Knowledge graphs serve as indispensable computational tools for converting fragmented, unstructured text into organized and machine-readable relational representations across academic, cultural, and enterprise domains. Building these structured networks depends on relation extraction, a foundational natural language processing task that identifies specific entities and determines the precise semantic connections linking them within raw prose. Conventional extraction workflows have historically depended on supervised deep learning models, including specialized transformer architectures and bidirectional encoders, that require extensive, meticulously annotated training corpora. In dynamic environments and specialized disciplines, however, human labeling of long-tail relational distributions is prohibitively expensive, and existing supervised systems struggle to adapt when encountering emergent entity types without full parameter retraining. While contemporary large language models possess broad semantic comprehension and in-context learning capabilities that bypass task-specific weight updates, directly deploying them for knowledge graph construction presents severe operational hurdles. Generative language models frequently suffer from stochastic output instability, factual hallucinations, and pronounced semantic drift, generating lexically distinct yet synonymous labels that fragment relational registries. To resolve this tension, computer scientists have designed a closed-loop framework designated as Dynamic Schema Evolution with Large Language Models for Iterative Relation Extraction, or DSE-RE, which dynamically maintains evolving schema tables to constrain generative extraction across iterative processing cycles.
The methodological framework of DSE-RE eliminates the traditional requirement for static, pre-engineered ontologies by establishing an automated extract-refine-inject architectural loop comprising three coordinated modules: LLM-Driven Inference, an Evolving Schema Registry, and a Graph-Based Semantic Clustering Algorithm. During the initial inference phase, the system segments unstructured domain documents into individual sentences and executes conditional probability generation using structured prompts composed of hierarchical instructions. A system-level constraint establishes formal behavioral boundaries and requires the model to output relational candidate tuples adhering strictly to a machine-readable, recursive JSON format. Concurrently, a user-level context prompt supplies few-shot demonstrative examples, task extraction guidelines, the specific input sentence, and dynamic schema tables retrieved directly from the memory registry. Rather than classifying text against a fixed, immutable label space, the generative inference engine identifies relational triples alongside candidate entity and relation types. These newly extracted items are routed into the schema registry, which maintains separate evolving tables for entity categories and relational ties, functioning as an active semantic anchor that prevents subsequent generation rounds from drifting arbitrarily into uncoordinated terminological variations.
Because probabilistic generative models naturally produce disparate surface expressions for identical concepts, simply accumulating raw textual strings in the registry would rapidly saturate contextual memory and degrade prompt guidance through redundant synonyms. To maintain schema quality and control registry growth, the architecture triggers a specialized Graph-Based Semantic Clustering Algorithm whenever schema table sizes exceed predetermined numerical thresholds. Unlike geometric clustering methods that rely on Euclidean distances or assume spherical data distributions, this algorithm evaluates strict semantic equivalence in high-dimensional vector spaces. A pre-trained language model maps each candidate type into dense embedding vectors, whereupon pairwise cosine similarities are calculated across the collected categories. The algorithm then constructs an undirected semantic graph wherein vertices represent individual type labels and edges connect pairs whose semantic similarity surpasses an empirically calibrated threshold. Applying depth-first search traversal across the resulting graph topology efficiently partitions the network into distinct connected components, each isolating a cohesive cluster of synonymous expressions. The system then selects a canonical representative term for each component, merging redundant variants into consolidated types before feeding the streamlined schema back into subsequent prompt contexts.
Algorithmic complexity analysis underscores the computational tractability of this graph-based clustering procedure within automated processing pipelines. For a set of unique types numbering n mapped into semantic vectors of dimensionality d, calculating pairwise cosine similarities across all candidate pairings requires evaluating n(n minus 1) divided by 2 combinations, yielding a graph construction time complexity of O(n squared multiplied by d). The subsequent depth-first search traversal identifying connected components exhibits a worst-case computational complexity of O(n plus the cardinality of edge set E), which simplifies to O(n squared). Consequently, the overall runtime complexity of the clustering algorithm remains bounded at approximately O(n squared multiplied by d). The operational memory requirement similarly scales as O(n multiplied by d) for storing high-dimensional semantic vectors alongside an O(n squared) allocation for the semantic graph adjacency list. Because the number of distinct schema types in targeted domains remains manageable and the vector dimension represents a fixed constant associated with the chosen embedding model, these computational and memory footprints introduce negligible computational overhead on standard modern computing infrastructure.
The FOBIE corpus comprises computational biology literature encompassing approximately 1,500 sentences and three predefined relation types, structured into 1,248 training samples and 300 test sentences. In contrast, the newly constructed HSA benchmark reflects complex cultural heritage documentation derived from multi-source traditional Chinese architecture records, containing 1,716 verified entity-relation triples spanning 19 entity classifications and 41 distinct relation types. Performance assessment employed standard precision, recall, and F1 metrics alongside a specialized Subjective Question Correction Score, denoted as the SQC-Score. This complementary metric addresses the known limitations of strict string-matching evaluation by employing a two-stage alignment mechanism that utilizes model synonym matching and natural language inference entailment scoring, incorporating an adaptive threshold to reward semantically valid predictions that express ground-truth relations through valid alternative phrasing.
Experimental comparisons against mainstream supervised deep learning relation extraction baselines—specifically PRGC, OneRel, and GPLinker—demonstrated severe recall bottlenecks inherent to traditional supervised models when operating in low-resource or long-tail settings. On the architectural HSA dataset, supervised models achieved recall values of only 0.493 for PRGC, 0.467 for OneRel, and 0.503 for GPLinker, with overall F1 scores remaining constrained between 0.56 and 0.70. Supervised recall improved somewhat on the biomedical FOBIE dataset, registering 0.633, 0.653, and 0.697 across the respective architectures, though the models remained limited by insufficient training coverage of sparse relation instances. Standalone deployments of foundation models, specifically GPT-4o and DeepSeek, significantly enhanced extraction recall, reaching 0.706 and 0.740 on HSA and raising baseline F1 metrics to 0.740 and 0.763, respectively. However, without schema evolution feedback, baseline foundation models still failed to extract heterogeneous expressions consistently, exhibiting substantial false-negative rates and schema fragmentation that limited automated knowledge graph construction quality.
Integrating the closed-loop schema evolution architecture produced marked, statistically significant performance improvements across both benchmark domains. When coupled with GPT-4o on the HSA dataset, the framework drove extraction recall up from 0.706 to 0.871, yielding an improved F1 score of 0.781 despite a slight drop in strict precision from 0.815 to 0.743. On the FOBIE corpus, the framework increased recall from 0.743 to 0.800, raising the overall F1 measure from 0.741 to 0.771. Pairing the framework with DeepSeek delivered even stronger gains on the architectural dataset, elevating recall from 0.740 to 0.864 and achieving a peak F1 score of 0.808, compared to 0.763 for the unaugmented baseline. Multiple independent experimental runs demonstrated low variance, confirming the statistical stability of the schema feedback mechanism. The researchers noted that while precision exhibited modest reductions due to broader surface matching, this trade-off was overwhelmingly compensated by dramatic recall expansions, establishing the framework as an effective tool for recovering sparse relational facts that standalone language models systematically miss.
Ablation investigations systematically substantiated the essential functional interdependence between the Evolving Schema Registry and the Graph-Based Semantic Clustering Algorithm. On the HSA benchmark, completely removing both modules reduced model performance to the baseline native extraction score of 0.740. When the custom graph clustering module was replaced with standard density-based clustering techniques, model performance deteriorated markedly: substituting DBSCAN dropped the overall F1 score to 0.745, whereas implementing HDBSCAN caused severe system degradation, plunging the F1 score to 0.430 due to catastrophic drops in both precision and recall. Tracking registry expansion across extraction iterations revealed that omitting semantic clustering caused both entity and relation type counts to expand in an uncontrolled, near-linear trajectory. In contrast, incorporating the graph-based clustering module constrained registry growth, merging synonymous variations and preserving contextual space. Comparative evaluations on the 20 Newsgroups benchmark further verified clustering efficacy: the graph-based algorithm achieved an Adjusted Rand Index of 0.392, markedly outperforming KMeans at 0.179 and Birch at 0.136, alongside superior accuracy, Normalized Mutual Information, and V-Measure scores of 0.465, 0.505, and 0.505, respectively.
To validate real-world practical utility, the investigators deployed the framework to construct an end-to-end domain knowledge graph from authoritative historical texts on Huizhou-style architecture. A source corpus comprising ten authoritative domain books was digitized via systematic optical character recognition, curated by eliminating indices and footnotes, and segmented into 4,330 granular sentences containing 1,108,440 Chinese characters. Operating with DeepSeek as the generative backbone, the framework automatically extracted domain-specific architectural concepts, historical periods, and craft techniques while dynamically organizing the schema without pre-engineered ontological rules. The resulting structured knowledge was imported into a Neo4j graph database, establishing indexed labels and constraints capable of supporting real-time complex queries. The completed architectural graph encompasses 64,361 individual entity nodes and 62,403 verified entity-relation triples spanning 27 entity categories and 37 relational classifications, modeling intricate architectural relationships such as the fire-prevention utility of traditional horse-head walls and supporting semantic retrieval, automated question answering, and digital heritage conservation planning.
Runtime scalability analyses demonstrated that processing latency remains largely governed by language model API response times, averaging under four seconds per sentence and exhibiting a consistent linear relationship relative to corpus volume. Despite these operational advantages, the researchers highlighted several technical boundaries and limitations governing the current implementation. The finite context window capacity of underlying foundation models imposes an upper ceiling on the volume of historical schema categories that can be dynamically inserted into user prompts during prolonged extraction runs. Furthermore, empirical validation has thus far focused on two moderately sized specialized corpora, leaving cross-domain transferability to massive or linguistically heterogeneous multilingual repositories unverified. Most crucially, because extraction veracity remains fundamentally bounded by the factual accuracy and internal domain coverage of the base language model, generative hallucinations cannot be entirely eliminated. To address these operational constraints, the authors outlined future research priorities focused on engineering hierarchical context management systems, incorporating active learning loops with minimal human supervision to arbitrate ambiguous classifications, and integrating multimodal computer vision features to facilitate automated multimodal knowledge graph construction.
Subject of Research: Iterative relation extraction and dynamic schema evolution using large language models
Article Title: DSE-RE: Dynamic Schema Evolution with LLM for Iterative Relation Extraction
Article References: Wang, T., Shi, D., Aguilar, J., & Zurada, J. (2026). DSE-RE: Dynamic Schema Evolution with LLM for Iterative Relation Extraction. Human-Centric Intelligent Systems, 6(3), 429-449. https://doi.org/10.1007/s44230-026-00158-1
Image Credits: AI Generated
DOI: 10.1007/s44230-026-00158-1
Keywords: Relation extraction, Large language models, Knowledge graphs, Dynamic schema evolution, Semantic clustering, In-context learning, Graph-based clustering, Depth-first search, Schema drift, Information extraction, FOBIE benchmark, Huizhou architecture, Neo4j, Cosine similarity, Natural language processing, Entity type tables, Relation type tables
Cite Scienmag News
Gavin Prescott. (November 18, 2026). Dynamic Schema Evolution Enables Language Models to Build Reliable Knowledge Graphs. Scienmag. https://scienmag.com/dynamic-schema-evolution-enables-language-models-to-build-reliable-knowledge-graphs/
Gavin Prescott. "Dynamic Schema Evolution Enables Language Models to Build Reliable Knowledge Graphs." Scienmag, 18 November 2026, https://scienmag.com/dynamic-schema-evolution-enables-language-models-to-build-reliable-knowledge-graphs/. Accessed 18 November 2026.
Gavin Prescott. "Dynamic Schema Evolution Enables Language Models to Build Reliable Knowledge Graphs." Scienmag. November 18, 2026. https://scienmag.com/dynamic-schema-evolution-enables-language-models-to-build-reliable-knowledge-graphs/








