Graph Fusion Across Languages using Large Language Models

Kaung Myat Kyaw; Khush Agarwal; Jonathan Chan

cs.CLMar 22, 2026

Graph Fusion Across Languages using Large Language Models

Authors: Kaung Myat Kyaw, Khush Agarwal, Jonathan Chan

Organizations: Innovative Cognitive Computing Research Center (IC2) School of Information Technology, KMUTT

Abstract

Combining multiple knowledge graphs (KGs) across linguistic boundaries is a persistent challenge due to semantic heterogeneity and the complexity of graph environments. We propose a framework for cross-lingual graph fusion, leveraging the in-context reasoning and multilingual semantic priors of Large Language Models (LLMs). The framework implements structural linearization by mapping triplets directly into natural language sequences (e.g., [head] [relation] [tail]), enabling the LLM to map relations and reconcile entities between an evolving fused graph ( $G_{c}^{(t-1)}$ ) and a new candidate graph ( $G_{t}$ ). Evaluated on the DBP15K dataset, this exploratory study demonstrates that LLMs can serve as a universal semantic bridge to resolve cross-lingual discrepancies. Results show the successful sequential agglomeration of multiple heterogeneous graphs, offering a scalable, modular solution for continuous knowledge synthesis in multi-source, multilingual environments.

Figures & tables

Figure 1: Conceptual framework for LLM-based cross-lingual knowledge graph fusion. Heterogeneous input graphs in English ( $G_{en}$ ), Chinese ( $G_{zh}$ ), and French ( $G_{fr}$ ) are processed through a fusion framework that linearizes structural triplets into natural language and the LLM acts as a semantic bridge to reconcile entities.

Figure 2: Overview of the proposed modular framework for $N$ -Graph Knowledge Graph Fusion.

Panel of Figure 3 — Figure 3: Details of the LLM prompting interface. The system prompt (left) defines the operational constraints and confidence thresholds, while the user prompt (right) serves as the data payload, containing the linearized triples from the source and target knowledge graphs.

Metric	Value
Total Batch Pairs	768
Unique Alignment Predictions	5,416
True Positives (TP)	3,543
False Positives (FP)	1,873
Precision	65.4%
Recall	23.6%

Table 1: Primary Alignment Metrics (DBP15K

zh\_en

)

k	Hits / 15,000	Rate	Acc. Predicted
1	3,516	23.4%	88.3%
5	3,543	23.6%	89.0%
10	3,543	23.6%	89.0%

Table 2: Hits@k Evaluation

$\tau$	Pred.	TP	Prec.	Rec.	F1
0.00	5,416	3,543	65.4%	23.6%	34.7%
0.80	4,309	3,540	82.2%	23.6%	36.6%
0.90	4,002	3,520	88.0%	23.5%	37.0%
0.95	3,483	3,219	92.4%	21.4%	34.8%

Table 3: Confidence Threshold (

\tau

) Sensitivity Sweep

Knowledge Graphs

Knowledge graphs represent entities and relationships as structured, queryable networks for integrating data and supporting reasoning.

See all 232 papers

The Graph Language: How Knowledge Graphs Speak to Large Language Models
Aug 2, 2026Giuseppe PirròLLM Grounding Knowledge Augmentation for Language Models
DeMaCS, University of Calabria
VL-KGE: Vision-Language Models Meet Knowledge Graph Embeddings
Mar 2, 2026Athanasios Efthymiou, Stevan Rudinac, Monika Kackovic +2KG Embedding Multimodal KGs
University of Amsterdam, Amsterdam, The Netherlands
Tokenization, Fusion and Decoupling: Bridging the Granularity Mismatch Between Large Language Models and Knowledge Graphs
Feb 26, 2026Siyue Su, Jian Yang, Bo Li +1KG Completion LLMs for Graph Learning
Beihang University

Aug 2, 2026J/K move · Enter open · S save

The Graph Language: How Knowledge Graphs Speak to Large Language Models

Giuseppe Pirrò

DeMaCS, University of Calabria

LLM Grounding Knowledge Augmentation for Language Models LLM Reasoning with Graphs Knowledge Graph Question Answering

Open paper AskPDF arXiv

Large Language Models (LLMs) excel at reasoning but benefit from grounding provided by Knowledge Graphs (KGs). However, integrating these paradigms is challenging. We introduce GRALAN, which enables KGs to speak directly in the LLM's semantic space through relational tokens that preserve graph structure. GRALAN-s trainable language mediator generates structured tokens for any frozen LLM, creating a foundation for knowledge-intensive applications. We demonstrate its effectiveness in question-answering by re-framing the task as entity classification over question-focused subgraphs. Experiments show that GRALAN significantly outperforms existing methods, particularly on complex multi-hop reasoning tasks, establishing a new paradigm for KG-LLM integration that maintains structural fidelity while leveraging LLMs' reasoning capabilities.

Abstract

Knowledge Graphs

Explore similar work

The Graph Language: How Knowledge Graphs Speak to Large Language Models

VL-KGE: Vision-Language Models Meet Knowledge Graph Embeddings

Tokenization, Fusion and Decoupling: Bridging the Granularity Mismatch Between Large Language Models and Knowledge Graphs