LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation

Haffoudhi, Samy; Dobričić, Nikola; Suchanek, Fabian; Holzenberger, Nils

doi:10.24963/ijcai.2026/966

LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation

Samy Haffoudhi, Nikola Dobričić, Fabian Suchanek, Nils Holzenberger

Download BibTeX Cite

Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence. Demo Track. Pages 8401-8405.

https://doi.org/10.24963/ijcai.2026/966

Keywords

AI: Natural Language Processing

Abstract

Entity linking is a key component of many downstream NLP systems, yet existing approaches are often tied to the specific target knowledge bases and domains, limiting their real world application. In this paper, we extend LELA, a modular and domain-agnostic LLM-based entity disambiguation method, into a practical Python library that integrates zero-shot Named Entity Recognition (NER) -- thereby providing a complete end-to-end pipeline for entity-linking in real-world usage. We provide experimental results validating LELA's performance and robustness across diverse entity linking settings. In our demo, users can play with the system on their own input texts. All code is publicly available at https://github.com/dig-team/LELA, and a video is at https://www.youtube.com/watch?v=WdupiRjLbR4.

[Auer et al., 2007] Sören Auer, Christian Bizer, Georgi Ko- bilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. DBpedia: A nucleus for a web of open data. In Pro- ceedings of the 6th International The Semantic Web and 2nd Asian Conference on Asian Semantic Web Conference, ISWC'07/ASWC'07, pages 722-735, Berlin, Heidelberg, November 2007. Springer-Verlag.
[Ayoola et al., 2022] Tom Ayoola, Shubhi Tyagi, Joseph Fisher, Christos Christodoulopoulos, and Andrea Pier- leoni. ReFinED: An Efficient Zero-shot-capable Approach to End-to-End Entity Linking. In Anastassia Loukina, Rashmi Gangadharaiah, and Bonan Min, editors, Proceed- ings of the 2022 Conference of the North American Chap- ter of the Association for Computational Linguistics: Hu- man Language Technologies: Industry Track, pages 209- 220, Hybrid: Seattle, Washington + Online, July 2022. As- sociation for Computational Linguistics.
[Bast et al., 2022] Hannah Bast, Matthias Hertel, and Natalie Prange. ELEVANT: A Fully Automatic Fine-Grained En- tity Linking Evaluation and Analysis Tool. In Wanxiang Che and Ekaterina Shutova, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations, pages 72-79, Abu Dhabi, UAE, December 2022. Association for Com- putational Linguistics.
[Bogdanov et al., 2024] Sergei Bogdanov, Alexandre Con- stantin, Timothée Bernard, Benoit Crabbé, and Etienne P Bernard. NuNER: Entity Recognition Encoder Pre- training via LLM-Annotated Data. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing, pages 11829-11841, Miami, Florida, USA, 2024. Association for Computational Linguistics.
[Cocchieri et al., 2025] Alessio Cocchieri, Marcos Martínez Galindo, Giacomo Frisoni, Gianluca Moro, Claudio Sartori, and Giuseppe Tagliavini. ZeroNER: Fueling Zero-Shot Named Entity Recognition via Entity Type Descriptions. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Findings of the Association for Computational Linguistics: ACL 2025, pages 15594-15616, Vienna, Austria, July 2025. Association for Computational Linguistics.
[Graciotti et al., 2025] Arianna Graciotti, Nicolas Lazzari, Valentina Presutti, and Rocco Tripodi. Musical heritage historical entity linking. Artificial Intelligence Review, 58(5):140, February 2025.
[Haffoudhi et al., 2026] Samy Haffoudhi, Fabian M. Suchanek, and Nils Holzenberger. LELA: an LLM- based Entity Linking Approach with Zero-Shot Domain Adaptation, January 2026. arXiv:2601.05192 [cs].
[Islamaj et al., 2021] Rezarta Islamaj, Chih-Hsuan Wei, David Cissel, Nicholas Miliaras, Olga Printseva, Oleg Rodionov, Keiko Sekiya, Janice Ward, and Zhiyong Lu. NLM-Gene, a richly annotated gold standard dataset for gene entities that addresses ambiguity and multi-species gene recognition. Journal of Biomedical Informatics, 118:103779, June 2021.
[Ji and Grishman, 2011] Heng Ji and Ralph Grishman. Knowledge base population: Successful approaches and challenges. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human lan- guage technologies, pages 1148-1158, 2011.
[Logeswaran et al., 2019a] Lajanugen Logeswaran, Ming- Wei Chang, Kenton Lee, Kristina Toutanova, Jacob De- vlin, and Honglak Lee. Zero-Shot Entity Linking by Read- ing Entity Descriptions, June 2019.
[Logeswaran et al., 2019b] Lajanugen Logeswaran, Ming- Wei Chang, Kenton Lee, Kristina Toutanova, Jacob De- vlin, and Honglak Lee. Zero-Shot Entity Linking by Read- ing Entity Descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3449-3460, Florence, Italy, 2019. Association for Computational Linguistics.
[Martinez-Rodriguez et al., 2020] Jose L. Martinez- Rodriguez, Aidan Hogan, and Ivan Lopez-Arevalo. Information extraction meets the Semantic Web: A survey. Semantic Web, 11(2):255-335, February 2020.
[Orlando et al., 2024] Riccardo Orlando, Pere-Lluís Huguet Cabot, Edoardo Barba, and Roberto Navigli. ReLiK: Retrieve and LinK, Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget.
In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, pages 14114-14132, Bangkok, Thailand, August 2024. Association for Computational Linguistics.
[Picco et al., 2023] Gabriele Picco, Marcos Mar- tinez Galindo, Alberto Purpura, Leopold Fuchs, Vanessa Lopez, and Thanh Lam Hoang. Zshot: An Open-source Framework for Zero-Shot Named Entity Recognition and Relation Extraction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 357-368, Toronto, Canada, 2023. Association for Computational Linguistics.
[Röder et al., 2018] Michael Röder, Ricardo Usbeck, and Axel-Cyrille Ngonga Ngomo. GERBIL -Benchmarking Named Entity Recognition and Linking consistently. Se- mantic Web, 9(5):605-625, August 2018.
[Stepanov and Shtopko, 2024] Ihor Stepanov and Mykhailo Shtopko. GLiNER multi-task: Generalist Lightweight Model for Various Information Extraction Tasks, August 2024.
[Stepanov et al., 2026] Ihor Stepanov, Mykhailo Shtopko, Dmytro Vodianytskyi, and Oleksandr Lukashov. The Million-Label NER: Breaking Scale Barriers with GLiNER bi-encoder, February 2026.
[Suchanek et al., 2024] Fabian M. Suchanek, Mehwish Alam, Thomas Bonald, Lihu Chen, Pierre-Henri Paris, and Jules Soria. YAGO 4.5: A Large and Clean Knowl- edge Base with a Rich Taxonomy. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval, pages 131-140, Washington DC USA, July 2024. ACM.
Denny Vrandečić. Wikidata: A new plat- form for collaborative data collection. In Proceedings of the 21st International Conference on World Wide Web, WWW '12 Companion, pages 1063-1064, New York, NY, USA, April 2012. Association for Computing Machinery.
[Welty et al., 2012] Chris Welty, J. William Murdock, Aditya Kalyanpur, and James Fan. A comparison of hard filters and soft evidence for answer typing in wat- son. In Philippe Cudré-Mauroux, Jeff Heflin, Evren Sirin, Tania Tudorache, Jérôme Euzenat, Manfred Hauswirth, Josiane Xavier Parreira, Jim Hendler, Guus Schreiber, Abraham Bernstein, and Eva Blomqvist, editors, The Se- mantic Web -ISWC 2012 -11th International Seman- tic Web Conference, Boston, MA, USA, November 11-15, 2012, Proceedings, Part II, volume 7650 of Lecture Notes in Computer Science, pages 243-256. Springer, 2012.
[Wu et al., 2020a] Ledell Wu, Fabio Petroni, Martin Josi- foski, Sebastian Riedel, and Luke Zettlemoyer. Scalable Zero-shot Entity Linking with Dense Entity Retrieval. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, Proceedings of the 2020 Conference on Empiri- cal Methods in Natural Language Processing (EMNLP), pages 6397-6407, Online, November 2020. Association for Computational Linguistics.
[Wu et al., 2020b] Ledell Wu, Fabio Petroni, Martin Josi- foski, Sebastian Riedel, and Luke Zettlemoyer. Scal- able Zero-shot Entity Linking with Dense Entity Retrieval, September 2020. arXiv:1911.03814 [cs].
[Yang et al., 2025] An Yang, Anfeng Li, Baosong Yang, Be- ichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayi- heng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Hao- ran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jin- gren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 Technical Report, May 2025. arXiv:2505.09388 [cs].
[Zaratiana et al., 2024] Urchade Zaratiana, Nadi Tomeh, Pierre Holat, and Thierry Charnois. GLiNER: General- ist Model for Named Entity Recognition using Bidirec- tional Transformer. In Proceedings of the 2024 Confer- ence of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), pages 5364-5376, Mexico City, Mexico, 2024. Association for Computational Lin- guistics. Zhang et al., 2025] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 Embedding: Advancing Text Em- bedding and Reranking Through Foundation Models, June 2025. arXiv:2506.05176 [cs].

References were extracted automatically from the PDF and may contain errors.

Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.

1 Introduction

Entity linking (EL) is the task of identifying and mapping ambiguous mentions of entities in a natural language text to reference entities in a knowledge base (KB). For example, the text can be:

France hosted the Olympics in Paris.

The KB contains each entity, associated with a short textual description, e.g.:

Paris (city) Capital city of France Paris (novel) Novel by Emile Zola France Country in Europe France Gall French singer

Entity linking consists of two sub-tasks, mention detection (MD) and entity disambiguation (ED). Mention detection aims to identify the mentions of entities (here: "France", "the Olympics", and "Paris"). Mention detection often boils down to Named Entity Recognition (NER). Entity disambiguation, on the other hand, aims to map the identified mentions to their correct entity in the KB, if they exist (here: "France" to France and "Paris" to

from lela import Lela # Choose each component of LELA config = { "loader": { "name": "text" # or: pdf, docx, ... }, "ner": { "name": "gliner", # or: regex, spacy "params": {"labels": ["person", "location"]}, }, "candidate_generator": {"name": "bm25"}, # or: fuzzy, dense, openai_api_dense "reranker": {"name": "llama_server"}, # or: none, cross_encoder_vllm... "disambiguator": { "name": "vllm", # or: first, openai_api, transformers "params": {"model_name": "Qwen/Qwen3-4B"}, }, "knowledge_base": { "name": "jsonl", "params": {"path": "my_kb.jsonl"} }, } lela = Lela(config) # Run the pipeline on a document results = lela.run("docs/file1.txt")

Paris (city), with no mapping for "Olympics"). Entity Linking is an important preprocessing step in tasks such as information extraction [Martinez-Rodriguez et al., 2020], knowledge-based question answering [Welty et al., 2012], and knowledge graph completion [Ji and Grishman, 2011].

Most entity linking approaches focus on linking to general KBs such as Wikidata [Vrandeˇci´c, 2012], DBpedia [Auer et al., 2007] or Yago [Suchanek et al., 2024]. However, in real-world applications, the KB is often proprietary or domain-specific, as in the legal or biomedical domain, or inside a company. Entity disambiguation for such KBs has been addressed recently by the LELA system [Haffoudhi et al., 2026], which can disambiguate entities in a true zeroshot fashion without any need for training data or fine-tuning. However, LELA performs only entity disambiguation, not mention detection, i.e., it assumes that entity mentions have already been identified in the text. In real-world scenarios, however, the mentions of the entities have to be detected.

In this demo paper, we extend LELA to an end-to-end EL system that can ingest text and a KB and output a disambiguation - with no need for either training-data, fine-tuning, or the prior identification of the mentions. We contribute a modular and fully open framework, where different named-entity extractors, candidate retrievers, rerankers, and disambiguators can be used interchangeably. Our demo lets users run LELA on their own input texts with different KBs.

2 RelatedWork

Current research typically bifurcates the problem into two distinct tasks: zero-shot NER, which identifies mentions of unseen entity types [Zaratiana et al., 2024; Bogdanov et al., 2024; Cocchieri et al., 2025], and zero-shot ED, which maps mentions to a KB not encountered during training [Logeswaran et al., 2019a; Wu et al., 2020a; Haffoudhi et al., 2026]. To our knowledge, a unified, end-to-end architecture for zero-shot EL remains unexplored. Consequently, existing frameworks such as spaCy 1 , Zshot [Picco et al., 2023], and GLINKER 2 rely on modular, multi-stage pipelines. However, these are often tightly coupled to specific architectures. LELA differentiates itself by offering a flexible LLM-native architecture and integrating the state-of-the-art MD and ED components, leading to unparalleled robustness in real-world scenarios. Our experiments confirm that the capabilities of LLMs are essential for handling the ambiguity of true zeroshot ED, allowing LELA to significantly outperform existing frameworks on complex, unseen domains.

3 SystemDesign

Our work extends LELA from a disambiguation-only method to a complete end-to-end entity linking framework. Our system is built as a modular pipeline on top of spaCy's component architecture, where each stage is a pluggable component that can be changed independently. The pipeline works in four stages:

Named Entity Recognition. The first stage identifies entity mentions in the input text. We provide several interchangeable NER components: a zero-shot GLiNER model [Zaratiana et al., 2024] (default: NuNER Zerospan [Bogdanov et al., 2024]), spaCy's pretrained NER models, and a lightweight regex-based recognizer for rapid prototyping or well-defined mention formats. Long documents are automatically chunked with overlap to respect model context limits.
Candidate Generation. For each detected mention, the system retrieves candidates from the KB. Available retrievers include: BM25 (with rank-bm25 3 ), dense retrieval with FAISS 4 indexing, and fuzzy string matching via RapidFuzz 5 .

1 https://spacy.io/api/large-language-models#el-v1

2 a recently-proposed production-oriented EL framework that is built around GLiNER [Stepanov and Shtopko, 2024; Stepanov et al., 2026]: https://github.com/Knowledgator/GLinker?tab= readme-ov-file

3 https://github.com/dorianbrown/rank bm25/tree/master

4 https://github.com/facebookresearch/faiss

5 https://github.com/rapidfuzz/RapidFuzz

Reranking (optional). Candidates can be reranked to filter the set passed to the disambiguator using cross-encoder models with top-k cutoff. This stage can also be skipped. See [Haffoudhi et al., 2026] for a full study of the use of reranking for entity disambiguation.

4. Disambiguation. The final stage selects the correct entity from the candidates using LLM reasoning, following the LELA methodology [Haffoudhi et al., 2026]. The LLM receives the top-k candidates from the reranker along with the input context (with the mention marked in square brackets) and selects the most likely entity by index. LLM-based disambiguation supports linking a detected mention to "NIL" via an explicit "None of the candidates" option. A baseline disambiguator that selects by retrieval rank is also provided. The entity is then displayed when hovering over the highlighted span.

Backends. The components 2-4 can also be disabled for use cases that require only mention detection. For a given component, the framework provides different backend options: vLLM 6 for fast LLM inference using the offline API, the Transformers 7 library, SentenceTransformers 8 for the embedding and reranking models, as well as HTTP requests to OpenAI API compatible endpoints 9 , for interacting with models served using vLLM (online API), llama.cpp 10 (for quantized models, potentially running on CPUs), or proprietary models hosted remotely.

Interfaces. The KB is given as a simple JSONL file, with each line representing an entity with an identifier, a label, and a description. Our system can be used via three interfaces:

a Python API for programmatic use (as in Figure 1),
a command-line interface (CLI) (lela --config config.json --input doc.txt) for batch processing of documents in various formats (text, PDF, DOCX, HTML, JSON, JSONL), and
an interactive Web application built with Gradio 11 (described in Section 5).

All three interfaces share the same underlying spaCy pipeline and configuration system. The system supports caching, progress tracking, and GPU memory estimation to help users select appropriate model configurations.

Extensibility. Newcomponents can be added by registering a spaCy factory (for pipeline stages) or a registry entry (for loaders and knowledge bases). Each component follows a simple protocol: (1) NER components populate doc.ents, (2) candidate generators populate a custom candidates extension, and (3) disambiguators set a resolved entity extension - making it straightforward to integrate domainspecific models without modifying the core framework.

6 https://github.com/vllm-project/vllm

7 https://github.com/huggingface/transformers

8 https://github.com/huggingface/sentence-transformers

9 https://developers.openai.com/api/reference/overview

10 https://github.com/ggml-org/llama.cpp

11 https://www.gradio.app/

4 Experimental Validation

Datasets. We evaluate our framework on two different benchmark datasets: Elgold [Islamaj et al., 2021], which evaluates entity linking to Wikipedia across seven domains; and MHERCL [Graciotti et al., 2025] which focuses on Wikidata long-tail entities and the musical heritage domain.

LELA Configuration. We report results for LELA using the Qwen3-Embedding-4B retriever, the Qwen3-Reranker-4B reranker [Zhang et al., 2025] and the Qwen3-30B-A3B reasoning LLM [Yang et al., 2025]. We retrieve 100 candidates per mention, set the top-k cutoff to 10 candidates, and sample three outputs for self-consistency. We use the NuNER Zero-span model [Bogdanov et al., 2024] for mention-detection with a fixed NER label set, based on the set of mention types present in the benchmark.

Baselines. We evaluate BLINK 12 [Wu et al., 2020b], as another representative of zero-shot entity disambiguation methods, and the two state-of-the-art end-to-end entity linking methods for Wikipedia: ReFinED [Ayoola et al., 2022] and Relik [Orlando et al., 2024]. On MHERCL, we also compare to the recently-proposed GLiNKER framework.

Evaluation Metrics. We measure performance using ELEVANT [Bast et al., 2022]. On Elgold, we report InKB EL F1 score [R¨oder et al., 2018]. On MHERCL, we follow prior work [Graciotti et al., 2025] and report EL F1.

Results. On the Elgold benchmark (Table 1), LELA performs on par with the fully supervised state-of-the-art models, ReFinED and Relik. LELA achieves this competitiveness without the required extensive task-specific training on Wikipedia. This advantage is most visible in the complex domains such as the Science paper abstracts, where LELA surpasses ReFinED by nearly 18 percentage points, demonstrating superior robustness. Furthermore, LELA systematically outperforms BLINK, when both rely on the same MD component, confirming the performance gains induced by the reasoning capabilities of our LLM-based ED compared to standard encoder-only baselines.

On the MHERCL benchmark (Table 2), which targets long-tail entities in the musical domain, LELA establishes a new state-of-the-art. It outperforms ReFinED and significantly surpasses other zero-shot or modular baselines. These results confirm that while supervised methods struggle to adapt to specialized domains without retraining, LELA handles arbitrary distributions robustly.

5 Demo

Our demo allows users to link entities in arbitrary documents with different KBs using LELA's Web application. Our interface shows the different steps of the disambiguation process and measures the execution time. Finally, it shows the recognized mentions with their link to the entities in the KB, or to NIL. The Web interface mirrors the full modularity of our approach: Users can experiment with different NER systems, candidate retrievers, rerankers, and reasoning LLMs trading off speed against performance. We provide a large range of reference KBs to try out: Users can work with both general-purpose KBs (such as YAGO [Suchanek et al., 2024]) and domain-specific KBs (such as the Crossref Funder Registry 13 or the 16 different Wikia domains from ZESHEL [Logeswaran et al., 2019b]).

12 Coupled with NuNER Zero-span for mention detection.

Table 1: F1 values (InKB, in %) on Elgold across the seven domains, with 95% confidence intervals. Gray results have overlapping confidence intervals with the best results. Unlike competing methods, LELA was not fine-tuned for entity linking on Wikipedia.
Domain	BLINK	Relik	ReFinED	LELA
1 (News) 2 (Jobs) 3 (Movie) 4 (Auto) 5 6 (Science) 8 (Historic)	68.2±3.1 43.3±4.2 69.2±4.8 63.7±5.4 63.4±4.7 29.3±3.3	76.5±2.9 67.0±5.1 72.5±5.0 65.8±6.0 67.8±4.9 31.5±4.0	78.4±2.8 72.0±4.6 74.3±4.8 74.9±5.1 66.9±4.9 23.8±3.8 72.2±6.9	74.7±3.0 60.2±4.7 75.3±4.7 66.5±5.6 71.5±4.6 41.6±3.8
Macro	57.6±1.8
(Amazon)
	65.7±7.0	68.2±7.4		69.9±7.0
		64.2±2.0	66.1±1.8	65.7±1.8

Table 2: F1 values (in %) on the MHERCL benchmark, with 95% confidence intervals.
Method	EL F1
GLiNKER [Stepanov and Shtopko, 2024]	16±1.5
BLINK [Wu et al., 2020a]	47±2.0
Relik [Orlando et al., 2024]	44±2.0
ReFinED [Ayoola et al., 2022]	49±2.0
LELA (ours)	56±2.0

In a twist that goes beyond existing demos, users can come up with their own KBs. For example, a user interested in biology can create a small dictionary with two meanings of the word "culture" (the process of growing cells in the lab vs. the ensemble of arts, customs, and traditions), and ask the system to identify mentions to these entities and determine which meaning is intended in a sentence. Similarly, users from industry can create a dictionary of technical terms that their own approaches struggle with (such as terms that have specific meanings in industry jargon), and see if our approach copes, experimenting with different pipeline configurations.

6 Conclusion

LELA is a Python library for entity linking, handling both mention detection and entity disambiguation. It is true zeroshot, in the sense that it needs neither training data nor finetuning, and works on domain-specific text and knowledge bases out-of-the-box. LELA can be used through a Python API, a CLI or a Web Interface, making it suited for realworld applications. It is modular, offering different options for the various stages of the entity linking pipeline, and can be extended with additional components. Future work can integrate additional components, add support for KBs in Turtle format, as well as automate the generation of textual descriptions and NER labels.

13 https://www.crossref.org/services/funder-registry/

Acknowledgements

The work was partially supported by Agence de l'Innovation de D´efense - AID - via Centre Interdisciplinaire d'Etudes pour la D´efense et la S´ecurit´e - CIEDS - (project 2024 - KBLM).

Samy Haffoudhi, Nikola Dobričić, Fabian Suchanek, and Nils Holzenberger. 2026. LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation. In Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, IJCAI-26. IJCAI, 8401–8405. https://doi.org/10.24963/ijcai.2026/966

@inproceedings{ijcai2026p966,
  title     = {LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation},
  author    = {Haffoudhi, Samy and Dobričić, Nikola and Suchanek, Fabian and Holzenberger, Nils},
  booktitle = {Proceedings of the Thirty-Fifth International Joint Conference on
               Artificial Intelligence, {IJCAI-26}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Diego Calvanese},
  pages     = {8401--8405},
  year      = {2026},
  month     = {8},
  note      = {Demo Track},
  doi       = {10.24963/ijcai.2026/966},
  url       = {https://doi.org/10.24963/ijcai.2026/966},
}

TY  - CPAPER
TI  - LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation
AU  - Haffoudhi, Samy
AU  - Dobričić, Nikola
AU  - Suchanek, Fabian
AU  - Holzenberger, Nils
T2  - Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, IJCAI-26
PB  - International Joint Conferences on Artificial Intelligence Organization
PY  - 2026
SP  - 8401
EP  - 8405
DO  - 10.24963/ijcai.2026/966
UR  - https://doi.org/10.24963/ijcai.2026/966
ER  -

[1] [Auer et al., 2007] Sören Auer, Christian Bizer, Georgi Ko- bilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. DBpedia: A nucleus for a web of open data. In Pro- ceedings of the 6th International The Semantic Web and 2nd Asian Conference on Asian Semantic Web Conference, ISWC'07/ASWC'07, pages 722-735, Berlin, Heidelberg, November 2007. Springer-Verlag.

[2] [Ayoola et al., 2022] Tom Ayoola, Shubhi Tyagi, Joseph Fisher, Christos Christodoulopoulos, and Andrea Pier- leoni. ReFinED: An Efficient Zero-shot-capable Approach to End-to-End Entity Linking. In Anastassia Loukina, Rashmi Gangadharaiah, and Bonan Min, editors, Proceed- ings of the 2022 Conference of the North American Chap- ter of the Association for Computational Linguistics: Hu- man Language Technologies: Industry Track, pages 209- 220, Hybrid: Seattle, Washington + Online, July 2022. As- sociation for Computational Linguistics.

[3] [Bast et al., 2022] Hannah Bast, Matthias Hertel, and Natalie Prange. ELEVANT: A Fully Automatic Fine-Grained En- tity Linking Evaluation and Analysis Tool. In Wanxiang Che and Ekaterina Shutova, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations, pages 72-79, Abu Dhabi, UAE, December 2022. Association for Com- putational Linguistics.

[4] [Bogdanov et al., 2024] Sergei Bogdanov, Alexandre Con- stantin, Timothée Bernard, Benoit Crabbé, and Etienne P Bernard. NuNER: Entity Recognition Encoder Pre- training via LLM-Annotated Data. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan- guage Processing, pages 11829-11841, Miami, Florida, USA, 2024. Association for Computational Linguistics.

[5] [Cocchieri et al., 2025] Alessio Cocchieri, Marcos Martínez Galindo, Giacomo Frisoni, Gianluca Moro, Claudio Sartori, and Giuseppe Tagliavini. ZeroNER: Fueling Zero-Shot Named Entity Recognition via Entity Type Descriptions. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Findings of the Association for Computational Linguistics: ACL 2025, pages 15594-15616, Vienna, Austria, July 2025. Association for Computational Linguistics.

[6] [Graciotti et al., 2025] Arianna Graciotti, Nicolas Lazzari, Valentina Presutti, and Rocco Tripodi. Musical heritage historical entity linking. Artificial Intelligence Review, 58(5):140, February 2025.

[7] [Haffoudhi et al., 2026] Samy Haffoudhi, Fabian M. Suchanek, and Nils Holzenberger. LELA: an LLM- based Entity Linking Approach with Zero-Shot Domain Adaptation, January 2026. arXiv:2601.05192 [cs].

[8] [Islamaj et al., 2021] Rezarta Islamaj, Chih-Hsuan Wei, David Cissel, Nicholas Miliaras, Olga Printseva, Oleg Rodionov, Keiko Sekiya, Janice Ward, and Zhiyong Lu. NLM-Gene, a richly annotated gold standard dataset for gene entities that addresses ambiguity and multi-species gene recognition. Journal of Biomedical Informatics, 118:103779, June 2021.

[9] [Ji and Grishman, 2011] Heng Ji and Ralph Grishman. Knowledge base population: Successful approaches and challenges. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human lan- guage technologies, pages 1148-1158, 2011.

[10] [Logeswaran et al., 2019a] Lajanugen Logeswaran, Ming- Wei Chang, Kenton Lee, Kristina Toutanova, Jacob De- vlin, and Honglak Lee. Zero-Shot Entity Linking by Read- ing Entity Descriptions, June 2019.

[11] [Logeswaran et al., 2019b] Lajanugen Logeswaran, Ming- Wei Chang, Kenton Lee, Kristina Toutanova, Jacob De- vlin, and Honglak Lee. Zero-Shot Entity Linking by Read- ing Entity Descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3449-3460, Florence, Italy, 2019. Association for Computational Linguistics.

[12] [Martinez-Rodriguez et al., 2020] Jose L. Martinez- Rodriguez, Aidan Hogan, and Ivan Lopez-Arevalo. Information extraction meets the Semantic Web: A survey. Semantic Web, 11(2):255-335, February 2020.

[13] [Orlando et al., 2024] Riccardo Orlando, Pere-Lluís Huguet Cabot, Edoardo Barba, and Roberto Navigli. ReLiK: Retrieve and LinK, Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget.

[14] In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, pages 14114-14132, Bangkok, Thailand, August 2024. Association for Computational Linguistics.

[15] [Picco et al., 2023] Gabriele Picco, Marcos Mar- tinez Galindo, Alberto Purpura, Leopold Fuchs, Vanessa Lopez, and Thanh Lam Hoang. Zshot: An Open-source Framework for Zero-Shot Named Entity Recognition and Relation Extraction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 357-368, Toronto, Canada, 2023. Association for Computational Linguistics.

[16] [Röder et al., 2018] Michael Röder, Ricardo Usbeck, and Axel-Cyrille Ngonga Ngomo. GERBIL -Benchmarking Named Entity Recognition and Linking consistently. Se- mantic Web, 9(5):605-625, August 2018.

[17] [Stepanov and Shtopko, 2024] Ihor Stepanov and Mykhailo Shtopko. GLiNER multi-task: Generalist Lightweight Model for Various Information Extraction Tasks, August 2024.

[18] [Stepanov et al., 2026] Ihor Stepanov, Mykhailo Shtopko, Dmytro Vodianytskyi, and Oleksandr Lukashov. The Million-Label NER: Breaking Scale Barriers with GLiNER bi-encoder, February 2026.

[19] [Suchanek et al., 2024] Fabian M. Suchanek, Mehwish Alam, Thomas Bonald, Lihu Chen, Pierre-Henri Paris, and Jules Soria. YAGO 4.5: A Large and Clean Knowl- edge Base with a Rich Taxonomy. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval, pages 131-140, Washington DC USA, July 2024. ACM.

[20] Denny Vrandečić. Wikidata: A new plat- form for collaborative data collection. In Proceedings of the 21st International Conference on World Wide Web, WWW '12 Companion, pages 1063-1064, New York, NY, USA, April 2012. Association for Computing Machinery.

[21] [Welty et al., 2012] Chris Welty, J. William Murdock, Aditya Kalyanpur, and James Fan. A comparison of hard filters and soft evidence for answer typing in wat- son. In Philippe Cudré-Mauroux, Jeff Heflin, Evren Sirin, Tania Tudorache, Jérôme Euzenat, Manfred Hauswirth, Josiane Xavier Parreira, Jim Hendler, Guus Schreiber, Abraham Bernstein, and Eva Blomqvist, editors, The Se- mantic Web -ISWC 2012 -11th International Seman- tic Web Conference, Boston, MA, USA, November 11-15, 2012, Proceedings, Part II, volume 7650 of Lecture Notes in Computer Science, pages 243-256. Springer, 2012.

[22] [Wu et al., 2020a] Ledell Wu, Fabio Petroni, Martin Josi- foski, Sebastian Riedel, and Luke Zettlemoyer. Scalable Zero-shot Entity Linking with Dense Entity Retrieval. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, Proceedings of the 2020 Conference on Empiri- cal Methods in Natural Language Processing (EMNLP), pages 6397-6407, Online, November 2020. Association for Computational Linguistics.

[23] [Wu et al., 2020b] Ledell Wu, Fabio Petroni, Martin Josi- foski, Sebastian Riedel, and Luke Zettlemoyer. Scal- able Zero-shot Entity Linking with Dense Entity Retrieval, September 2020. arXiv:1911.03814 [cs].

[24] [Yang et al., 2025] An Yang, Anfeng Li, Baosong Yang, Be- ichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayi- heng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Hao- ran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jin- gren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 Technical Report, May 2025. arXiv:2505.09388 [cs].

[25] [Zaratiana et al., 2024] Urchade Zaratiana, Nadi Tomeh, Pierre Holat, and Thierry Charnois. GLiNER: General- ist Model for Named Entity Recognition using Bidirec- tional Transformer. In Proceedings of the 2024 Confer- ence of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), pages 5364-5376, Mexico City, Mexico, 2024. Association for Computational Lin- guistics. Zhang et al., 2025] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 Embedding: Advancing Text Em- bedding and Reranking Through Foundation Models, June 2025. arXiv:2506.05176 [cs].