FedV-KGQA in Practice: Design Lessons and an Interactive Prototype

# FedV-KGQA in Practice: Design Lessons and an Interactive Prototype
Source: [https://arxiv.org/html/2609.13661](https://arxiv.org/html/2609.13661)
###### Abstract

Knowledge graph question answering usually assumes that one system can reach the whole graph\. In practice, facts are often held by organizations that share entity identifiers but own disjoint relation types, so no single party sees a complete reasoning chain\. This poster presents the empirical findings of FedV\-KGQA on multi\-hop question answering over such vertically partitioned graphs\. Each silo enriches its local graph and trains a knowledge graph embedding on its own triples\. A server then concatenates the silo\-specific entity views, anchors the projected question at the topic entity, and ranks candidates by similarity\. Raw triples and relation embeddings never leave a silo\. Comparing the FedV\-KGQA experiments with one another yields three results\. First, federated fusion recovers most of the centralized accuracy, while a single silo recovers little\. Second, anchoring and enrichment matter more than the choice of embedding model\. Third, the cheapest encoder depends on the target accuracy rather than on parameter count\. This poster paper contributes that cross\-experiment comparison, four design lessons drawn from it, and an interactive prototype that runs real inference and traces the full pipeline, per question, on released checkpoints\.

###### keywords

Distributed Knowledge Graphs ,Knowledge Graph Question Answering ,Multihop Reasoning

††copyrightyear:2026††copyright:Copyright for this paper by its authors\. Use permitted under Creative Commons License Attribution 4\.0 International \(CC BY 4\.0\)\.††venue:ISWC 2026 Companion Volume, October 25–29, 2026, Bari, Italy††email:islamm9@rpi\.edu††email:senevo@rpi\.edu††address:Rensselaer Polytechnic Institute, Troy, NY 12180, USA## 1Introduction

Multi\-hop knowledge graph question answering \(KGQA\) chains several facts to reach an answer entity: answering “which actors starred in films directed by Christopher Nolan?” takes one hop to the films and a second to their cast\. Almost all KGQA systems assume a single system can query the whole knowledge graph \(KG\) to follow such a chain[Saxena et al\. \(2020\)](https://arxiv.org/html/2609.13661#bib.bib1);[Jiang et al\. \(2023\)](https://arxiv.org/html/2609.13661#bib.bib2), and recent methods that add large language models keep that assumption, since the evidence is gathered in one place[Luo et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib3);[Ma et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib4)\. That assumption fails when the hops reside in different organizations: a film studio may holddirectorrelations, a streaming platformcastrelations, and a metadata servicegenrerelations, so each party owns a slice of the same entities and none can trace a chain across the slices, while governance, sovereignty, and commercial intellectual property constraints usually forbid merging the triples\. This is a*vertical*partition of the KG, in which parties share entities as samples but hold disjoint relations as feature views[Khan et al\. \(2024\)](https://arxiv.org/html/2609.13661#bib.bib23);[Tran et al\. \(2024\)](https://arxiv.org/html/2609.13661#bib.bib24)\. Existing federated knowledge graph embedding \(KGE\) work instead assumes horizontal partitions, where clients share the relation vocabulary but hold different triples, and targets link prediction rather than multi\-hop QA[Chen et al\. \(2021\)](https://arxiv.org/html/2609.13661#bib.bib5);[Hu et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib6);[Zhu et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib7)\. Federated SPARQL engines answer queries across autonomous sources, and recent work continues to advance source selection and scale to large federations[Ogura et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib22);[Aimonier\-Davat et al\. \(2024\)](https://arxiv.org/html/2609.13661#bib.bib25)\. These engines require queryable endpoints and a structured query, whereas our silos expose only embeddings and run no query over their triples\. Neither line covers a vertical split answered from natural language, which is where FedV\-KGQA sits, ranking answers across relation silos while keeping triples and relation parameters inside each silo[Bappy and Seneviratne \(2026\)](https://arxiv.org/html/2609.13661#bib.bib8)\.

FedV\-KGQA[Bappy and Seneviratne \(2026\)](https://arxiv.org/html/2609.13661#bib.bib8)answers six research questions in separate sections but never compares them, and that comparison is what this poster adds: model choice looks decisive until the ablation shows anchoring and enrichment matter more; silo count is a communication parameter in one section and a representation\-quality parameter in another; and encoder choice looks settled by accuracy until the communication experiment shows the cheapest encoder depends on the target\. Section[6](https://arxiv.org/html/2609.13661#S6)turns that comparison into four design lessons, and an interactive prototype \(Section[5](https://arxiv.org/html/2609.13661#S5)\) runs live inference on released checkpoints, tracing a single question through every stage of the pipeline, a per\-question view FedV\-KGQA[Bappy and Seneviratne \(2026\)](https://arxiv.org/html/2609.13661#bib.bib8)does not provide\.

## 2Framework Overview

Figure[1](https://arxiv.org/html/2609.13661#S2.F1)shows the four stages: local enrichment, local KGE training, server\-side fusion with QA training, and inference\. Each silo first enriches its local graph with inverse\-property and property\-chain rules, without any triple leaving the silo, which improves entity representation quality and answer reachability\. The inverse rules matter more than they look, because under a vertical split an answer entity often appears only as a tail in its own silo, so its embedding is trained from one role only and gives a weak similarity signal; adding the inverse triple restores that signal\. Each silo also precomputes a topic\-conditioned candidate set by bounded neighborhood expansion, then trains a local KGE model on its enriched triples, for which we test TransE[Bordes et al\. \(2013\)](https://arxiv.org/html/2609.13661#bib.bib12), DistMult[Yang et al\. \(2014\)](https://arxiv.org/html/2609.13661#bib.bib13), ComplEx[Trouillon et al\. \(2016\)](https://arxiv.org/html/2609.13661#bib.bib15), and RotatE[Sun et al\. \(2019\)](https://arxiv.org/html/2609.13661#bib.bib14)\. Raw triples and relation embeddings stay local, and only the entity embedding matrix is shared\.

The server concatenates the silo\-specific entity embeddings into a joint matrix𝐇joint\\mathbf\{H\}\_\{\\mathrm\{joint\}\}instead of averaging them, which preserves the geometric view of each silo\. A frozen encoder maps the question to a vector, a multilayer perceptron \(MLP\) projects it into the joint entity space, and the question is anchored at the topic entitye0e\_\{0\},

𝐪anch=MLP⁡\(Enc⁡\(q\)\)\+𝐇joint​\[e0\],\\mathbf\{q\}\_\{\\mathrm\{anch\}\}=\\mathrm\{MLP\}\(\\mathrm\{Enc\}\(q\)\)\+\\mathbf\{H\}\_\{\\mathrm\{joint\}\}\[e\_\{0\}\],\(1\)which grounds ranking in the relevant neighborhood\. The topic entitye0e\_\{0\}is supplied with each question by the benchmarks and is not predicted, so entity linking lies outside the scope of this evaluation\. Candidates are scored by cosine similarity with𝐪anch\\mathbf\{q\}\_\{\\mathrm\{anch\}\}\. The MLP is the only trainable component\. Given a question whose gold answers form the set𝒜\\mathcal\{A\}, QA training minimizes a margin ranking loss over the precomputed candidate set,

ℒQA=max⁡\(0,γ\+maxe−∈𝒞⁡\(e0\)∖𝒜⁡score⁡\(e−\)−maxe\+∈𝒜⁡score⁡\(e\+\)\),\\mathcal\{L\}^\{\\mathrm\{QA\}\}=\\max\\\!\\Big\(0,\\;\\gamma\+\\\!\\\!\\max\_\{e^\{\-\}\\in\\,\\mathcal\{C\}\(e\_\{0\}\)\\setminus\\mathcal\{A\}\}\\\!\\\!\\mathrm\{score\}\(e^\{\-\}\)\\;\-\\;\\max\_\{e^\{\+\}\\in\\,\\mathcal\{A\}\}\\mathrm\{score\}\(e^\{\+\}\)\\Big\),\(2\)which penalizes the model whenever the hardest negative candidate scores close to the best gold answer\. Because the encoder is frozen, the projection head alone learns to place question vectors near the answer region of the joint space, while relation embeddings receive no gradient and stay at their local values\. The gradient with respect to the concatenated entity matrix decomposes into silo\-specific column slices, so each slice goes back only to its own silo\. At inference the server reuses the fine\-tuned entity matrices and the precomputed candidates, without runtime graph traversal or cross\-silo triple exchange\.

### Scope\.

The method answers path\-shaped questions\. Each question comes with a topic entity, and the answer must be an entity that the bounded expansion reaches\. Questions that need aggregation, comparison, counting, or time constraints fall outside a ranking formulation, as do questions whose answer is not an entity\. Accuracy also depends on how much structure each silo holds and on how well entity identifiers align across silos\.

![Refer to caption](https://arxiv.org/html/2609.13661v1/workflow2.png)Figure 1:Workflow of FedV\-KGQA\. Only entity embeddings and silo\-specific gradient slices cross the boundary\.

## 3Experimental Design

We use three benchmarks of differing scale and domain: MetaQA \(43,000\+ movie entities, one\- to three\-hop questions\)[Zhang et al\. \(2018\)](https://arxiv.org/html/2609.13661#bib.bib9), PathQuestion \(person\-centric two\- and three\-hop chains from Freebase\)[Zhou et al\. \(2018\)](https://arxiv.org/html/2609.13661#bib.bib10), and WebQuestionsSP \(WebQSP; 985,000\+ entities, open\-domain, up to two hops\)[Yih et al\. \(2016\)](https://arxiv.org/html/2609.13661#bib.bib11)\. Relations are assigned by semantic category to three, five, or seven silos with shared entity identifiers, so each relation belongs to exactly one silo and the silos jointly cover the full vocabulary\. Twelve configurations pair four KGE models with a frozen BERT[Devlin et al\. \(2019\)](https://arxiv.org/html/2609.13661#bib.bib16), DistilBERT[Sanh et al\. \(2019\)](https://arxiv.org/html/2609.13661#bib.bib17), or RoBERTa[Liu et al\. \(2019\)](https://arxiv.org/html/2609.13661#bib.bib18)encoder, evaluated by mean reciprocal rank \(MRR\) and Hits@KK\. The experiments span two\- and three\-hop accuracy, ablations of anchoring and enrichment, adapted federated and centralized baselines, Gaussian perturbation of transmitted entity embeddings, and communication costs required to reach a target Hits@3, with three silos as the default partition\.

Partitions differ substantially in what each silo can learn\. Under the three\-silo split, and counting inverse relations, MetaQA silos hold 4, 3, and 8 relations over 63K, 129K, and 84K enriched triples; PathQuestion silos hold 3, 5, and 5 relations over 18K, 168K, and 191K triples; and WebQSP silos hold 332, 847, and 2,098 relations over 0\.59M, 0\.99M, and 1\.89M triples\. Silo count and per\-silo structure therefore vary together, which is why silo count alone does not predict accuracy\. The*local\-only*baseline used below removes fusion entirely: each silo trains, expands candidates, and ranks using only its own embeddings and its own triples, and we report the best single silo\. Built from one silo alone, the candidate set covers at most 54%, 35%, and 46% of gold answers on MetaQA, PathQuestion, and WebQSP, against 99%, 100%, and 78% when the silos are combined\. Coverage caps attainable accuracy, so WebQSP’s results are measured against a ceiling of 0\.78\.

## 4Empirical Findings

Finding 1: Fusion works, and the model matters less than the mechanism\.With DistilBERT\+TransE, FedV\-KGQA trails a centralized upper bound \(Figure[2](https://arxiv.org/html/2609.13661#S4.F2)\(a\)\) by 0\.04 MRR on MetaQA and 0\.03 on PathQuestion and WebQSP, whereas the best single\-silo baseline drops to 0\.40, 0\.20, and 0\.32\. No single partition holds the full chain\. Yet concatenated entity views recover most of what centralized access gives, reaching 0\.76, 0\.65, and 0\.54 MRR and beating adapted FedE[Chen et al\. \(2021\)](https://arxiv.org/html/2609.13661#bib.bib5)by up to 0\.08 and adapted RelChain[Jin et al\. \(2023\)](https://arxiv.org/html/2609.13661#bib.bib21)by up to 0\.12\. The KGE model is secondary: TransE is most stable \(0\.54 MRR on WebQSP with BERT, against 0\.41 for DistMult, 0\.48 for ComplEx, 0\.49 for RotatE\), but on MetaQA all four range from 0\.71 to 0\.76\. BERT and DistilBERT stay close despite DistilBERT’s smaller size, while RoBERTa trails throughout, suggesting its\[CLS\]vectors align poorly with the KGE spaces\. We evaluate translational and bilinear scoring functions only, since the question here is how fusion and anchoring behave under a vertical split rather than which scoring function is strongest\. Message\-passing encoders, such as graph neural networks \(GNNs\), are excluded for a structural reason: each silo would aggregate over a graph missing most of its edges by construction, so the neighborhood a GNN exploits is precisely what the partition removes\. Whether enrichment restores enough structure to make aggregation worthwhile is left to future work\.

Finding 2: Partitioning and path length behave non\-uniformly\.Silo count is not a simple knob\. With BERT\+TransE, WebQSP falls only from 0\.54 to 0\.51 MRR from three to seven silos and MetaQA peaks at 0\.83 with five, while PathQuestion rises to 0\.68 at five silos then drops to 0\.55 at seven\. So moderate partitioning can aid specialization while heavy fragmentation weakens embeddings, and relation placement shapes representation quality as much as silo count does, as the partition statistics in Section[3](https://arxiv.org/html/2609.13661#S3)indicate\. Longer paths need no architectural change \(Figure[2](https://arxiv.org/html/2609.13661#S4.F2)\(b\)\)\. On MetaQA, BERT\+TransE falls only from 0\.76 to 0\.74 MRR \(0\.93 Hits@10 held\), but on PathQuestion it falls 0\.65 to 0\.57 and 0\.96 to 0\.90, tracking its smaller, more heterogeneous training set\. So longer\-path generalization should not be inferred from one benchmark\.

Finding 3: Both components help, and cost is target\-dependent\.Removing anchoring lowers MRR from 0\.64 to 0\.57 on PathQuestion and 0\.53 to 0\.47 on WebQSP, while removing enrichment causes larger drops on WebQSP \(0\.19 MRR, 0\.28 Hits@10\)\. Enrichment strengthens local structure and candidate reachability, anchoring focuses ranking on the right neighborhood\. Robustness holds at low noise \(Figure[2](https://arxiv.org/html/2609.13661#S4.F2)\(c\)\): atσ=0\.05\\sigma=0\.05, MRR drops only to 0\.79, 0\.59, and 0\.48, and decline steepens aboveσ=0\.10\\sigma=0\.10\. This is a robustness analysis, not a privacy guarantee[Peng et al\. \(2021\)](https://arxiv.org/html/2609.13661#bib.bib19)\. Total communication \(Figure[2](https://arxiv.org/html/2609.13661#S4.F2)\(d\)\) isCtotal=2​T​K​\|ℰ\|​d⋅4C\_\{\\mathrm\{total\}\}=2TK\|\\mathcal\{E\}\|d\\cdot 4bytes, so cost is target\-dependent\. DistilBERT reaches Hits@3 of 0\.35 on WebQSP for 96\.8 vs 121\.1 GB, but BERT is cheaper at higher targets \(23\.1 vs 31\.4 GB on PathQuestion at Hits@3 of 0\.70\), making encoder efficiency a cost\-to\-target question[Zhu et al\. \(2025\)](https://arxiv.org/html/2609.13661#bib.bib7)\. Cost scales with the entity vocabulary, not the triple count, so sparse updates restricted to the candidate sets should reduce it substantially\. We have not measured this\.

![Refer to caption](https://arxiv.org/html/2609.13661v1/result.png)Figure 2:Cross\-experiment synthesis\. \(a\) DistilBERT\+TransE baseline comparison, three silos\. \(b\) BERT\+TransE from two to three hops, three silos\. \(c\) BERT\+TransE robustness, five silos\. \(d\) Communication for selected Hits@3 targets\.
## 5Interactive Prototype

The prototype turns the pipeline into something a viewer can drive\. It loads the released checkpoints \(TransE with BERT, three silos, two\-hop MetaQA\) and runs real inference\. The user picks a question, and each stage streams in turn: the topic entity, the private triple each silo holds, two\-hop candidate filtering, fusion, and the ranked answers with per\-question MRR and Hits\. The silo triples are displayed for demonstration only, to show what each silo holds and how the answer is assembled across silos; in deployment, neither the user nor the server sees them\. The same trace exposes properties that aggregate metrics hide\. First, the*cross\-silo dependency*is visible per query\. One can see which silo supplies each hop and confirm that no silo holds the whole chain\. Second, the*privacy boundary*is visible in what moves: only the shared entity embeddings, never the per\-silo triples\. FedV\-KGQA[Bappy and Seneviratne \(2026\)](https://arxiv.org/html/2609.13661#bib.bib8)does not show this per\-question view, so a reader cannot see how any single question was answered\.

Table[1](https://arxiv.org/html/2609.13661#S5.T1)is one such run\. For*“who are the actors in the films written by John Travis?”*the first hopwritten\_bysits in Silo A and the second hopstarred\_actorsin Silo B, so the question is unanswerable inside any one of them\. Fusion still places a gold answer at rank 2 and recovers all three gold answers in the top six, while only entity embeddings ever leave a silo\. Stepping through other questions shows both where the mechanism fails and succeeds\. The coverage reported in the trace distinguishes a failure at the filtering stage from one where the answer was present but scored too low\. The averages in Section[4](https://arxiv.org/html/2609.13661#S4)cannot make that distinction\.111Source and prototype:[https://github\.com/brains\-group/fedv\-kgqa\-source\.git](https://github.com/brains-group/fedv-kgqa-source.git)and[https://github\.com/brains\-group/fedv\-kgqa\-prototype\.git](https://github.com/brains-group/fedv-kgqa-prototype.git)

Table 1:A live prototype run\.User selects:*“who are the actors in the films written by John Travis?”*The prototype then shows each stage in order\. Gold answers:Haley Bennett,Chace Crawford,Jake Weber\.StageLive output❶Topic entityJohn Travis\(id 15311\), seeks*person*❷Silo A\(private\)\(The Haunting of Molly Hartley,written\_by, John Travis\)Silo B\(private\)\(The Haunting of Molly Hartley,starred\_actors, Chace Crawford\)Silo C\(private\)\(The Haunting of Molly Hartley,associated\_genre, Horror\)❸2\-hop filter58 / 39,170 entities scored \(0\.1%\); 3/3 gold covered❹Fuse \+ encode𝐡joint​\[John Travis\]∈ℝ768\\mathbf\{h\}\_\{\\mathrm\{joint\}\}\[\\text\{John Travis\}\]\\in\\mathbb\{R\}^\{768\}; BERT\+MLP question vector❺Ranked outputRobert Duvall 0\.29\>\>Jake Weber0\.23 ✓\>\>Sarah Polley 0\.15\>\>…Result:best rank 2 — MRR 0\.50 — Hits@10 ✓*\(only entity embeddings left each silo\)*
## 6Cross\-Experiment Lessons and Limitations

The results support four design lessons\. Concatenation should preserve each silo’s geometric view, since averaging removes it \(Finding 1\)\. Silo count and relation composition both matter, as moderate partitioning may aid specialization while heavy fragmentation costs accuracy \(Finding 2\)\. Questions should be anchored to the topic entity, since language semantics alone do not identify the correct neighborhood \(Finding 3\)\. And communication should be reported against a target, since the cheapest encoder depends on the operating point \(Finding 3\)\. These lessons hold under the evaluation’s assumptions of a static graph, a shared and aligned entity vocabulary, and semantic relation partitions\. Entity embeddings reach the server, so the boundary is structural, not a formal privacy guarantee\. The candidate sets also leak information: building𝒞⁡\(e0\)\\mathcal\{C\}\(e\_\{0\}\)across silos shows the server that an entity sits within two hops of the topic entity, even though the relation types stay hidden\. How much a server could recover from the embeddings and the candidate sets is an open question\. Future work should study incremental updates, uncertain entity alignment[Chen et al\. \(2024\)](https://arxiv.org/html/2609.13661#bib.bib20), alternative partitioning, message\-passing encoders under a vertical split, secure aggregation, and differential privacy\. Silos may also be incomplete, since a missing edge can break a cross\-silo path that no single silo can repair locally\.

## 7Conclusion

FedV\-KGQA shows that multi\-hop KGQA stays feasible when relation types are split vertically across organizations, approaching centralized accuracy, extending to three\-hop reasoning, and holding up under moderate embedding noise\. Comparing the experiments directly, rather than one at a time, is what surfaces the design lessons, and the prototype makes the mechanism inspectable one question at a time\. These findings carry one message: effective federation takes more than keeping triples local, since enrichment, geometry\-preserving fusion, topic anchoring, and cost\-aware model selection all decide whether it works\. The vertical split changes what each silo can learn, not only where the data sits, and that is what these mechanisms have to compensate for\. Whether the same holds when graphs change over time, or when entity alignment is imperfect, remains an important direction for future evaluation\.

## Declaration on Generative AI

During the preparation of this work, the authors used Claude \(Anthropic\) and OpenAI Codex for language editing, grammar correction, and structural refinement\. The authors reviewed and edited all content and take full responsibility for the submission\. No generative AI tool was used to generate research findings, experimental results, tables, or citations\. This disclosure complies with CEUR’s Policy on AI\-Assisting Tools\.

## References

- Aimonier\-Davatet al\.\(2024\)J\. Aimonier\-Davat, B\. Nédelec, M\. Dang, P\. Molli, and H\. Skaf\-MolliFedUP: querying large\-scale federations of SPARQL endpoints\.InProceedings of the ACM Web Conference 2024,pp\. 2315–2324\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Bappy and Seneviratne \(2026\)M\. S\. I\. K\. Bappy and O\. SeneviratneFedV\-KGQA: Multi\-Hop Question Answering over Vertically Partitioned Knowledge Graphs\.arXiv preprint arXiv:2608\.24846\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1),[§1](https://arxiv.org/html/2609.13661#S1.p2.1),[§5](https://arxiv.org/html/2609.13661#S5.p1.1)\.
- Bordeset al\.\(2013\)A\. Bordes, N\. Usunier, A\. Garcia\-Duran, J\. Weston, and O\. YakhnenkoTranslating embeddings for modeling multi\-relational data\.Advances in neural information processing systems26\.Cited by:[§2](https://arxiv.org/html/2609.13661#S2.p1.1)\.
- Chenet al\.\(2024\)D\. Chen, H\. Zhu, J\. Gu, R\. Chen, and M\. XieUnaligned federated knowledge graph embedding\.InInternational Semantic Web Conference,pp\. 180–198\.Cited by:[§6](https://arxiv.org/html/2609.13661#S6.p1.1)\.
- Chenet al\.\(2021\)M\. Chen, W\. Zhang, Z\. Yuan, Y\. Jia, and H\. ChenFedE: Embedding knowledge graphs in federated setting\.InProceedings of the 10th international joint conference on knowledge graphs,pp\. 80–88\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1),[§4](https://arxiv.org/html/2609.13661#S4.p1.1)\.
- Devlinet al\.\(2019\)J\. Devlin, M\. Chang, K\. Lee, and K\. ToutanovaBERT: Pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 \(long and short papers\),pp\. 4171–4186\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Huet al\.\(2025\)Q\. Hu, W\. Jiang, H\. Li, Z\. Wang, J\. Bai, Q\. Mao, Y\. Song, L\. Fan, and J\. LiLearning federated neural graph databases for answering complex queries from distributed knowledge graphs\.Transactions on Machine Learning Research\.External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=3K1LRetR6Y)Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Jianget al\.\(2023\)J\. Jiang, K\. Zhou, X\. Zhao, and J\. WenUniKGQA: unified retrieval and reasoning for solving multi\-hop question answering over knowledge graph\.InThe Eleventh International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=Z63RvyAZ2Vh)Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Jinet al\.\(2023\)W\. Jin, B\. Zhao, H\. Yu, X\. Tao, R\. Yin, and G\. LiuImproving embedded knowledge graph multi\-hop question answering by introducing relational chain reasoning\.Data Mining and Knowledge Discovery37\(1\),pp\. 255–288\.Cited by:[§4](https://arxiv.org/html/2609.13661#S4.p1.1)\.
- Khanet al\.\(2024\)M\. S\. I\. Khan, A\. Gupta, O\. Seneviratne, and S\. PattersonFed\-RD: Privacy\-preserving federated learning for financial crime detection\.In2024 IEEE Symposium on Computational Intelligence for Financial Engineering and Economics \(CIFEr\),pp\. 1–9\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Liuet al\.\(2019\)Y\. Liu, M\. Ott, N\. Goyal, J\. Du, M\. Joshi, D\. Chen, O\. Levy, M\. Lewis, L\. Zettlemoyer, and V\. StoyanovRoBERTa: A robustly optimized BERT pretraining approach\.arXiv preprint arXiv:1907\.11692\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Luoet al\.\(2025\)S\. Luo, X\. Lu, Q\. Zhao, and W\. RaoBridging the Gap between Knowledge Graphs and LLMs for Multi\-hop Question Answering\.InProceedings of the 34th ACM International Conference on Information and Knowledge Management,pp\. 5006–5010\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Maet al\.\(2025\)C\. Ma, Y\. Chen, T\. Wu, A\. Khan, and H\. WangLarge language models meet knowledge graphs for question answering: synthesis and opportunities\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 24589–24608\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Oguraet al\.\(2025\)Y\. Ogura, T\. Masuda, and T\. AmagasaEfficient Source Selection for Federated SPARQL Queries Using Adjacent Predicate Information\.InInternational Conference on Database and Expert Systems Applications,pp\. 115–129\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Penget al\.\(2021\)H\. Peng, H\. Li, Y\. Song, V\. Zheng, and J\. LiDifferentially private federated knowledge graphs embedding\.InProceedings of the 30th ACM international conference on information & knowledge management,pp\. 1416–1425\.Cited by:[§4](https://arxiv.org/html/2609.13661#S4.p3.1)\.
- Sanhet al\.\(2019\)V\. Sanh, L\. Debut, J\. Chaumond, and T\. WolfDistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter\.arXiv preprint arXiv:1910\.01108\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Saxenaet al\.\(2020\)A\. Saxena, A\. Tripathi, and P\. TalukdarImproving multi\-hop question answering over knowledge graphs using knowledge base embeddings\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 4498–4507\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Sunet al\.\(2019\)Z\. Sun, Z\. Deng, J\. Nie, and J\. TangRotatE: Knowledge graph embedding by relational rotation in complex space\.arXiv preprint arXiv:1902\.10197\.Cited by:[§2](https://arxiv.org/html/2609.13661#S2.p1.1)\.
- Tranet al\.\(2024\)L\. Tran, S\. Chari, M\. S\. I\. Khan, A\. Zachariah, S\. Patterson, and O\. SeneviratneA differentially private blockchain\-based approach for vertical federated learning\.In2024 IEEE International Conference on Decentralized Applications and Infrastructures \(DAPPS\),pp\. 86–92\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1)\.
- Trouillonet al\.\(2016\)T\. Trouillon, J\. Welbl, S\. Riedel, É\. Gaussier, and G\. BouchardComplex embeddings for simple link prediction\.InInternational conference on machine learning,pp\. 2071–2080\.Cited by:[§2](https://arxiv.org/html/2609.13661#S2.p1.1)\.
- Yanget al\.\(2014\)B\. Yang, W\. Yih, X\. He, J\. Gao, and L\. DengEmbedding entities and relations for learning and inference in knowledge bases\.arXiv preprint arXiv:1412\.6575\.Cited by:[§2](https://arxiv.org/html/2609.13661#S2.p1.1)\.
- Yihet al\.\(2016\)W\. Yih, M\. Richardson, C\. Meek, M\. Chang, and J\. SuhThe value of semantic parse labeling for knowledge base question answering\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),pp\. 201–206\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Zhanget al\.\(2018\)Y\. Zhang, H\. Dai, Z\. Kozareva, A\. Smola, and L\. SongVariational reasoning for question answering with knowledge graph\.InProceedings of the AAAI conference on artificial intelligence,Vol\.32\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Zhouet al\.\(2018\)M\. Zhou, M\. Huang, and X\. ZhuAn interpretable reasoning network for multi\-relation question answering\.InProceedings of the 27th international conference on computational linguistics,pp\. 2010–2022\.Cited by:[§3](https://arxiv.org/html/2609.13661#S3.p1.1)\.
- Zhuet al\.\(2025\)X\. Zhu, Y\. Xie, Y\. Liu, Y\. Li, and W\. HuParameter\-efficient federated knowledge graph embedding learning and unlearning\.InInternational Semantic Web Conference,pp\. 462–480\.Cited by:[§1](https://arxiv.org/html/2609.13661#S1.p1.1),[§4](https://arxiv.org/html/2609.13661#S4.p3.1)\.