Recent advancements in the representation of knowledge via knowledge graphs have paved an intuitive way to build scholarly knowledge for users and for artificial agents. Despite the expressiveness of knowledge graphs, accessing these knowledge graphs requires proficiency in a query language such as SPARQL, presenting a barrier for a multitude of users. In this study, we present a hybrid, end-to-end framework that (i) interprets user questions expressed in natural language, (ii) classifies each query into one of four target categories (conferences, authors, organizations, or papers) using a fine-tuned RoBERTa-Large model, (iii) synthesizes candidate SPARQL queries via a large language model (GPT-4o-mini) augmented with few-shot examples, and (iv) refines the raw query results by reranking either the SPARQL output or, when necessary, fallback candidate items retrieved through vector-space embeddings (indexed with FAISS). On a testbed of 92 manually crafted “gold-standard” SPARQL queries, our automated pipeline achieved over 96% overlap with expert results (≥ 70% overlap in 89/92 cases), with perfect consistency on conference, author, and organization queries and 90% coverage on paper queries given the semantic nature of queries. Moreover, our query-type classifier achieved 99% accuracy, demonstrating the reliability of schema selection. These results indicate that combining LLM-driven query synthesis with embeddings-based reranking delivers a robust, user-centric interface to scholarly knowledge graphs, enabling complex information retrieval without SPARQL expertise.
The contemporary development and widespread adoption of knowledge graphs have transformed the
way scholarly knowledge is represented and retrieved. These knowledge graphs, in the context of
scholarly knowledge, ofer powerful tools for organizing and interpreting the vast and growing body of
scholarly content [
graphs to harmonize diverse metadata sources, enable sophisticated analytical queries, and foster deeper insights into the structure and dynamics of academic research.
Despite their evident potential, these scholarly KGs primarily rely on SPARQL, a standardized query
language for RDF data, which is inherently complex and poses a significant barrier for widespread
adoption [
In an efort to bridge this accessibility gap, various methods have been proposed to translate natural
language (NL) queries directly into structured query formats [
CEUR
ceur-ws.org
extensive maintenance [
Furthermore, complementary research employing embedding-based retrieval techniques has demonstrated significant success in handling semantic search and retrieval problems within knowledge bases [15, 16]. Embeddings ofer powerful mechanisms to approximate semantic similarity, enabling rapid approximate-nearest-neighbor searches using high-dimensional vector representations, thus circumventing the strict matching constraints of purely symbolic query methods [17]. Tools such as FAISS exemplify these advances by eficiently managing large-scale embedding indexes and enabling hybrid or fallback retrieval strategies when symbolic queries fail [18].
Motivated by these recent developments, this paper introduces a novel hybrid approach designed
explicitly to simplify access to scholarly KGs for non-expert users. Our method synergistically combines
multiple contemporary techniques:
• Natural-language classification leveraging RoBERTa-Large [
We empirically demonstrate the efectiveness of our hybrid system through rigorous evaluations against a set of 92 manually curated (”gold-standard”) SPARQL queries. Our pipeline achieves over 96% overlap with expert-generated results, surpassing the 70% overlap threshold consistently, and demonstrating perfect consistency in queries targeting conferences, authors, and organizations, alongside robust performance on queries targeting research papers.
The subsequent sections are organized as follows: Section 2 presents related literature on Natural Language-to-SPARQL translation, embedding-based retrieval, and hybrid approaches. Section 3 dives into the background associated with this study, section 4 details our hybrid system architecture. Section 5 outlines the experimental setup, datasets, and evaluation metrics, followed by a detailed discussion of results. Section 6 discusses limitations and future extensions. Section 7 concludes the study with an intuitive summarization of findings.
Early approaches to natural-language interfaces for querying structured knowledge graphs primarily
relied on rule-based grammars, templates, and manually engineered mappings between language
patterns and ontological constructs [
Recent advancements have shifted toward leveraging large language models (LLMs) to enhance natural language interfaces for knowledge graphs. For instance, the EDGE system integrates LLMs to facilitate natural language interactions with educational knowledge graphs, improving user experience in data exploration [21]. Additionally, following framework utilizes Retrieval-Augmented Generation (RAG) to enhance SPARQL query generation, reducing semantic errors and improving robustness [22].
The advent of pre-trained LLMs such as GPT-3 and GPT-4 has significantly reshaped the landscape of semantic parsing and query synthesis. These models can translate natural-language queries into formal queries like SPARQL with minimal annotated examples [23, 24]. Few-shot prompting techniques have been shown to generalize query generation across diverse knowledge schemas, reducing the need for extensive domain-specific annotations [ 24].
To address challenges like hallucinations and out-of-distribution errors in LLM-generated queries, Sharma et al. introduced PGMR (Post-Generation Memory Retrieval), a modular framework that separates query structure generation from knowledge retrieval. This approach significantly reduces the incidence of hallucinated URIs in SPARQL query generation [25]. Frameworks like FRASE leverage frame-semantic structured representations to improve generalization in SPARQL query generation, particularly in scenarios involving naturally phrased, template-free questions [26]. Moreover, SPARKLE integrates knowledge graph structures directly into the decoding process of LLMs, reducing the occurrence of inoperative query generations [27].
Embedding-based retrieval methods provide an efective complementary mechanism for structured
queries by representing entities and their relationships within vector spaces [
Hybrid approaches combining symbolic and neural methods have recently emerged, demonstrating superior performance by reranking symbolic query results using embeddings [29, 30]. Graph-based reranking methods have been explored to enhance the selection of optimal query graphs in Knowledge Base Question Answering systems. Jia and Chen proposed a two-step approach involving initial ranking followed by reranking of query graphs, leading to improved retrieval accuracy [31]. The KGR3 framework integrates retrieval, reasoning, and reranking components to enhance knowledge graph completion tasks, using context-enriched modules to improve prediction accuracy [32]. Additionally, ReranKGC introduces a cooperative retrieve-and-rerank framework for multi-hop knowledge graph completion, improving accuracy and eficiency [ 33].
Our research advances the state-of-the-art by integrating several successful strategies into a unified pipeline tailored specifically for scholarly knowledge graphs. Specifically, our hybrid architecture synthesizes: • Natural-language classification using RoBERTa-Large, achieving near-perfect query-type identification accuracy. • LLM-driven (GPT-4o-mini) SPARQL generation augmented by few-shot examples, significantly reducing reliance on manually annotated datasets. • Embedding-based refinement (FAISS) employed both as a fallback mechanism and a reranking technique to enhance retrieval accuracy.
• LLM-based summarization of the results as according to the user query Our work represents the systematic efort to integrate LLM-based SPARQL generation with embeddingbased reranking specifically targeted at scholarly knowledge graphs, achieving expert-level retrieval quality without extensive manual annotations or user SPARQL expertise.
The past decade has witnessed substantial evolution in scholarly data management, marked by a shift
from isolated bibliographic databases to richly structured, interconnected knowledge graphs (KGs).
Unlike traditional relational databases, scholarly knowledge graphs utilize semantic web standards such as
RDF (Resource Description Framework) and OWL (Web Ontology Language) to integrate heterogeneous
academic metadata seamlessly [34, 35, 36]. These structured graphs empower researchers, libraries,
and institutions to discover relationships among authors, publications, conferences, and institutions
previously concealed in disparate databases [
Yet, the complexity inherent in these rich semantic structures introduces a notable barrier:
traditional querying mechanisms, particularly SPARQL (the standardized language for querying RDF-based
knowledge graphs), are notoriously dificult to master without extensive training [
The growing recognition of this limitation has prompted eforts toward developing intuitive,
userfriendly query interfaces. Central to this vision is the concept of natural language querying, a paradigm
that seeks to leverage users’ inherent linguistic capabilities to interact with complex data structures
without explicit technical knowledge [37]. However, natural language interfaces introduce their own
set of challenges: ambiguity, variability in linguistic expression, and dificulty translating informal
questions into precise symbolic queries [
Simultaneously, advancements in deep learning and natural language processing have opened pathways toward bridging this gap. Particularly transformative are pre-trained large language models (LLMs), capable of generating structured queries such as SPARQL from plain-language user prompts, thereby drastically simplifying user interactions with semantic web resources [38]. Nonetheless, the efective application of LLMs to scholarly data introduces practical concerns, especially regarding accuracy, consistency, and the need for validation mechanisms ensuring generated queries yield meaningful, high-quality results [38].
A parallel and complementary technological advancement has emerged through embedding-based retrieval methods. Embeddings provide semantic vector-space representations of entities (authors, papers, conferences, organizations), capturing subtle semantic relationships beyond direct symbolic matches [35]. By transforming entities into continuous vector representations, embeddings facilitate rapid approximate nearest-neighbor searches, thus enabling eficient retrieval of semantically similar results without rigid symbolic constraints [28]. This technology, operationalized through libraries like FAISS, has become essential in scenarios demanding real-time semantic retrieval at scale [28].
However, embedding-based approaches alone cannot fully leverage structured relationships explicitly encoded in scholarly KGs. Thus, an ideal query mechanism combines symbolic querying (SPARQL) and embedding-based semantic retrieval into a unified, hybrid framework that maximizes the strengths of both paradigms: symbolic precision with embedding-driven semantic flexibility.
Existing methods in scholarly KG querying predominantly focus either on precise symbolic querying (via SPARQL) or purely semantic embedding-based retrieval, with relatively limited exploration of how best to integrate these paradigms efectively in scholarly contexts. The complexity, diversity, and specialized nature of scholarly metadata call for a bespoke hybrid solution specifically designed to address domain-specific queries reliably and intuitively [ 36].
The necessity for a robust hybrid approach is driven by practical considerations: researchers require both accurate and comprehensive query results that align closely with their semantic intent. This demands a mechanism that dynamically utilizes the strengths of symbolic SPARQL querying for structured precision, complemented by embeddings to maintain semantic coherence and manage cases where symbolic retrieval falls short or returns inadequate results [37].
Responding to this pressing need, our work presents a tailored hybrid architecture explicitly designed for scholarly knowledge graphs. We incorporate advanced machine learning methods, specifically a RoBERTa-Large classifier for robust query categorization, GPT-4o-mini for reliable generation of SPARQL queries from natural language inputs, and FAISS-based embeddings for semantic refinement and fallback retrieval. This synthesis uniquely positions our work, ofering a comprehensive solution that addresses existing gaps by seamlessly integrating symbolic query precision and embedding-based semantic flexibility in scholarly knowledge graphs.
The subsequent sections of this paper detail this integration and empirically validate its efectiveness, demonstrating that such a hybrid approach significantly enhances the accessibility and usability of scholarly knowledge graphs for non-technical users.
This study proposes a comprehensive methodology to introduce a natural language interface for querying the scholarly knowledge graph. It employs natural language processing (NLP) techniques, large language models (LLM), semantic embeddings, and symbolic query execution to constitute into a GraphRAG system underlying the natural language query interface.
4.1.1. Dataset Acquisition and Graph Setup We initiated the study by acquiring the ScholarlyData knowledge graph, a prominent RDF-based dataset encompassing structured metadata on conferences, research papers, authors, and organizations. The ScholarlyData graph was imported into a semantic graph database environment (e.g., Apache Jena, GraphDB), enabling eficient storage, indexing, and SPARQL querying. 4.1.2. Semantic Embeddings Creation To support embedding-based retrieval, we computed separate semantic embeddings for each of the four primary entity categories within the scholarly knowledge graph: • Conferences • Authors • Organizations • Papers
Embeddings were generated using a pre-trained transformer-based language model, specifically Sentence-BERT, to convert textual descriptions (e.g., titles, abstracts, names, organizational descriptions) into continuous, dense vector representations capturing semantic meaning. These embeddings were indexed using Facebook’s FAISS library, allowing eficient approximate nearest-neighbor retrieval at query time.
To determine the category of entities targeted by user-submitted natural-language queries, we finetuned a supervised classification method based on RoBERTa-Large, a robust transformer-based language model optimized for text classification tasks. 4.2.1. Automated Federated Dataset Construction Training data for query-type classification was auto generated using GPT-4. We created a federated training/evaluation dataset containing diverse natural-language query examples synthetically paired with their corresponding entity types (conference, author, organization, or paper). This automated process ensured extensive linguistic coverage without manual annotation eforts. Following is the distribution of the dataset:
Model RoBERTa-large (fine-tuned)
Training Samples 772
Test Samples 100 4.2.2. Fine-Tuning RoBERTa-Large Classifier The RoBERTa-Large model was fine-tuned on the generated dataset using a supervised training approach, employing standard hyperparameters (learning rate: 2 × 10−5, batch size: 16, maximum token length: 128). RoBERTa-Large was chosen for its strong performance on a wide range of natural language understanding tasks, particularly in scenarios requiring nuanced contextual representations, making it well-suited for accurately interpreting diverse scholarly queries. This resulted in a high-performance model capable of reliably classifying user queries into the defined entity types, facilitating targeted downstream retrieval.
Following query-type identification, the natural-language queries were transformed into executable SPARQL queries using the GPT-4o-mini language model. 4.3.1. Few-Shot Prompt Engineering To guide GPT-4o-mini in generating accurate SPARQL queries, we implemented a few-shot learning approach tailored for each query type. Specifically, we provided carefully curated examples of naturallanguage queries paired with correctly structured SPARQL queries corresponding to scholarly KG schemas. These prompts ensured GPT-4o-mini could efectively generalize from limited examples, generating syntactically valid and semantically accurate SPARQL queries. 4.3.2. Query Execution The SPARQL queries generated by GPT-4o-mini were directly executed against the semantic knowledge graph database. The results retrieved from these symbolic queries formed the primary candidate set for answering user queries.
To ensure semantic relevance and retrieval quality, we incorporated a hybrid retrieval and reranking strategy that dynamically integrated symbolic SPARQL results with embedding-based semantic similarity computations. 4.4.1. Embedding-based Result Reranking When SPARQL queries returned a suficient number of results ( ≥ 5), the retrieved entities were semantically reranked using cosine similarity scores calculated between their embeddings and the embedding of the user’s natural-language query. This step enhanced semantic coherence and reduced ambiguity inherent in purely symbolic query results. 4.4.2. Embedding-only Retrieval (Fallback Mechanism) If SPARQL queries returned insuficient results ( < 5) or no results, indicating either semantic mismatches or query complexity, our system defaulted directly to embedding-based retrieval. Specifically, we queried the FAISS-indexed embeddings of the identified entity type using the original natural-language query, retrieving the top- nearest entities based on semantic similarity scores. 4.4.3. GraphRAG-based Answer Generation Following the reranking and retrieval of the final results, the conclusive step of the GraphRAG was triggered. The top- embedding results and the original natural-language query were provided back into GPT-4o-mini. Leveraging Retrieval-Augmented Generation, GPT-4o-mini synthesized these retrieved entities into a coherent, contextually relevant, natural-language summary that directly addressed the user’s query.
To validate the efectiveness and accuracy of our proposed hybrid query system, we designed a comprehensive evaluation involving rigorous experimental setups, comparative analyses against manually constructed ground-truth queries, and a detailed assessment of results across multiple dimensions.
5.1.1. Dataset and Gold-Standard Queries We prepared a dataset comprising 92 manually crafted natural-language queries, each paired with expertly constructed corresponding SPARQL queries, hereafter referred to as the gold-standard queries. These queries were selected to reflect realistic academic information needs and were distributed across four distinct categories: • Conferences (22 queries) • Authors (25 queries) • Organizations (20 queries) • Papers (25 queries)
The diversity and representativeness of the queries ensured robust coverage of typical scenarios encountered in scholarly querying tasks. 5.1.2. Evaluation Metrics To quantitatively measure retrieval efectiveness, we employed the following metrics:
Overlap Percentage: Defined as the intersection of retrieved entities by our automated pipeline and the manually retrieved entities (gold-standard), normalized by the size of the gold-standard result set: Overlap = | auto ∩ manual| × 100%
| manual|
Threshold Accuracy: We established a performance threshold at 70% overlap, considering queries achieving this threshold as successfully answered. This threshold was selected to balance recall and precision while remaining practically useful for typical scholarly use cases. 5.1.3. Comparative Baselines For thoroughness, our automated approach was compared primarily against: • Manually constructed SPARQL queries.
• A pure embedding-based retrieval baseline using FAISS without symbolic query generation.
This allowed us to highlight the advantages of our hybrid model in both structured and semantic contexts.
5.2.1. Overall Performance Out of the 92 queries evaluated, our hybrid retrieval system achieved the following outcomes: • Successful Retrieval (≥70% overlap): 89 queries (96%) • Mean Overlap Percentage (all queries): 92.5%
These results indicate that the vast majority of automated queries closely approximated expert-level query quality. 5.2.2. Performance by Entity Type We further analyzed performance separately for each query category to assess consistency across diferent scholarly query contexts.
These breakdowns illustrate particularly robust consistency in conference, author, and organization queries, with only slightly lower performance on paper-related queries, possibly due to higher semantic complexity inherent in publication metadata. 5.2.3. Query Type Classification Accuracy The RoBERTa-Large classifier demonstrated exceptional efectiveness, achieving an accuracy of 99% in predicting query categories on the federated validation set. This high classification accuracy was instrumental in routing each query to the correct generation and retrieval pipeline.
To better understand the benefit of our hybrid strategy, we compared it with a baseline using FAISS embeddings only (no symbolic query generation):
Our hybrid approach significantly outperformed the pure embedding baseline, underscoring the value of combining symbolic SPARQL generation with embedding-based reranking and fallback mechanisms.
The results clearly demonstrate that our hybrid retrieval method efectively bridges the gap between SPARQL expertise and user-friendly natural-language interfaces. Key findings include: • Robust query-type classification: The 99% classification accuracy ensured precise identification of query intent, enabling targeted downstream processing. • Balanced hybrid strategy: The integration of symbolic querying with embedding-based reranking delivered consistent results, addressing weaknesses of either method alone. • Semantic flexibility: Embedding-based reranking improved semantic coherence and compensated for variability in query phrasing.
• The number of test dataset was short, it comprised of 92 manually curated examples. • The federated dataset created for the training of RobertaLarge might contain some bias or lack of proper generalization when tested over a huge number of queries.
Future work will include having a larger evaluation dataset curated by experts, improving entity embeddings, expanding few-shot prompt coverage for LLMs and explore other finetuned llms to generate SPARQL query, and exploring advanced reranking strategies to further enhance the accuracy and robustness of results across all entity types.
In this study, we presented a robust hybrid framework that enables natural-language querying of scholarly knowledge graphs by combining the precision of symbolic SPARQL generation with the lfexibility of embedding-based retrieval. Our system leverages a fine-tuned RoBERTa-Large classifier for query type prediction, GPT-4o-mini for SPARQL generation guided by few-shot prompting, and FAISS-indexed semantic embeddings for result reranking and fallback retrieval. This architecture bridges the gap between expert-level structured querying and accessible, intuitive user interfaces.
Through extensive evaluation against a curated set of 92 gold-standard SPARQL queries, our approach achieved a 96% success rate in matching expert results, with a mean overlap of 92.5%, demonstrating its efectiveness in replicating expert performance. The integration of GraphRAG further ensured meaningful answers even in cases where symbolic retrieval alone was insuficient.
Our findings highlight the practical potential of combining LLM-driven query generation with embedding-based semantic reasoning to improve the accessibility and usability of scholarly knowledge graphs for non-expert users. Future work will explore multilingual query support, dynamic schema adaptation, user feedback integration, and expansion to other domains beyond scholarly metadata.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 955569. The opinions expressed in this document reflect only the author’s view and in no way reflect the European Commission’s opinions. The European Commission is not responsible for any use that may be made of the information it contains.
During the preparation of this work, the author(s) used GPT-4 and Grammarly to assist with paraphrasing, grammar correction, and spelling checks. [14] A. Meloni, D. R. Recupero, F. Osborne, A. Salatino, E. Motta, S. Vahadati, J. Lehmann, Assessing large language models for sparql query generation in scientific question answering, in: CEUR Workshop Proceedings, volume 3953, 2025. [15] H. Peng, H. Li, Y. Song, V. Zheng, J. Li, Diferentially private federated knowledge graphs embedding, in: Proceedings of the 30th ACM international conference on information & knowledge management, 2021, pp. 1416–1425. [16] L. Yue, Y. Zhang, Q. Yao, Y. Li, X. Wu, Z. Zhang, Z. Lin, Y. Zheng, Relation-aware ensemble learning for knowledge graph embedding, arXiv preprint arXiv:2310.08917 (2023). [17] H.-J. Cha, S.-W. Choi, E.-B. Lee, D.-M. Lee, Knowledge retrieval model based on a graph database for semantic search in equipment purchase order specifications for steel plants, Sustainability 15 (2023) 6319. [18] I. A. Ebeid, Medgraph: A semantic biomedical information retrieval framework using knowledge graph embedding for pubmed, Frontiers in big Data 5 (2022) 965619. [19] J. Qi, C. Su, Z. Guo, L. Wu, Z. Shen, L. Fu, X. Wang, C. Zhou, Enhancing sparql query generation for knowledge base question answering systems by learning to correct triplets, Applied Sciences 14 (2024) 1521. [20] S. Ferré, Sparklis: An expressive query builder for sparql endpoints with guidance in natural language, Semantic Web 8 (2016) 405–418. [21] N. Afreen, G. Balloccu, L. Boratto, G. Fenu, F. M. Malloci, M. Marras, A. G. Martis, Edge: A conversational interface driven by large language models for educational knowledge graphs exploration, in: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 5159–5163. [22] M. Arazzi, D. Ligari, S. Nicolazzo, A. Nocera, Augmented knowledge graph querying leveraging llms, arXiv preprint arXiv:2502.01298 (2025). [23] D. Bustamante, H. Takeda, Sparql generation with entity pre-trained gpt for kg question answering, arXiv preprint arXiv:2402.00969 (2024). [24] S. Xu, S. Liu, T. Culhane, E. Pertseva, M.-H. Wu, S. J. Semnani, M. S. Lam, Fine-tuned llms know more, hallucinate less with few-shot sequence-to-sequence semantic parsing over wikidata, arXiv preprint arXiv:2305.14202 (2023). [25] A. Sharma, L. Lara, C. J. Pal, A. Zouaq, Reducing hallucinations in language model-based sparql query generation using post-generation memory retrieval, arXiv preprint arXiv:2502.13369 (2025). [26] P. A. K. Karou Diallo, A. Zouaq, Frase: Structured representations for generalizable sparql query generation, arXiv e-prints (2025) arXiv–2503. [27] J. Lee, H. Shin, Sparkle: Enhancing sparql generation with direct kg integration in decoding,
Expert Systems with Applications (2025) 128263. [28] J. Johnson, M. Douze, H. Jégou, Billion-scale similarity search with gpus, IEEE Transactions on
Big Data 7 (2019) 535–547. [29] J. Lu, K. Hall, J. Ma, J. Ni, Hyrr: Hybrid infused reranking for passage retrieval, arXiv preprint arXiv:2212.10528 (2022). [30] M. Acosta, C. Qin, T. Schwabe, Neuro-symbolic query optimization in knowledge graphs, arXiv preprint arXiv:2411.14277 (2024). [31] Y. Jia, W. Chen, A method of query graph reranking for knowledge base question answering, arXiv preprint arXiv:2204.12808 (2022). [32] M. Li, C. Yang, C. Xu, X. Jiang, Y. Qi, J. Guo, H.-f. Leung, I. King, Retrieval, reasoning, re-ranking: A context-enriched framework for knowledge graph completion, arXiv preprint arXiv:2411.08165 (2024). [33] M. Gao, Y. Xie, W. Chen, F. Zhang, F. Ding, T. Wang, J. Yao, J. Zheng, K.-F. Wong, Rerankgc: A cooperative retrieve-and-rerank framework for multi-modal knowledge graph completion, Neural Networks 188 (2025) 107467. [34] S. Verma, R. Bhatia, S. Harit, S. Batish, Scholarly knowledge graphs through structuring scholarly communication: a review, Complex & intelligent systems 9 (2023) 1059–1095. [35] A. Hogan, E. Blomqvist, M. Cochez, C. d’Amato, G. D. Melo, C. Gutierrez, S. Kirrane, J. E. L. Gayo,
R. Navigli, S. Neumaier, et al., Knowledge graphs, ACM Computing Surveys (Csur) 54 (2021) 1–37. [36] D. Fensel, U. Şimşek, K. Angele, E. Huaman, E. Kärle, O. Panasiuk, I. Toma, J. Umbrich, A. Wahler, D. Fensel, et al., Introduction: what is a knowledge graph?, Knowledge graphs: Methodology, tools and selected use cases (2020) 1–10. [37] R. Usbeck, A.-C. Ngonga Ngomo, M. Röder, D. Gerber, S. A. Coelho, S. Auer, A. Both, Agdistisgraph-based disambiguation of named entities using linked data, in: The Semantic Web–ISWC 2014: 13th International Semantic Web Conference, Riva del Garda, Italy, October 19-23, 2014.
Proceedings, Part I 13, Springer, 2014, pp. 457–471. [38] H. Chen, Large knowledge model: Perspectives and challenges, arXiv preprint arXiv:2312.02706 (2023).