Skip to content

Medical Bilingualism: Cross-Ontology Graph Retrieval for Incommensurable Knowledge Bases

Conference: EMNLP 2026
Paper: ACL Anthology
Area: Medical LLM
Keywords: Cross-Ontology Graph Retrieval, Incommensurable Knowledge Bases, Dual-Ontology Graph, Relational Graph Attention Network, Medical RAG

TL;DR

Addressing semantic collapse caused by forcing incommensurable medical paradigms (Traditional Chinese Medicine and Western Medicine) into a shared embedding space, XONTO-RAG constructs a dual-ontology graph with verified typed bridging relations and applies a query-aware RGAT retriever to assemble provenance-grounded multi-hop evidence subgraphs.

Background & Motivation

Large Language Models (LLMs) demonstrate remarkable promise across biomedical question answering and clinical decision support, yet they remain vulnerable to critical knowledge gaps and factual hallucinations in specialized and interdisciplinary medical domains. Supervised fine-tuning is computationally expensive and introduces catastrophic forgetting risks, while factual knowledge editing struggles to reconcile systemic philosophical and structural divergences between contradictory knowledge paradigms at scale. Consequently, external evidence-grounding paradigms like Retrieval-Augmented Generation (RAG) and GraphRAG have become standard approaches. However, prevailing retrieval pipelines operate under an unstated yet pervasive assumption: that heterogeneous domain concepts can be projected, ranked, and aggregated within a single shared semantic space or unified scoring function without friction.

This forced unification fundamentally breaks down when confronting what philosophy of science and linguistic anthropology define as ontological incommensurability. Taking Traditional Chinese Medicine (TCM) and Western Medicine (WM) as a quintessential example, TCM operates through functional, systemic models of pattern differentiation and qi dynamic balances, whereas WM relies upon anatomical, material, and molecular pathophysiological mechanisms. Forcing these disparate explanatory primitives into a single continuous vector space induces severe semantic collapse, which surfaces in three recurring failure modes: erroneous alignment (spuriously equating functionally distinct concepts, e.g., mapping "Qi deficiency" directly to anatomical "hypoxia"), conflated reasoning (interweaving incompatible evidence without provenance tracking), and fragmented retrieval (breaking multi-hop reasoning chains across the semantic divide and inducing retrieval derailment).

Prior ontology-grounded RAG approaches either project evidence onto a single authoritative ontology or apply unconstrained embedding-space alignments, failing to model translation uncertainty or to distinguish non-equivalence correspondences from genuine identity. The angle of attack in this work is to preserve explicit ontology labels while inducing structured, auditable cross-paradigm transitions. Core idea: construct an ontology-preserving dual-ontology graph with typed bridging relations induced via an LLM-assisted verify-and-vote pipeline, optimized through a query-aware Relational Graph Attention Network (RGAT) that dynamically injects query nodes to assemble compact, provenance-tracked multi-hop evidence subgraphs.

Method

Overall Architecture

The XONTO-RAG pipeline comprises four tightly coupled stages: dual-ontology graph integration with typed bridge induction, structure-aware representation learning, query-aware dynamic graph retrieval, and interpretable grounded generation. The system takes a natural language medical query as input and outputs a coherent answer backed by step-wise provenance across the dual paradigms.

Initially, the framework maintains the independent node sets and intra-domain relations of both \(\mathcal{D}_{\text{TCM}}\) and \(\mathcal{D}_{\text{WM}}\), ensuring source ontology labels are strictly preserved. Candidate cross-ontology concept pairs are retrieved from curated intermediate lexicons and filtered via dense semantic similarity. Subsequently, an LLM-assisted verify-and-vote mechanism infers fine-grained logical edge types between paired entities, establishing explicit bridging edges \(E_{\text{bridge}}\). During representation learning, a Relational Graph Attention Network (RGAT) learns structure-aware entity embeddings under a joint objective incorporating link prediction, an auxiliary ontology preservation loss to counteract representation collapse, and a pseudo-query contrastive objective. At inference time, the static graph is dynamically contextualized by injecting the user query as a virtual node, driving a query-conditioned dynamic evidence expansion that prioritizes paths containing verified cross-paradigm bridges. Finally, the induced evidence subgraph is linearized into an ontology-tagged narrative and provided to the LLM generator for provenance-grounded synthesis.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["User Medical Query Input"] --> B["Stage 1: Semantic Filtering & Verify-and-Vote Bridge Induction<br/>Construct Dual-Ontology Graph E_bridge"]
    B --> C["Stage 2: Structure-Aware Representation Learning & Anti-Collapse<br/>RGAT + L_onto + L_query"]
    C --> D["Stage 3: Query Injection & Dynamic Multi-Hop Evidence Expansion<br/>Extract Connected Subgraph G_sub"]
    D --> E["Stage 4: Provenance-Aware Linearization & Grounded Generation<br/>Synthesize Dual-Ontology Grounded Answer"]

Key Designs

1. Semantic Filtering & Verify-and-Vote Bridge Induction: Preventing Spurious Equivalences Across Heterogeneous Primitives To overcome the severe distortion caused by naive entity alignment that misinterprets surface-level lexical similarity as strict semantic equivalence, this design decomposes cross-ontology edge creation into semantic proposal and LLM arbitration. First, candidate entity pairs \(\mathcal{C}_0\) are extracted from intermediate medical lexicons linking TCM sources (SymMap, HERB, TCMSP) and WM sources (UMLS, MeSH, DOID). To suppress lexical noise, dense semantic embeddings \(e_u\) and \(e_v\) derived from textual concept definitions are evaluated, retaining pairs exceeding a cosine similarity threshold \(\tau\): $\(\mathcal{C} = \left\{(u, v) \in \mathcal{C}_0 \mid \frac{e_u^\top e_v}{\|e_u\|_2 \|e_v\|_2} \ge \tau \right\}\)$ For each filtered candidate \((u, v) \in \mathcal{C}\), a Large Language Model inspects full medical definitions across multiple decoding seeds or prompt templates to assign an explicit relation \(r\) from a predefined bridging vocabulary \(\mathcal{R}_{\text{bridge}}\) (e.g., functional overlap, approximate equivalence, causal influence, conflict). A bridging edge \((u, r, v) \in E_{\text{bridge}}\) is instantiated if and only if relation \(r\) secures strict majority consensus across \(M\) votes: $\(E_{\text{bridge}} = \left\{(u, r, v) \mid (u, v) \in \mathcal{C},\, \sum_{m=1}^M \mathbb{I}[r^{(m)}(u, v) = r] \ge \delta \right\} \quad (\delta > M/2)\)$ This neuro-symbolic mechanism prunes millions of potential false alignments, ensuring that every cross-paradigm edge conveys an auditable, non-trivial semantic relationship.

2. Structure-Aware Representation Learning & Anti-Collapse Training: Preserving Domain Distinction During Message Passing Standard GNNs treat all neighbors homogenously, causing cross-domain message passing to homogenize distinct explanatory standards and induce latent semantic collapse. To address this, XONTO-RAG adopts an RGAT encoder that assigns relation-specific projection weights \(W_r^{(\ell)}\) and attention vectors \(a_r\) to each edge type \(r \in \mathcal{R}\). At layer \(\ell\), node \(v\) computes normalized attention weights \(\alpha_{vu}^{(r,\ell)}\) over its relation-specific neighbors \(u \in \mathcal{N}_r(v)\): $\(\alpha_{vu}^{(r,\ell)} = \frac{\exp\left(a_r^\top [W_r^{(\ell)} h_v^{(\ell)} \,;\, W_r^{(\ell)} h_u^{(\ell)}]\right)}{\sum_{u' \in \mathcal{N}_r(v)} \exp\left(a_r^\top [W_r^{(\ell)} h_v^{(\ell)} \,;\, W_r^{(\ell)} h_{u'}^{(\ell)}]\right)}\)$ To explicitly penalize representations that erase domain boundaries, the training objective complements standard DistMult link prediction \(\mathcal{L}_{\text{link}}\) with an auxiliary ontology preservation loss \(\mathcal{L}_{\text{onto}}\), which forces a classification head to recover the ground-truth ontology label \(\phi(v)\) directly from final node embeddings \(h_v^{(L)}\): $\(\mathcal{L}_{\text{onto}} = - \sum_{v \in \mathcal{V}} \log P(\phi(v) \mid h_v^{(L)})\)$ Furthermore, to align pretraining representations with inference-time query injection, a pseudo-query contrastive objective \(\mathcal{L}_{\text{query}}\) is optimized over synthetically sampled entity-query pairs. The composite training objective \(\mathcal{L} = \mathcal{L}_{\text{link}} + \lambda \mathcal{L}_{\text{onto}} + \mu \mathcal{L}_{\text{query}}\) enforces structure-guided message passing while preventing cross-ontology feature collapse.

3. Query Injection & Dynamic Multi-Hop Evidence Expansion: Assembling Coherent, Cross-Ontology Subgraphs Static subgraphs fail to adapt to query-specific diagnostic intents and are easily derailed by irrelevant dense clusters. At inference time, the user query \(q\) is temporarily injected into the dual-ontology graph as node \(v_q\), connected to top-\(k\) initial semantic seeds \(\mathcal{S}_{\text{seed}}\) via typed query edges. A single forward pass through the pretrained RGAT produces query-conditioned representations \(h_v^{(L)}(q)\), yielding dynamic relevance scores: $\(s(v \mid q) = \cos\left(h_v^{(L)}(q), h_{v_q}^{(L)}(q)\right)\)$ Starting from \(\mathcal{S}_{\text{seed}}\), multi-hop expansion ranks candidate neighbors using \(s(v \mid q)\) while explicitly prioritizing frontier nodes connected via bridging edges \(E_{\text{bridge}}\). Traversal continues until the induced subgraph contains at least one continuous, bridge-traversing path between paradigms (or reaches a budget limit of \(H=3\) hops and \(k=20\) context entities). Linearizing the subgraph via BFS with \([TCM]\) and \([WM]\) provenance tags delivers compact, highly informative context that prevents evidence pollution and single-feature diagnostic anchoring.

Key Experimental Results

Main Results

Main experiments evaluate across 8 benchmark subsets from MLEC-QA, CMB, and MedQA, spanning Chinese and English exams across Clinic, TCM, WM, and Integrated TCMโ€“WM (TCWM) tasks. All retrieval baselines utilize Qwen2.5-7B as the fixed generation backbone. Evaluated metrics are Answer Accuracy (AC, %) and Faithfulness (FF, %) verified via an NLI protocol using Erlangshen-Roberta-330M-NLI.

Dataset Task Subset Metric XONTO-RAG (Ours) Second-Best Baseline Relative Gain over 2nd Best
MLEC-QA Clinic AC / FF 69.4 / 83.2 63.6 (OG-RAG) / 77.4 (MedGraphRAG) +9.1% / +7.5%
MLEC-QA TCM AC / FF 72.2 / 85.1 64.6 (GraphRAG) / 79.1 (MedGraphRAG) +11.8% / +7.5%
MLEC-QA TCWM AC / FF 77.0 / 78.4 63.0 (OG-RAG) / 75.9 (RAG) +22.2% / +3.3%
CMB Clinic AC / FF 69.4 / 82.5 62.8 (GraphRAG) / 76.1 (MedGraphRAG) +10.5% / +8.4%
CMB TCM AC / FF 73.4 / 81.8 62.2 (GraphRAG) / 79.8 (RAG) +18.0% / +2.5%
CMB WM AC / FF 69.6 / 81.5 63.4 (MedGraphRAG) / 76.7 (RAG) +9.8% / +6.3%
CMB TCWM AC / FF 76.0 / 84.8 71.2 (MedGraphRAG) / 80.4 (GraphRAG) +6.7% / +5.5%
MedQA WM AC / FF 62.4 / 80.2 51.6 (MedGraphRAG) / 77.9 (MedGraphRAG) +20.9% / +2.9%

Ablation Study

The ablation study assesses the individual contributions of typed bridging edges, query-aware dynamic graph walking, and verify-and-vote induction across representative benchmarks (reported in AC, %).

Configuration Variant MLEC-QA TCWM CMB Clinic CMB TCM MedQA WM Mechanism Description
XONTO-RAG (Full Model) 77.0 69.4 73.4 62.4 Full pipeline with typed bridges & query-aware RGAT
w/o Typed Bridges (R-GAT) 68.6 65.4 66.2 55.6 Removes cross-ontology edges; intra-ontology reasoning only
w/o Query-Aware Walker (Dual-RAG) 63.4 64.2 64.2 54.4 Replaces graph walker with dual-corpus dense retrieval
Replaced with Embedding Alignment (GCN-Align) 63.0 61.2 60.4 52.6 Budget-matched embedding alignment without typed relations

Key Findings

  • Peak gains occur on integrated cross-paradigm reasoning: The largest improvement is observed on MLEC-QA TCWM, where XONTO-RAG improves accuracy by +22.2% over OG-RAG (77.0% vs. 63.0%), confirming that explicit typed transitions are vital when reasoning must span incommensurable boundaries.
  • The embedding alignment failure hypothesis is verified: Replacing verify-and-vote typed bridges with continuous embedding alignment (GCN-Align) triggers a dramatic 14.0% drop on MLEC-TCWM (77.0% โ†’ 63.0%), proving that arbitrary connectivity without explicit relational semantics exacerbates semantic collapse rather than mitigating it.
  • Compact retrieval accelerates downstream generation: Although XONTO-RAG spends slightly more time on graph traversal compared to GraphRAG, its highly compact, noise-free evidence subgraphs reduce generation latency, keeping end-to-end latency tightly bounded (2614.3 ms ~ 2664.9 ms vs. 2129.2 ms for GraphRAG).

Highlights & Insights

  • Decoupling rather than violent unification: Instead of attempting a reductionist mapping of traditional medicine concepts onto molecular pathways, the paper embraces ontological incommensurability from the philosophy of science, providing a principled paradigm for interdisciplinary AI.
  • Two-tiered candidate verification: Filtering candidate pairs with dense semantic similarity before subjecting them to multi-round LLM verify-and-vote consensus balances the computational tractability of billion-scale graph induction with high edge precision (achieving 94.25% precision in manual audits).
  • Cross-disciplinary transferability: The dual-ontology framework extends readily beyond medicine. When evaluated on a non-medical MMLU benchmark spanning legal/governance and religion/ethics ontologies, XONTO-RAG achieved 75.0% accuracy (outperforming the strongest baseline by 4.0 points) without architectural modification.

Limitations & Future Work

  • Lexicon dependency for seed proposals: Candidate generation relies upon curated intermediate lexicons; for niche domains devoid of cross-reference indices, initial seed recall may be constrained.
  • Preprocessing overhead for knowledge updates: The offline verify-and-vote pipeline consumed approximately \(1.08 \times 10^8\) tokens during graph construction, presenting non-trivial computational costs for continuously streaming knowledge updates.
  • Need for clinical expert validation: While audited manually across 400 edges (94.25% precision) and verified via GPT-5 consistency checks (92.4% agreement), formal large-scale validation by practicing clinicians remains desirable for high-stakes deployment.
  • vs OG-RAG (Sharma et al., 2025): OG-RAG projects evidence onto ontology-guided hypergraphs but assumes a single authoritative ontology as ground truth; XONTO-RAG maintains a dual-ontology graph with typed bridges and provenance tracking, preventing cross-ontology reasoning derailment.
  • vs MedGraphRAG (Wu et al., 2025): MedGraphRAG focuses on constructing audit triplets within a unified medical KG; XONTO-RAG explicitly addresses inter-ontology incommensurability through anti-collapse losses (\(\mathcal{L}_{\text{onto}}\)) and dynamic query injection.
  • vs Embedding-based KG Alignment (GCN-Align / MTransE): Conventional aligners assume semantic equivalence across graphs; XONTO-RAG demonstrates that forced metric-space alignment induces semantic collapse in incommensurable domains, establishing neuro-symbolic typed bridges as a superior alternative.

Rating

  • Novelty: โญโญโญโญโญ Pioneering application of ontological incommensurability to resolve semantic collapse in cross-paradigm medical GraphRAG.
  • Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation across 8 benchmark subsets, rigorous NLI faithfulness scoring, and extensive ablations including non-medical transfer.
  • Writing Quality: โญโญโญโญโญ Exceptional clarity in problem formulation, tight integration between theoretical philosophy and algorithmic design, and exhaustive experimental reporting.
  • Value: โญโญโญโญโญ Provides an impactful, principled framework for cross-paradigm clinical decision support, multi-ontology RAG, and heterogeneous knowledge integration.