Construct, Align, and Reason: Large Ontology Models for Enterprise Knowledge Management

Enterprise-scale knowledge management faces significant challenges in integrating multi-source heterogeneous data and enabling effective semantic reasoning. Traditional knowledge graphs often struggle with implicit relationship discovery and lack sufficient semantic understanding for complex question answering. To address these limitations, we introduce a unified construct–align–reason framework, the large ontology model (LOM). We first build a dual-layer enterprise ontology from structured databases and unstructured text, subsequently fusing these sources into a comprehensive enterprise ontology. To enable instruction-aligned reasoning, we propose a unified three-stage training pipeline: ontology instruction fine-tuning to improve structural understanding; text-ontology grounding to strengthen node semantic encoding; and multi-task instruction tuning on ontology-language pairs with curriculum learning to enhance semantic reasoning and generation. We also construct comprehensive training and evaluation datasets covering diverse ontology reasoning tasks. On this benchmark, our 4B-parameter LOM achieves 89.47% accuracy and outperforms DeepSeek-V3.2 on complex graph reasoning, indicating effective fusion of ontology structure and language.

💡 Research Summary

The paper addresses two fundamental challenges in enterprise‑scale knowledge management: integrating heterogeneous data sources (structured relational databases and unstructured textual documents) and performing deep semantic reasoning over the resulting knowledge base. To solve these problems, the authors propose a unified “Construct‑Align‑Reason” framework called the Large Ontology Model (LOM).

In the Construct phase, a dual‑layer ontology is built. From structured databases, a two‑stage scanning process extracts explicit keys and discovers implicit relationships using a multi‑factor algorithm that analyzes both schema metadata and data overlap. This yields a schema layer (abstract) and an instance layer (concrete). From unstructured text, an LLM‑LOM pipeline performs entity‑relation extraction, link prediction, and robust entity disambiguation by combining symbolic rules with semantic embeddings. The two ontologies are then fused via cross‑modal alignment based on tag‑description matching, producing a unified enterprise ontology.

The Align phase introduces a heterogeneous graph encoder (graph transformer) and a text encoder, linked by a linear alignment projector. The model is trained on intra‑type alignment (tokens of the same meta‑type) and inter‑type alignment (tokens of different meta‑types) using next‑token prediction with cross‑entropy loss. This step forces the language model to embed graph‑structured information in the same space as natural language, bridging the semantic gap between graph topology and textual meaning.

In the Reason phase, a three‑stage instruction‑tuning pipeline is applied. First, ontology‑instruction fine‑tuning endows the base LLM with foundational understanding of ontology structures. Second, the text‑ontology grounding stage refines the alignment projector and graph‑token embeddings. Third, multi‑task instruction tuning is performed on a large set of ontology‑language instruction pairs, covering predictive, reasoning, and generative tasks. Curriculum learning orders the training samples from simple prediction to complex generative reasoning, allowing the model to progressively acquire multi‑hop reasoning capabilities.

To support this training regime, the authors construct a comprehensive Chain‑of‑Thought (CoT)‑enhanced dataset containing 115 k samples across 19 graph reasoning tasks, ranging from basic traversals (BFS, DFS) to advanced algorithms (shortest path, minimum spanning tree, PageRank, topological sorting). Each sample includes a problem description, step‑by‑step reasoning trace, and final answer, enabling the model to learn the underlying logic rather than memorizing outputs. The dataset is provided in PyG format with node feature vectors (768‑dim sentence‑transformer embeddings) and graph structure, supporting both homogeneous and heterogeneous graphs.

Experimental evaluation shows that the 4‑billion‑parameter LOM achieves 89.47 % accuracy on the benchmark, outperforming the strong baseline DeepSeek‑V3.2, especially on complex graph reasoning tasks. The results demonstrate that fusing ontology structure with language modeling yields superior reasoning performance compared to prior approaches that treat knowledge graphs as external retrieval tools or rely solely on graph neural networks.

Overall, the paper delivers a coherent end‑to‑end system that can be directly deployed in enterprise environments for real‑time query answering, decision support, and knowledge discovery, effectively bridging data integration and semantic reasoning under a single, scalable architecture.