yakdas/turklink-corpus

TurkLink is a comprehensive, large-scale Entity Linking corpus specifically designed for the Turkish language. It bridges the gap between unstructured Turkish text and structured knowledge bases by aligning Turkish Wikipedia articles with Wikidata entities. Furthermore, the dataset features extensive linguistic enrichments generated by the Turkish.AI NLP Tool, making it a highly valuable resource for advanced Natural Language Processing (NLP) tasks in Turkish.

Dataset Details

• Language: Turkish (tr)
• Primary Task: Entity Linking (EL)
• Secondary Tasks: Named Entity Recognition (NER), Part-of-Speech (POS) Tagging, Morphological Analysis, Dependency Parsing

Data Splits

The dataset is divided into training, validation, and test sets to facilitate standardized model evaluation.

Table 1. The data distribution of TurkLink corpus

Set	Number of samples	Percentage (%)
Train	472,074	80
Validation	59,009	10
Test	59,009	10
Total	590,092	100

Dataset Statistics

Below are the detailed statistics representing the sheer scale and richness of the Wikipedia portion of the corpus.

Table 2. Statistics of the Wikipedia part of TurkLink corpus

Statistic	Value
Article Count	590,092
Sentence Count	6,361,587
Token Count	96,647,654
Unique Entities (Q-codes)	258,773
Total Entity Mentions	4,367,587
Total NER Tags	15,186,429
Mention-to-Entity Ratio	16.88

Source Data

The corpus is constructed from robust, high-quality knowledge bases:

• Wikipedia: Based on the Turkish Wikipedia dump from April 2024.
• Wikidata: Based on the Wikidata dump from October 2025.

Dataset Volume and Scale

TurkLink is built to support large-scale machine learning applications and contains:

• > 500,000 Turkish Wikipedia articles.
• > 3,000,000 Wikidata entities that feature Turkish labels.

Linguistic Annotations

A standout feature of TurkLink is its deep linguistic enrichment. Every article within the corpus has been comprehensively analyzed using the Turkish.AI NLP Tool. The annotations provided for the text include:

• Named Entity Recognition (NER): Beyond standard broad categories, the corpus features highly granular NER annotations. Every article is deeply tagged using a comprehensive taxonomy of over 60 fine-grained entity types, allowing for highly precise identification and classification of entities across varied domains.
• Morphological Analysis: Because Turkish is a highly agglutinative language where a single word can convey the meaning of an entire English sentence, deep morphological breakdown is essential. Every single token in the dataset is extensively analyzed to extract its exact root (lemma) and its sequence of inflectional and derivational affixes, strictly adhering to the widely recognized Oflazer standard [1].
• Part-of-Speech (POS) Tagging: Working in tandem with the morphological analysis, each token is assigned a precise grammatical tag based on its specific syntactic role within the sentence context. Aligned with the Oflazer standard [1], this ensures accurate grammatical representation and helps disambiguate words that share roots but serve different functions.
• Dependency Parse Trees: To capture the exact syntactic architecture of the Turkish text, a complete structural analysis is performed. A full dependency parse tree—strictly compliant with Universal Dependencies v2 (UD v2) standards [2]—is provided for every single sentence, mapping out the precise grammatical relationships and hierarchical dependencies between words.

Knowledge Base & Entity Features

To maximize its utility for Entity Linking and Knowledge Graph tasks, TurkLink includes rich, structured metadata seamlessly integrated from Wikidata:

• Entity-Mention Prior Probabilities: The dataset includes pre-calculated prior probabilities for entity-mentions, mapping the statistical likelihood of a specific text mention resolving to a particular Wikidata entity within the corpus.
• Wikidata Triples & Properties: The corpus is enriched with structured knowledge representation, providing Wikidata triples and relational properties for the linked entities.
• Entity Attributes: Every Wikidata entity referenced in the dataset is accompanied by its corresponding Turkish label, detailed description, and sitelink count which serves as a valuable proxy for entity prominence/popularity.
• Machine-Translated Descriptions: To ensure maximum coverage and handle missing data, Wikidata entities lacking a native Turkish description were augmented using the Meta NLLB-200 [3] model, which translated their English descriptions into Turkish. For transparency and data integrity, all entities with machine-translated descriptions are explicitly flagged within the dataset.

Citation

Please consider citing the following if you use this dataset in your work.

@article{AKDAS2026333,
    title = {TurkLink: A Morphologically-Aware and Syntactically Enriched Corpus for Entity Linking in Turkish},
    journal = {Procedia Computer Science},
    volume = {275},
    pages = {333-342},
    year = {2026},
    note = {7th International Conference on AI in Computational Linguistics},
    issn = {1877-0509},
    doi = {https://doi.org/10.1016/j.procs.2026.01.041},
    url = {https://www.sciencedirect.com/science/article/pii/S1877050926000414},
    author = {Yusuf Akdaş and Ahmet Cüneyd Tantuğ},
    keywords = {Natural Language Processing, Corpus, Turkish NLP, Entity Linking, Named Entity Recognition, Morphological Analysis, Part-of-Speech Tagging, Dependency Parsing},
}

References

• [1] KEMAL OFLAZER, Two-level Description of Turkish Morphology, Literary and Linguistic Computing, Volume 9, Issue 2, 1994, Pages 137–148, https://doi.org/10.1093/llc/9.2.137
• [2] Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajič, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034–4043, Marseille, France. European Language Resources Association.
• [3] team, N., Costa-jussà, M.R., Cross, J., cCelebi, O., Elbayad, M., Heafield, K., Heffernan, K., Kalbassi, E., Lam, J., Licht, D., Maillard, J., Sun, A., Wang, S., Wenzek, G., Youngblood, A., Akula, B., Barrault, L., Gonzalez, G.M., Hansanti, P., Hoffman, J., Jarrett, S., Sadagopan, K., Rowe, D., Spruit, S.L., Tran, C., Andrews, P.Y., Ayan, N.F., Bhosale, S., Edunov, S., Fan, A., Gao, C., Goswami, V., Guzmán, F.(., Koehn, P., Mourachko, A., Ropers, C., Saleem, S., Schwenk, H., & Wang, J. (2022). No Language Left Behind: Scaling Human-Centered Machine Translation. ArXiv, abs/2207.04672.