Entity Resolution and Knowledge Graph Work for Japan

A Japanese brand name has four or five valid written forms: kanji, hiragana, katakana, full-width romaji, half-width romaji. Until something links them, retrieval systems treat them as different things. Of the thirty-odd Japanese consumer brands we checked against Wikidata before starting work, most had no item at all, and two of the ones that did had been merged with a differently-owned company sharing a 商号. That is the failure this engagement addresses: not that a model says something bad about you, but that it does not know which company you are. What we anchor the entity to: Identifiers that exist outside your marketing and can be checked by anyone. The 13-digit corporate number from the National Tax Agency register, which is public, permanent and almost never used by the agencies that should be using it. The EDINET code if you are listed. GS1 Japan prefixes and JAN codes for food and cosmetics SKUs, which is how product-level entities resolve in retail data. For energy clients, the registered retail electricity supplier number from METI's published list, which settles in one line whether you are a supplier or a reseller, a question assistants get wrong about 新電力 constantly. Then the graph work: Organization and Brand markup whose sameAs points at those identifiers rather than at your own social accounts, which prove nothing. A Wikidata item with sourced statements and correct qualifiers, including the name variants, so the katakana form and the romaji form resolve to one thing. Name forms made consistent across every property you control, including the ones marketing forgot: the recruitment site, the IR pages, the regional distributor microsites. Founder and officer entities where the founder is the retrieval hook, which in beauty and in craft food is often the case, and which is the difference between a model knowing a company and knowing a story. What we refuse: We do not edit Wikipedia articles about clients. Paid editing that hides itself gets the article deleted, gets the brand a permanent record on a public noticeboard, and is worse than having no article. We prepare sourced material, disclose the engagement, and post it on the talk page for independent editors to accept or reject. Sometimes they reject it. We do not buy mentions, place paid citations in databases, or create Wikidata items for companies with no independent sources, which is against policy and pointless, since unsourced items get deleted and the deletion is itself a signal. Who this is not for: Companies whose legal name, brand names and product names are still moving. Fix the naming first; entity work done on a name that changes in six months gets done twice and you pay for both. Groups with unresolved trademark conflicts in Japan, where the disambiguation question is legal before it is technical. And brands who want a Knowledge Panel as the deliverable. Panels are Google's decision, they appear when the underlying entity is well formed and well sourced, and a firm promising you one is promising something it cannot cause. What ships: The identifier map, live and cross-referenced. Schema for your properties, written out and ready to deploy, with the reasoning for every sameAs so your developers do not quietly drop half of it. The Wikidata submission and its outcome, whichever way it goes. A name-form standard for your organisation in Japanese and English that procurement and HR can follow as well as marketing. And a re-check three months later, included, because a share of what gets filed is edited by someone else afterwards and you should know which share.