Multilingual Knowledge Graph Construction

• Multilingual Knowledge Graph Construction (mKGC) is the process of building and aligning knowledge graphs using multilingual texts, semi-structured data, and graph fusion techniques.
• It employs diverse methodologies such as dual-encoder retrieval, PLM-based generation, federated learning, and LLM-mediated fusion to overcome challenges in low-resource settings and schema drift.
• The approach improves precision, recall, and structural consistency in knowledge graphs while addressing issues like uneven language coverage and noisy seed alignments.

Multilingual Knowledge Graph Construction (mKGC) denotes the construction, enrichment, alignment, completion, and grounding of knowledge graphs across multiple languages. In this setting, the central difficulty is not merely that knowledge graphs are incomplete, but that multilingual evidence is distributed unevenly across language-specific graphs and corpora, while the textual realizations of otherwise language-agnostic facts are highly language-dependent. Recent work frames mKGC as a family of closely related problems: linking multilingual text to existing facts, predicting missing entities or triples in multilingual KGs, aligning and fusing partially overlapping graphs, and building provenance-preserving multilingual resources for downstream reasoning and verification (Kolluru et al., 2021, Song et al., 2024, Shen et al., 3 Feb 2026).

1. Conceptual scope and formal problem formulations

A recurring formalization treats a KG as $G=(\mathcal{E},\mathcal{R},\mathcal{T})$, where $\mathcal{E}$ is the entity set, $\mathcal{R}$ the relation set, and $\mathcal{T}$ the triple set; multilingual KGC then predicts a missing tail $t$ from $(h,r,?)$ or a missing head $h$ from $(?,r,t)$ across multiple language-specific KGs $G_1,\dots,G_M$ (Tang et al., 2023). In parallel, multilingual fact linking formulates the problem at the text–KG interface. Given a sentence $T_m$ in language $\mathcal{E}$0, the objective is to recover the subset of KG facts it expresses, even when the corresponding fact label is unavailable in that language: $\mathcal{E}$1 This formulation explicitly targets multilingual alignment between raw text and structured facts rather than monolingual relation extraction (Kolluru et al., 2021).

A second axis of formal variation concerns whether the multilingual signal comes from aligned KGs, from unstructured evidence, or from iterative graph fusion. In graph-fusion settings, the task is to merge an evolving fused graph $\mathcal{E}$2 with an incoming language graph $\mathcal{E}$3 by identifying alignment pairs over entities and relations, then collapsing matched nodes and folding equivalent predicates into a shared schema (Kyaw et al., 22 Mar 2026). Resource-centric constructions such as FactNet define still another variant: atomic assertions $\mathcal{E}$4 are canonicalized into equivalence classes and linked to language-specific evidence pointers, so that multilingual construction is simultaneously structural and provenance-preserving (Shen et al., 3 Feb 2026).

This range of formulations matters because mKGC is often conflated with either entity alignment alone or monolingual link prediction with multilingual labels. The literature does not support that reduction. The available work instead suggests a broader design space in which multilingual construction can proceed from multilingual text to KG facts, from multiple partially aligned KGs to completed graphs, or from multilingual evidence corpora to large grounded factual infrastructures.

2. Construction from multilingual text, semi-structured corpora, and KG translation

Multilingual Fact Linking (MFL) addresses the case where text is multilingual but fact labels are sparse. The key observation is that a fact label in a language $\mathcal{E}$5 exists only if the subject, relation, and object labels all exist in $\mathcal{E}$6; otherwise the fact has no usable label in that language. To study this problem, IndicLink was introduced with 11,293 linked Wikidata facts and 6,429 sentences spanning English and six Indian languages, with test coverage reported as 1,002 English examples, 889 Hindi, 888 Telugu, 881 Tamil, 1,001 Urdu, 881 Gujarati, and 887 Assamese. The corresponding fact-label coverage is sharply skewed: English has 4.6M facts, Hindi 230K, Telugu 145K, Tamil 248K, Urdu 361K, Gujarati 91K, and Assamese 257. The proposed ReFCoG architecture combines dual-encoder retrieval over millions of facts via FAISS with constrained Seq2Seq generation over a prefix trie of valid English fact labels; its best retrieval strategy is All-Sum, and the full Retrieval+Generation system improves over classification-based reranking by 10.7 points in Precision@1 and 15.2 points in Recall@5, reaching an average P@1 of 52.1 and average R@5 of 51.5 across the seven languages (Kolluru et al., 2021).

Low-resource text-first construction appears in BanglaAutoKG, which is designed to construct Bangla KGs from arbitrary Bangla text. Its pipeline uses ChatGPT 3.5 for multilingual entity and relation extraction, a Bangla-to-English translation dictionary to identify English equivalents and related terms, pretrained BERT features for the initial graph, graph-attention-based feature denoising, local spectral filtering, and a GNN-based semantic filter that prunes weak edges. The reported ablations use Average Semantic Feature Alignment Score (A-SFAS): on a poem, the score rises from 0.451 without semantic filtering to 0.892 with it, and from 0.528 without feature denoising to 0.892 with denoising; on a Wikipedia case, the corresponding changes are 0.527 to 0.912 and 0.618 to 0.912 (Wasi et al., 2024). This suggests that, in very low-resource settings, multilingual semantic priors and graph filtering are used not only for transfer but also for basic graph denoising.

QuoteKG exemplifies multilingual KG construction from semi-structured web corpora. Its creation pipeline extracts quotes and contexts from Wikiquote language editions, enriches them with dates, sources, misattribution status, linked entities, and sentiment, then aligns mentions of the same quote across languages by clustering multilingual sentence embeddings from XLM-RoBERTa with cosine threshold 0.8 via UKPLab’s Fast Clustering algorithm. The resulting RDF graph contains 880,878 quotes, 961,535 quote mentions, 411,912 mentions with context, 69,467 people, and 55 languages; 13,104 quotes are labeled as misattributed or falsely claimed, and 38,931 quotes have mentions in different languages. On a manually aligned gold set over English, German, and Italian, the reported alignment result is Precision $\mathcal{E}$7, Recall $\mathcal{E}$8, and F1 $\mathcal{E}$9 (Kuculo et al., 2022).

A different route is cross-lingual KG translation and enrichment. In the THOTH pipeline, facts are treated as sentences in which URIs are tokens; English RDF triples are translated into German with RNN-based NMT, split between resource-only triples and literal-containing triples, and then merged into the target KG. THOTH reports a BLEU score of 65.47, up to 80% accuracy for subjects, predicates, and objects, and full-triple accuracy of 68.83%. The enriched German DBpedia improves FactBench fact validation by 18.4% and improves MAG-based entity linking by about 19% F-measure (Moussallem, 2020).

3. Completion through alignment, fusion, and collective reasoning over multiple KGs

The classical multilingual completion literature assumes several language-specific KGs with partial seed alignments and seeks to exploit complementary knowledge across them. KEnS embeds all KGs into a shared space, expands entity alignments by self-learning with CSLS mutual nearest neighbors, and performs ensemble inference over language-specific predictors. It evaluates three ensemble variants—majority voting, MRR-weighted combination, and a boosting-based scheme—and reports that the general ordering is $\mathcal{R}$0. On DBP-5L, the largest gains occur on the sparsest language-specific graph: for Greek with RotatE, baseline Hits@1 is 14.5, while $\mathcal{R}$1 reaches 27.5 (Chen et al., 2020). The broader implication is that cross-KG transfer is most useful when the target graph is structurally sparse.

SS-AGA replaces rigid alignment regularization with adaptive cross-graph reasoning. All KGs are fused into a single heterogeneous graph by converting each seed alignment pair into two directed facts under a new relation $\mathcal{R}$2, so multilingual correspondence becomes a graph edge rather than an external loss. A relation-aware attention GNN then controls information propagation over both factual and alignment edges, while a self-supervised new pair generator proposes additional alignments using structural embeddings, multilingual BERT text embeddings, CSLS, and a masked alignment recovery objective. On DBP-5L, SS-AGA reaches 30.8 Hits@1, 58.6 Hits@10, and 35.3 MRR on Greek; on the industrial multilingual E-commerce KG E-PKG, it reports 33.9 Hits@1, 74.1 Hits@10, and 48.3 MRR on Japanese (Huang et al., 2022). The model is explicitly motivated by KG inconsistency across languages, not only by seed scarcity.

Joint optimization of completion and alignment is the central idea of JMAC. The model uses two relation-aware GNN encoders, one for completion and one for alignment, together with a structural inconsistency reduction mechanism that injects completion embeddings into the alignment component, and EnTr, a mechanism for alignment seed enlargement and triple transferring. The completion module uses a TransE-style score at every layer, the alignment module learns cosine-distance-based alignment rankings, and training alternates between the two objectives. JMAC is reported to obtain new state-of-the-art results on both MKGC and MKGA on DBP-5L, and also achieves the best results on OpenEA 15K / DBP1.0 bilingual alignment benchmarks (Tong et al., 2022). The methodological claim is not merely that alignment helps completion, but that completion can reduce the structural inconsistency that degrades alignment.

CKGC-CKD formulates the same multilingual setting as collective completion over local KGs and one global fused KG. Each language-specific KG is encoded by a local CompGCN model, while a fused KG adds seed alignments as edges with a special alignment-specific message rule. Mutual knowledge distillation then transfers head-prediction, tail-prediction, and relation-prediction distributions between local and global models, subject to a performance-aware teacher rule. On DBP-5L, CKGC-CKD reports 49.2 / 88.5 / 63.9 on Greek and 48.4 / 84.6 / 60.9 on Japanese for Hit@1 / Hit@10 / MRR; on E-PKG it reports 47.5 / 85.5 / 62.1 on Japanese (Zhang et al., 2023). This line of work treats multilingual KGs as complementary views of a partially shared world, with alignment edges functioning as channels for message passing rather than only as constraints on final embeddings.

4. PLM-based generation, retrieval augmentation, and efficient multilingual knowledge sharing

A distinct family of methods formulates mKGC as conditional generation. In the knowledge-constrained PLM approach built on Prix-LM, a query $\mathcal{R}$3 is serialized with special role tokens $\mathcal{R}$4, $\mathcal{R}$5, and $\mathcal{R}$6, and the model is trained under three objectives: autoregressive answer generation, a global knowledge constraint enforcing triple-level semantic consistency through a TransE-like score,

$\mathcal{R}$7

and a local knowledge constraint that maximizes mutual information between the query words and the tail representation via a Jensen–Shannon MI estimator. Using XLM-R Base on the Prix-LM DBpedia benchmark over German, Finnish, French, Hungarian, Italian, Japanese, and Turkish, the reported average results are Hits@1 21.39 versus 19.04 for Prix-LM, Hits@3 31.99 versus 28.58, and Hits@10 43.50 versus 37.49, corresponding to +12.32% relative improvement on Hits@1 and +16.03% on Hits@10 (Song et al., 2024). The intended effect is to reduce type ambiguity and better align multilingual PLMs with KG reasoning.

mRAKL reformulates multilingual KG completion as multilingual question answering. Given a triple $\mathcal{R}$8, a manually prepared relation template turns $\mathcal{R}$9 into a question, the tail entity becomes the answer, and retrieval supplies a passage $\mathcal{T}$0 so that the generator models $\mathcal{T}$1. The retriever is either BM25 or LaBSE; the generator is AfriTeVa-base fine-tuned with LoRA. The study focuses on Tigrinya and Amharic, with Arabic and English as transfer languages. Zero-shot parametric probing remains below 6% H@1 for all tested LMs, while the RAG setting outperforms the no-context configuration. Under an idealized retrieval analysis, mRAKL improves H@1 by 4.92 percentage points for Tigrinya and 8.79 percentage points for Amharic, and multilingual self-context exceeds monolingual self-context by 4.69 percentage points for Tigrinya (Nigatu et al., 21 Jul 2025). The design assumption is that open-domain multilingual text can compensate for sparse structured supervision in low-resource languages.

Parameter-efficient LLM reranking is taken further in the KL-GMoE and IER framework. The paper constructs a new five-language dataset from Wikidata/Wikidata5M-style resources with EN, FR, IT, JA, and ZH, comprising 351,299 entities, 2,264 relations, and 3,028,767 triples. Candidate entities are first retrieved by a KGE model, then reranked by Llama-2-7b-chat-hf augmented with Knowledge-level Grouped Mixture of Experts in the FFN layers and Iterative Entity Reranking over 10 rounds. The full system reports average Hits@1 41.88, Hits@3 50.93, Hits@10 58.78, and MRR 47.86, with improvements over GC-PLM of +5.47% Hits@1, +3.27% Hits@3, +1.01% Hits@10, and +6.13% MRR (Mao et al., 9 Oct 2025). The routing analysis further reports that samples with the same relation across languages often activate the same experts, which the paper interprets as evidence of language-agnostic shared knowledge.

5. Privacy-preserving and LLM-mediated multilingual fusion

FedMKGC addresses a practical constraint absent from most alignment-based multilingual KGC systems: raw KG data may be sensitive or proprietary. In FedMKGC, each KG is treated as a client, each client trains a local text-based KGC model, and a central server aggregates model weights rather than triples or entity alignments. The model uses trainable relation parameters inspired by prefix-tuning, mean pooling with $\mathcal{T}$2 normalization, a SimKGC-style contrastive objective, and weighted parameter averaging analogous to FedAvg, simplified to uniform averaging in the implementation. On DBP-5L, FedMKGC reports Hits@1 / Hits@10 / MRR of 40.1 / 73.7 / 52.2 on Greek, 35.4 / 64.6 / 45.4 on Japanese, 34.6 / 64.7 / 45.3 on French, 34.8 / 64.0 / 45.2 on Spanish, and 30.0 / 61.2 / 40.8 on English; the paper states that these results are competitive with alignment-based baselines such as KEnS and SS-AGA without requiring labeled alignments or raw data sharing (Tang et al., 2023). The privacy claim rests on locality of raw data and the absence of explicit entity alignment.

LLM-mediated graph fusion removes the alignment-heavy assumption in a different way. In the sequential agglomeration framework for cross-lingual graph fusion, each triple is linearized as

$\mathcal{T}$3

and an LLM reasons over linearized neighborhoods from the current fused graph $\mathcal{T}$4 and an incoming graph $\mathcal{T}$5. The process groups triples by head entity, exhaustively pairs batches, prompts the LLM for JSON outputs containing entity_alignments and relation_alignments, aggregates repeated predictions by maximum confidence, and merges accepted alignments into the evolving graph. On DBP15K zh_en, the study reports 5,416 unique alignment predictions, 3,543 true positives, 1,873 false positives, 65.4% precision, 23.6% recall, and 34.7% F1; with confidence threshold $\mathcal{T}$6, the precision rises to 88.0% with 23.5% recall and 37.0% F1, and the reported Hits@1 accuracy is 88.3% (Kyaw et al., 22 Mar 2026). The method is explicitly described as exploratory, with low recall attributed to context window limits and exhaustive batching constraints.

These two paradigms—federated learning and LLM graph fusion—target different operational bottlenecks. FedMKGC is designed for privacy-sensitive multi-organization settings with non-IID client KGs, whereas sequential LLM fusion is designed for schema drift, lexical heterogeneity, and continual N-graph agglomeration. Both depart from the earlier assumption that multilingual construction must be driven by a fixed set of seed alignments and direct raw-data exchange.

6. Benchmarks, large-scale infrastructures, and recurring limitations

The benchmark ecosystem reflects the heterogeneity of the task. IndicLink supports multilingual fact linking with Precision@1, Recall@5, and macroP@1 (Kolluru et al., 2021). DBP-5L is a standard multilingual completion benchmark used by KEnS, SS-AGA, JMAC, and FedMKGC, with filtered-ranking metrics such as Hits@1, Hits@10, and MRR (Chen et al., 2020). E-PKG extends this setting to industrial multilingual e-commerce KGs (Huang et al., 2022). QuoteKG functions as both a multilingual KG and an alignment benchmark over quotes (Kuculo et al., 2022). FactNet-Bench introduces a broader evaluation suite covering KGC, multilingual KG question answering, and multilingual fact checking with evidence-based metrics and split safeguards at the FactSynset level (Shen et al., 3 Feb 2026).

FactNet is presently the largest resource in this set. It aligns Wikidata statements with native Wikipedia evidence across 316 Wikipedia editions and reports 1.70B FactStatements, 1.55B FactSynsets, 3.01B FactSenses, and 3.69B RelationEdges, spanning 12.1K properties and occupying 894 GB on disk. The resource is deterministic: it uses only Wikidata JSON dumps, Wikipedia XML dumps, and SQL link tables; stores provenance via snapshot page and revision IDs plus normalized offsets; and assigns split membership through stable hashing. The reported corpus-level FactSense grounding precision is $\mathcal{T}$7 with 95% CI $\mathcal{T}$8, low-resource Tier 3 precision is 0.885, and a 1M-item relocalization test reports 99.63% exact relocalization (Shen et al., 3 Feb 2026). For mKGC, this shifts emphasis from purely predictive completion toward auditable multilingual factual grounding.

Several limitations recur across otherwise different methods. Fact-label sparsity remains acute in low-resource languages, as shown by the discrepancy between English and Indian-language coverage in IndicLink (Kolluru et al., 2021). Generative PLM methods remain affected by English-centric pretraining and low knowledge density in non-English KGs (Song et al., 2024). Retrieval-augmented methods are bottlenecked by retrieval quality and sometimes rely on manual relation templates (Nigatu et al., 21 Jul 2025). Alignment-based graph methods remain sensitive to noisy or scarce seed alignments and to structural inconsistency between graphs (Huang et al., 2022). Federated methods face non-IID heterogeneity and client-participation trade-offs (Tang et al., 2023). LLM fusion methods are presently precision-oriented but recall-limited (Kyaw et al., 22 Mar 2026). A further source of confusion is terminological: in some papers, “MKGC” denotes multimodal knowledge graph completion rather than multilingual knowledge graph construction, as in DiffusionCom, which targets multimodal KGs with images and text rather than cross-lingual graphs (Huang et al., 9 Apr 2025).

Taken together, the literature portrays mKGC as a layered problem rather than a single benchmark task. Construction may begin from multilingual sentences, semi-structured quote pages, translated RDF triples, aligned multi-KG graphs, federated local models, or provenance-linked Wikipedia evidence. Completion and alignment are often mutually reinforcing, but they are not universally sufficient; multilingual text grounding, retrieval, and auditable evidence have become equally central. This suggests that future mKGC systems will continue to combine scalable retrieval, multilingual representations, graph-structural reasoning, and explicit validity or provenance constraints rather than relying on a single alignment or embedding mechanism.