Short answer: DBP15K is the closest standard starting point for cross-lingual entity alignment, but it contains Chinese–English and Japanese–English pairs—not Korean data or labeled matches between corporate records. The resources identified here do not establish a ready-made Korean–Japanese–Chinese corporate-name matching dataset. For that task, treat these benchmarks as references or baselines and build gold labels from the company records your system will actually resolve.
First, define what “entity resolution” means for your project
Dataset labels are useful only when they represent the decision your model needs to make. Corporate-name resolution usually asks whether two records refer to the same entity. That requires an explicit policy for cases such as subsidiaries, parent companies, joint ventures, aliases, ownership changes, and legal-entity changes.
As an Amazon Associate I earn from qualifying purchases.
That is different from several related tasks:
- Knowledge-graph entity alignment: deciding which nodes in two knowledge graphs correspond.
- Entity linking: mapping a mention in text to a node in a knowledge base.
- Named-entity recognition (NER): identifying and classifying spans of text.
- Entity classification: assigning a category to an entity.
Alignment, linking, and NER data can help with candidate generation, name normalization, or category recognition. Their labels are not automatically evidence that two corporate records identify the same legal or operating entity.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Best starting point: DBP15K, with important limits
DBP15K is a cross-lingual knowledge-graph alignment benchmark with Chinese–English, Japanese–English, and French–English subsets. The cited IJCAI 2019 paper reports 15,000 reference alignment links per subset. It also reports 66,469 Chinese-side entities and 65,744 Japanese-side entities; those are graph entity counts, not matched company-record counts. See the original DBP15K paper and dataset description.
#1 Best Overall
It is a useful baseline for general alignment experiments, but it does not supply Korean coverage, direct Korean–Japanese or Korean–Chinese gold pairs, or demonstrated corporate-registry labels. English can serve as a bridge for separate Chinese–English and Japanese–English experiments; that does not make the resulting links direct Korean–Japanese or Korean–Chinese labels.
EntMatcher’s DBP15K splits
The EntMatcher repository provides an implementation-oriented organization of DBP15K, including zh_en and ja_en folders and files for support, validation, reference links, and graph triples. It describes a split of 20% training, 10% validation, and 70% test for its listed gold alignment links. Check the specific release’s provenance and split before using it, particularly if comparing results or adapting it to a new task. EntMatcher repository.
Other useful datasets—and the different labels they provide
| Resource | Labels and language coverage | Fit for company-record resolution | Access and caveat |
|---|---|---|---|
| Hansel | Chinese entity-linking test data with 10,000 examples, using Wikidata as the knowledge base. Training and validation examples come from Wikipedia hyperlinks. | Useful for Chinese mention-to-knowledge-base linking, including tail and emerging entities; it does not label pairs of company records. | The repository states CC BY-SA for Hansel. Check terms for component data and the current release. Hansel project. |
| SHINRA2021-ML / SHINRA2020-ML | Japanese Wikipedia pages annotated with Extended Named Entity categories, language links, and target-language Wikipedia pages; the project also provides classification data and scripts. | Can support multilingual entity-category classification or help create language-linked examples; it does not establish whether two company records identify the same entity. | Files are distributed in multiple formats and sizes. Review project terms and data notices. SHINRA project. |
| Mewsli-9 | 289,087 linked entity mentions from 58,717 originally written WikiNews articles across nine languages, linked to Wikidata. Japanese is included; Korean and Chinese are not in the listed language set. | Useful for multilingual entity-linking evaluation and domain-shift study, not corporate-record pair matching. | The paper uses a WikiNews snapshot dated 2019-01-01. Its links are automatically extracted, trading annotation quality for scale and diversity. Mewsli-9 paper. |
| TAC KBP Chinese Cross-lingual Entity Linking 2011–2014 | An LDC collection with English and Chinese documents, queries, entity types, knowledge-base links, and NIL equivalence clusters. | Potentially useful for Chinese/English entity-linking training or evaluation; it does not provide Korean/Japanese coverage or corporate-name record linkage. | The LDC catalog lists a release date of November 17, 2017, and an LDC user agreement for non-members. LDC catalog entry. |
| MELD | A standardized collection of NER datasets across languages and domains, with annotations depending on each source dataset. | Can help with mention detection and entity-type recognition, but NER labels do not identify same-entity record pairs. | Licensing is source-specific; some datasets must be fetched from their original sources. MELD documentation. |
| KORE 50DYWC | An entity-linking evaluation set expanded to DBpedia, YAGO, Wikidata, and Crunchbase. | Relevant to entity linking across knowledge-base targets, not a three-language corporate-name pair corpus. | See the LREC 2020 paper and verify release terms and label compatibility before reuse. KORE 50DYWC paper. |
How to choose data for a corporate-name model
- Write an identity policy. Decide whether the target is the same legal entity, the same operating business, or a broader corporate group. Define how to label subsidiaries, joint ventures, aliases, and changes over time.
- Start with records from the systems you will use. Create or obtain adjudicated match and non-match pairs from the Korean, Japanese, and Chinese registries or source systems in scope. Include difficult negatives, such as different companies with similar names.
- Use public links as candidate signals, not automatic gold labels. Wikipedia or Wikidata links, language links, transliteration, and name similarity may help propose candidates. Preserve each candidate’s origin and confidence, then manually audit a sample before treating the links as training truth.
- Keep task labels separate. Do not merge graph-alignment, entity-linking, NER, and corporate pair-match labels into one target unless you have a documented method for converting them.
- Check splits, leakage, and reuse terms. Confirm how a resource partitions train, validation, and test data; check whether related entities can leak across splits; and review licensing or access agreements for the exact release and intended use.
What is established—and what is not
The resources described above offer useful ingredients for general cross-lingual alignment, Chinese or Japanese entity linking, entity classification, and multilingual NER. They do not demonstrate a labeled corpus of Korean–Japanese–Chinese corporate-name record matches. That is a limit of the identified resources, not proof that no specialized dataset exists in a particular industry or jurisdiction.
A title-matched article by Tae Kim, dated September 23, 2026, reports that the author could not find a public labeled dataset for this task and manually reviewed about 2,000 pairs. That is the author’s account, not a published benchmark statistic or a systematic census of every repository. Tae Kim’s article.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
