is Jinju City, South Gyeongsang Province, South Korea, which experiences recurring urban flooding driven by monsoon rainfall and the Nam River (Namgang) system (Fig. 2). The city’s topography (characterized by low-lying areas along the Namgang and Yeongcheongang rivers, densely developed commercial districts, and multiple bridge crossings) creates spatially concentrated flood vulnerabilities that frequently disrupt road networks.

Two types of flood-related textual data were collected: (1) 89 news articles from BIG KINDS, a Korean news archive platform, covering January 2001 to December 2024; and (2) 165 disaster alert messages from the Korean Disaster and Safety Data Sharing Platform, covering September 2023 to December 2024.

A two-stage filtering process was applied in Stage 1, a hierarchical keyword filter selected documents containing strong flood-related terms (flood, inundation, heavy rain, torrential rain, overflow, water level, dam discharge), reducing the corpus from 254 to 138 documents. Strong flood keywords overrode exclusion keywords (collision, festival, election, COVID-19, earthquake), while documents matching only supporting keywords (road closure, detour, river, levee, drainage) were subject to exclusion filtering. In Stage 2, a re-validation step removed 26 documents that passed through supporting keywords only but described non-flood events (sinkholes, road construction, festivals), yielding 112 final documents (82 news articles, 30 disaster alerts).

To verify thematic coherence, Latent Dirichlet Allocation (LDA) topic modeling was applied using Gensim (k=15). Four dominant thematic clusters were identified (heavy rainfall alerts, flood-induced traffic congestion, infrastructure damage, and weather forecasting), with zero off-topic documents detected. Structured flood event extraction yielded 67 flood-affected locations, 14 road closure records, and 27 spatial co-occurrence relationships, serving as the empirical foundation for KG construction.

A literature pool of relevant rainfall- and flood-traffic studies [3,4,6–9] was consolidated based on the criterion of providing explicit quantitative values for traffic flow changes as a function of rainfall intensity or flood depth.

Table 1 summarizes the synthesized values across five rainfall intensity levels: light (<2.5 mm/h; speed reduction 2%–5%, capacity reduction 4%–10%), moderate (2.5–7.6 mm/h; 4%–9%, 7%–11%), heavy (7.6–15 mm/h; 6%–17%, 10%–30%), very heavy (15–50 mm/h; 12%–25%+, 15%–32%+), and extreme (>50 mm/h; 21%–40%+, road closure). Ni et al. [4] reported that waterlogging amplifies traffic impacts by 1.3–2.0 times compared to rainfall alone, and Pregnolato et al. [3] developed a depth-disruption function relating flood depth to vehicle speed. These values were directly mapped to the TrafficImpact entities in the KG.

Training data was generated through constrained weighted graph walks following a seven-step path: TimeContext → WeatherCondition → FloodState → Location → Closed Road (via FLOOD_CLOSES) → TrafficImpact → ResponseAction. At each step, the next node is selected via weighted probabilistic sampling. For locations without FLOOD_CLOSES edges, an alternative mechanism selects spatially proximate NEAR neighbors.

Each graph walk produces an ordered 7-tuple (TimeContext → WeatherCondition → FloodState → Location → ClosedRoad → TrafficImpact → ResponseAction). The tuple is serialized into a JSON metadata object M with eight fact slots (Section 3.6) and then rendered into a natural-language prompt via a paragraph-level template tk selected uniformly from T = {t1, …, t256}.

The prompt, together with the tuple, forms a (prompt, reference) pair used for instruction fine-tuning; Fig. 5 illustrates the data flow, and Table 4 provides a concrete example of the resulting model input and output formats.

The LLM never accesses the KG at inference time: all KG information is compiled into the training prompt offline, which is precisely the mechanism that allows the fine-tuned sLLM to run locally without a graph store, a prerequisite for disaster-time deployment under connectivity loss.

Each scenario was rendered as a four-paragraph English narrative (situation overview, flood situation, traffic impact, response measures) using four template variants per paragraph (256 combinations). Quantitative values were randomly sampled within KG-defined ranges to produce continuous variation. A total of 800 scenarios were generated with 338 unique entity combinations, validated through 38 automated quality checks with 100% compliance. The dataset was split into training (644), validation (78), and test (78) sets via stratified sampling by weather condition.

Three sLLMs were selected to enable systematic comparison across architecture and parameter scale: Flan-T5-Large [13] (770M, encoder-decoder), Qwen2.5-3B-Instruct [14] (3B, decoder-only), and Qwen2.5-7B-Instruct [14] (7B, decoder-only). Flan-T5-Large serves as a cross-family reference baseline. The intra-family pairing of Qwen2.5-3B and Qwen2.5-7B provides a scale-controlled comparison under identical architecture; the implications and limits of this comparison, including the ceiling effect observed on the KG Fact Consistency metric, are discussed in Sections 5.1 and 5.4. All models operate within the 16 GB VRAM constraint of a single commodity GPU, addressing practical requirements for local deployment in disaster response systems.

All models were fine-tuned using QLoRA [15], combining 4-bit NF4 quantization with low-rank adaptation (LoRA) [21] (r = 16, α = 32, dropout = 0.05). Common settings included bfloat16 compute, 8-bit AdamW optimizer, cosine learning rate scheduler (warmup ratio 0.05), and effective batch size of 16. Model-specific configurations are detailed in Table 2: all three models were trained for 10 epochs with model-specific learning rates. Flan-T5-Large used a learning rate of 3×10−4 with LoRA on query/value projections (4.7M trainable parameters), while both Qwen models used 2×10−4 with LoRA on query/key/value/output projections (7.4 and 10.1M trainable parameters, respectively). Experiments were conducted on a single NVIDIA RTX 4060 Ti (16 GB VRAM) with AMD Ryzen 7 7800X3D, using PyTorch 2.9.1 (CUDA 12.8) and HuggingFace Transformers 5.2.0.

Four complementary metrics were employed: three standard text generation metrics and one domain-specific factual fidelity metric.

BLEU-4 [22] measures n-gram precision with a brevity penalty. Given a candidate sentence c and reference sentence r the score is defined as:(1)BLEU−N=BP×exp⁡(∑n=1N(1Nlog⁡pn)where pn is the modified n-gram precision (clipped to reference counts), and BP=min(1,e(1−|r||c|)) is the brevity penalty. Here, N=4.

ROUGE-L [23] measures the longest common subsequence (LCS) between candidate and reference:(2)Rlcs=LCS(r,c)|r|,Plcs=LCS(r,c)|c|,Flcs=(1+β2)PlcsRlcs/(Rlcs+β2Plcs)where |r| and |c| denote reference and candidate lengths, respectively, and β is set to favor recall.

BERTScore F1 [24] computes token-level semantic similarity using contextual embeddings. Given reference token embeddings x=<x1,…,xm> and candidate token embeddings y=<y1,…,ym>:(3)RBERT=(1m)∑i=1mmaxjcos⁡(xi,yj),PBERT=(1n)∑j=1nmaxicos⁡(xi,yj) (4)FBERT=2×(PBERT×RBERT)/(PBERT+RBERT)

The embedding model is roberta-large.

KG Fact Consistency Score: Standard metrics assess text quality but not factual fidelity to input conditions. In data-to-text generation, task-specific factual metrics are commonly employed alongside standard metrics: Wen et al. [25] introduced the slot error rate for dialogue natural language generation (NLG), providing precedent for task-specific evaluation metrics in data-to-text generation. Following this established practice, this study introduces a domain-specific metric that evaluates whether key facts from the input KG triples are faithfully preserved in generated text. For a generated scenario y^ and its corresponding input metadata m containing K fact categories, the score is defined as:(5)KG−FC(y^,m)=(1K)∑k=1Kfk(y^,mk)where each fk is a binary verification function that returns 1 if fact category k is correctly preserved in y^, and 0 otherwise. Eight fact categories (K=8) are defined: (1) weather condition, (2) rainfall intensity (mm/h), (3) flood location, (4) flood state, (5) closed road, (6) speed reduction percentage, (7) capacity reduction percentage, and (8) response actions. Categories 1–7 are verified via exact or partial string matching against the input values, while category 8 computes the fraction of response actions mentioned (f8=|Amatched||Atotal|).

The design uses a uniform averaging scheme: the KG Fact Consistency score is the arithmetic mean of the eight per-category verification values (each in [0, 1]), giving equal weight to all facets. Category weights were not learned or tuned, to avoid overfitting the score to the test distribution.

This design is in the same lineage as slot-based factual metrics in data-to-text generation [25,26], which evaluate whether structured input facts are faithfully preserved in surface text. Table 5 specifies the verification rule for each category as implemented in the evaluation code.

Three consequences of this design are explicitly acknowledged and deferred to future work: (a) the metric measures fact omission but not fabrication (partially addressed via the post-hoc analysis in Section 4.3.4); (b) string-matching verification saturates at 1.0 once a model reliably copies all input facts; and (c) the metric does not penalize contextual misplacement of otherwise correct values. A subsequent human-expert evaluation targets precisely these three dimensions.

All models were evaluated on the 78-sample test set stratified by weather condition, with both overall and per-condition breakdowns analyzed.

From 254 raw documents (89 news articles, 165 disaster alerts), two-stage filtering retained 112 documents (Table 6). Keyword filtering reduced the corpus to 138, and re-validation removed 26 false positives (sinkholes, construction, festivals). LDA topic modeling (k=15) confirmed four dominant thematic clusters related to urban flood-traffic scenarios with zero off-topic documents. Structured extraction yielded 67 flood-affected locations and 14 road closure records, with Namgang River as the most frequently mentioned location (62 occurrences). Table 6 summarizes the two-stage filtering pipeline and the stage-by-stage corpus reductions.

The domain KG comprises 58 nodes across six entity categories and 285 edges representing ten relationship types (Table 2, Fig. 3). Edge weights were derived from document frequency (76.5%), literature values (8.4%), and domain rules (15.1%). The FLOOD_CLOSES relationship (14 edges) encodes observed flood-to-road-closure spatial pairings, with three spatial clusters identified: the Namgang River corridor, the Yeongcheongang area, and the Sangpyeong lowland. Rule-based validation confirmed zero constraint violations.

Using the KG, 800 scenarios were generated in approximately 5 min (Table 7). The dataset exhibits realistic distributions: heavy to very heavy rainfall constitutes 62.3% of conditions, and Namgang River accounts for 27.5% of locations, proportional to its corpus frequency.

Table 10 compares model outputs for a representative Heavy Rainfall scenario (8.7 mm/h, Namgang River). All three models correctly preserve key facts: rainfall intensity, flood location, traffic impacts (15.6% speed reduction, 10.2% capacity reduction), and response actions. Qualitative differences emerge in style: Flan-T5-Large produces more verbose narratives with connective phrases, while Qwen models generate more structured, report-like text. Qwen2.5-7B produces the most concise output while preserving all factual elements.

Three qualitative observations complement the quantitative metrics above; they do not constitute a human evaluation and are intended to illustrate cases where automatic metrics align with, or diverge from, intuitive judgment. (a) Convergent case (Extreme Rainfall): Qwen-7B preserves all eight input facts and the narrative coherently escalates to a full closure directive, consistent with its KG Fact Consistency of 1.0000. (b) Metric-insensitive case (Heavy Rainfall): Qwen-3B and Qwen-7B both score 1.0000 on KG Fact Consistency, yet their outputs differ in stylistic register (the 7B variant is noticeably more concise), a distinction that the automatic metrics do not capture. (c) Blind-spot case (Light Rainfall): Flan-T5-Large preserves the numeric values (speed and capacity reductions) correctly, so they pass string-matching verification, yet the connective phrasing places these values in an operationally ambiguous context—a form of contextual misplacement that is invisible to the present metric and motivates the supplementary evaluation avenues discussed in Section 5.4. These qualitative observations are illustrative rather than a formal human evaluation, acknowledging that automatic metrics have known limitations in capturing factual and semantic adequacy [27,28].

The results reveal two factors potentially influencing performance: architecture and parameter scale. The Flan-T5 vs. Qwen contrast varies both factors simultaneously and cannot isolate an architectural effect; Flan-T5-Large is best interpreted as a cross-family reference baseline.

The intra-family Qwen2.5-3B vs. 7B contrast provides a scale-controlled comparison under identical architecture. The scale effect on KG Fact Consistency is obscured by ceiling saturation (both Qwen models reach 1.0000), and the observable scale effect is limited to modest improvements on lexical metrics (BLEU-4: +1.5%; ROUGE-L: +2.1%).

The controlled within-family comparison (Qwen2.5-3B vs. 7B) provides a scale-controlled contrast under identical architecture: scaling from 3 to 7B yields consistent but modest improve-ments on lexical metrics (BLEU-4: +1.5%, ROUGE-L: +2.1%), while both models saturate at KG Fact Consistency (1.0000), a consequence of the string-matching ceiling effect described in Section 5.4. Under this ceiling, the 3B model already matches the 7B model on the factual-preservation axis, so the additional parameters of the 7B model manifest primarily in text-generation quality rather than factual accuracy. This pattern of diminishing returns is consistent with scaling literature suggesting that task-specific improvements plateau when training data is domain-constrained.

The weather-condition breakdown provides additional insight. Flan-T5-Large’s KG Fact Consistency shows an inverse relationship with scenario complexity (0.792 for Light vs. 0.961 for Very Heavy Rainfall), suggesting difficulty with subtler conditions. Both Qwen models maintain consistency of 1.0000 across all weather conditions, demonstrating robust conditional generation. Notably, Qwen2.5-3B slightly outperforms Qwen2.5-7B in ROUGE-L for Light Rainfall (0.725 vs. 0.691), potentially reflecting minor differences in template pattern capture between model scales.

The design role of each constraint rule is analyzed through counterfactual reasoning grounded in the zero-violation outcome of the rule-based validation.

Removing P5 (FLOOD_CLOSES spatial consistency) would admit closures at geographically unrelated bridges (e.g., flooding at Yeongcheongang River paired with closure at Jinyang Bridge), violating the spatial pairings observed in the 112-document corpus. Removing prohibition rules P1–P4 would readmit physically implausible combinations such as Light rainfall producing Severe Inundation, which the corpus never documents. Allowance rules A1–A4 govern upper-bound co-occurrences; their removal would widen the scenario distribution at the cost of operational plausibility, particularly for safety-critical response directives.

A controlled retraining-based ablation quantifying these effects is an important direction for future work.

The results demonstrate that KG-generated synthetic training data provides sufficient signal for effective domain fine-tuning. Even the smallest model achieves BERTScore F1 of 0.9415, indicating strong semantic alignment. Both Qwen models achieve KG Fact Consistency of 1.0000 across all 78 test samples, preserving all eight fact categories (weather condition, rainfall intensity, flood location, flood state, closed road, speed reduction, capacity reduction, response actions)—validating the KG-driven approach as a reliable method for grounding model outputs in structured domain knowledge.

This design reflects a deliberate separation of concerns between the KG and sLLM components. At the 770M–7B parameter scale, sLLMs lack the capacity for reliable causal reasoning over physical constraints (e.g., inferring that light rainfall cannot cause severe inundation). Rather than relying on the model to learn such relationships from limited training data, the pipeline delegates constraint enforcement entirely to the KG (whose rule-based validation guarantees 100% compliance), while the sLLM is tasked solely with data-to-text generation: transforming structured KG-derived conditions into coherent natural language narratives. This division ensures that factual plausibility is deterministically guaranteed by the KG, independent of model scale or training data volume.

The KG-driven approach offers three key advantages: (1) constraint rules (P1–P5, A1–A4) prevent physically implausible scenarios critical for disaster management; (2) the FLOOD_CLOSES relationship ensures spatial consistency between flood locations and road closures, addressing a common limitation of template-based approaches; and (3) literature-derived traffic impact values provide empirically grounded quantitative parameters distinguishing generated scenarios from those of general-purpose LLMs.

The pipeline generates 800 scenarios in approximately 5 min after the one-time KG construction, enabling on-demand dataset expansion as new empirical data become available.

The comparative evaluation provides practical deployment recommendations. Qwen2.5-3B achieves 98.6% of the 7B’s BLEU-4 and identical KG Fact Consistency (1.0000) while requiring only 53% of peak VRAM (7.82 vs. 14.63 GB) and 71% of training time. For municipal systems with 8–12 GB VRAM, the 3B model represents the most cost-effective choice, achieving the same factual consistency as the 7B model at substantially lower computational cost.

Qwen2.5-7B achieved the highest text generation scores but operates at the practical limit of 16 GB GPUs (91.4% VRAM utilization), and is therefore recommended for higher-capacity GPUs or applications requiring maximum text quality. Flan-T5-Large (3.16 GB VRAM) is the most resource-efficient but exhibits substantially lower performance, suitable only for severely constrained hardware or as a baseline.

All models were trained for 10 epochs on a single GPU, with per-model training times ranging from 46 to 146 min, demonstrating practical feasibility for municipal disaster response deployment.

Five limitations should be acknowledged.

First, the study focuses on a single city (Jinju), and generalizability to different flooding mechanisms (coastal, snowmelt) has not been validated.

Second, the Flan-T5 vs. Qwen comparison varies architecture and parameter scale simultaneously and cannot isolate an architectural effect. Flan-T5-Large is interpreted as a cross-family reference baseline rather than as evidence of an encoder-decoder vs. decoder-only performance difference; the controlled comparison is restricted to the intra-family Qwen2.5-3B vs. 7B contrast. A matched decoder-only baseline near 770M parameters would further disentangle architectural and scale effects.

Third, the test set (78 samples) limits statistical power for per-weather-condition comparisons.

Fourth, the KG Fact Consistency metric is subject to a ceiling effect: once a model reliably copies all input facts, the score saturates at 1.0 and becomes insensitive to subsequent factual quality dimensions, in particular fabrication of facts not present in the input and contextual misplacement of otherwise correct values.

The post-hoc Location Fabrication Analysis (Section 4.3.4) showed zero cross-reference fabrication across all 234 test predictions, providing a second independent factual signal on the Location-entity axis; this analysis does not cover fabrication of other fact categories or broader semantic errors [5].

Fifth, evaluation relies on automated metrics without human expert assessment. Automatic text-generation metrics have known limitations in capturing deeper semantic dimensions such as rationality, coherence, and information completeness [27,28]. A blinded evaluation with ≥3 domain experts reporting Fleiss’ κ would complement the present automated and KG-based fidelity measures.

Notwithstanding these limitations, the study establishes that KG-driven training-data construction enables sLLMs fine-tuned on commodity hardware to achieve KG Fact Consistency of 1.0000 with zero Location-level cross-reference fabrication across 234 test predictions. The FLOOD_CLOSES spatial-consistency rule offers a reusable mechanism for injecting safety-critical spatial constraints into generation, and the nine-rule constraint framework provides a template adaptable to other urban-disaster domains.

The limitations enumerated above frame the scope of validity and motivate methodological refinement, rather than nullify the core finding that KG-grounded fine-tuning can constrain sLLMs to produce factually consistent scenarios for disaster preparedness. Scalability of the manually curated KG is addressed separately in Section 5.5.

The present KG was partially hand-curated, which raises a legitimate concern about transferability to other cities or flooding regimes beyond the monsoon-driven case of Jinju (e.g., coastal, snowmelt). Three elements are outlined below of a scalable path forward. (i)

Minimum data requirement. In the present study’s application to Jinju, reliable KG construction required on the order of 100 filtered local documents and 10+ quantitative literature sources; this serves as an operational lower bound, not a proven threshold, for replicating the pipeline in a new city.