Songwriter lookup: questions and answers

Why can’t I find a songwriter I know?

Usually one of four things, in roughly the order we see them:

• You typed one word. A lookup needs a full name. A surname on its own matches thousands of records, so it is refused rather than answered badly.
• You typed a stage or band name. Registrations carry the name the writer registered under, which is often their legal name (what a stage name can and cannot reach).
• The name is registered differently. A middle name, an initial, a romanisation or a misspelling is enough to miss it. Try the fuller legal name.
• Nobody has registered their songs. This is the common one, and no amount of searching fixes it. Registration data is not a directory of musicians: a person appears in it only once somebody registers a work with their name on it.

My friend writes songs. How does someone get into this data at all?

By having their works registered, which is a separate act from writing or releasing them. A songwriter normally affiliates with a performing rights organisation — that is where an IPI name number comes from — and their works are registered by a publisher or an administrator, which for a self-releasing writer is often the publishing side of their distributor.

Songlinks only reads registration data. We cannot add anyone to it, and being absent from it says nothing about someone as a writer.

Can I search for an artist or band name?

Sometimes. Songlinks searches songwriters, not performers, so a performer who does not write is not in the data at all, and a band is not a person. A stage name of more than one word can still reach its writer, through the songs released under it. A one-word name is not searched.

Why does one songwriter appear as several records?

Because each record is a name as somebody once registered it, with the works registered under it. Spellings, married names, romanisations, initials and more than one IPI name number all split one writer across records, and nothing in the data says the records are the same human being. Putting them back together is the whole job: the five things that split a writer.

How do you handle two songwriters with the same name?

By not treating the name as the person. Songlinks compares records in pairs and asks whether they are the same human being, and the name itself turns out to be the weakest signal in that decision. Most of the work is done by what is around it: shared catalogue, shared co-writers, and identifiers where they exist. Two Dave Smiths who never appear on a song together and never share a collaborator stay two Dave Smiths.

So a lookup often answers with several people under one name, each with their own records and works. Where the evidence genuinely does not settle a pair, Songlinks says it is unsettled rather than guessing: a wrong join is worse than a missed one when money is involved.

What kind of machine learning does this use?

First the thing that surprises people: there is no standard taxonomy to normalise names to. That absence is the problem. The same person is registered under several names and several IPI name numbers, and two genuinely different writers will happily share a name, a genre and a city. So the job is not name normalisation but entity resolution — score every pair of records for “is this the same human?”, then group the records that agree.

The model is a supervised model over those pairwise features, chosen so that a decision can be opened up and explained rather than only trusted. Two things about it were not obvious in advance: a few calls cannot be learned from this data at all and stay as explicit rules, and the system has to be able to say it is not sure instead of answering anyway.

Where do the records come from, and how current are they?

From public music rights-registration data, read as a snapshot. The snapshot does not state the date it was cut, so we do not claim one: a work registered in the last few months may not be in it yet. We do not publish which data it is.

How should I read an answer?

As evidence, and check it against what you already know. Every join comes with the shared works, co-writers and publishers it rests on, and every record Songlinks left apart is shown too, so the answer can be read rather than taken on trust. The case worth a second look is a stage name that reads like a legal name: there, the catalogue behind the answer is what tells a right answer from a near miss.

Songlinks is in beta and the grouping is a model’s judgement that no person has checked, so an answer is a lead to verify in the registrations themselves rather than a rights or royalty decision.

Something about me here is wrong. Can you change it?

We can change how records are grouped, and a wrong grouping is exactly what we want to hear about. We cannot change a registration itself — the name, the works, the numbers — because that belongs to whoever registered the work. Email alex@songlinks.to, or see the privacy notice.

Is it free? Do I need an account?

It is free and there is nothing to sign up for. There are no accounts on the site. Try a lookup. Matching a whole list of writers at once is a separate service we run for you.