My office has bike storage and a locker room for people to shower and change. Closing the locker door, my bag was sticking out, and as I squooshed it in, I heard the ting-ting-ting of something metal bouncing on the concrete floor.
I found a small nameplate reading Hollman, and looking closer, I saw one of the hinges of my locker had an identical name plate while the other was without a cover and showed it mounting hardware. I never noticed there were name plates.
Actually, I never even noticed there were hinges. Hollman considers itself to be “the world’s largest and premier manufacturer of team sports, fitness, and workspace lockers. Hollman lockers are built to inspire teamwork, collaboration, innovation and trust among those who use them.” I don’t design offices. A locker’s a locker. Shrug.
I usually do find it interesting when I learn about the semantic worlds of things that are hiding in plain sight though. In speech tech these days we call people the people who label, interpret and manipulate source language data “ontologists.” Maybe this is a thing about computer science and I’m biased because my focus has always been language. Either way, “ontologists” is a fascinating name to me because it’s impossible to separate it from a definition from metaphysics.
The nature of being. Isness. How do we know something is, and if we do, how can we describe it so that it’s unambiguous? It’s fun to think about.
Sometimes people get annoyed with me because I’m wont to delve into a philosophical tangent, in the spirt of truly and conclusively solving a problem by fully understanding it and agreeing on the understanding. And here is the interesting problem is ontologists: They usually decide that what is under consideration is a single thing. Whether it’s a symptom of the task definition, the toolkit, the environment, the schema, membership in multiple groups is rare. So is consensus. Judgement of a token, “ground truth,” as we call it, is by one.
In the world, nearly every thing is a member of multiple groups, which we learn as humans independently but also contextually. Designers (and linguists, of course) love to cry out, “Context!” and I’m no different: The context of appearance of a given token will help determine its label. And in speech data, in terms of a practical application of your work, it’s usually the case that only a single characteristic at a time is relevant.
Let’s say you wish to play a song, and for the sake of argument, let’s assume the song title is unique and there’s no issue with different versions or artists. You say to your speech interface, “Play Don’t Stop Believin’,” and the system plays the song. As long as the song is played, there’s nothing else to think about, right?
Except in the universe of this particular song, there were any number of releases that contain it. It was released as a 7″ vinyl single, and also featured on an original album, as well as some number of greatest hits collections in various formats, and finally, decades after its original release, a digital single. Semantically, the song belongs to all of these entities. In the universe of software however a single song played, and it was only one of these.
Now, I’m a big believer into functional equivalence. The song’s the song. But essentially my request that was not definable as a single entity was interpreted as one, and if that data is traced, the song as an entity will be a part of a specific release. This was not intended by the user, but functionally, it will have been forced as an interpretation.
What does it mean to force something into meaning in this way? At the given moment of the request, it was likely trivial, but in our universe of persistent data and AI that is not controlled by users, we don’t know what will be used or to what purpose at a later date or for other reasons.
In the case of Don’t Stop Believin’ it’s trivial, but let’s say I’m only familiar with the original album release, and later in my software I see my request was interpreted as a selection from a greatest hits album or a soundtrack.
(The audio nerds in the house will be going, “well yeah sure but what about the the actual qualities of the individual releases? The lp was likely mastered with a specific dynamic range for release as a record, but all of the subsequent formats and repackagings were likely remastered from that one to fit the new one, rather than mastered from the source. Source mastering is expensive and time-consuming, and many listeners won’t know the difference, and even if they did, most playback devices and environments won’t allow the difference to be heard. And so on. Shush, you.)
If you asked for x, the system sets likes x = true, Now you’re an x liker. What if the system subsequently serves anything entailed by x in a tag hierarchy? What if persistently in the user’s knowledge graph, x becomes a defining entity? Let’s say as long as you are a user of that system, for any task that’s been defined in relation to entities like x, new inferences are later made based on the “fact” of your x liking?
Ian Stewart’s thoughts on this landed in my reading list recently and that’s got to be discussed further at another time, but it’s heartening that the question is on many peoples’ lips right now.
I’m sure it’s been hiding in plain sight for some time and I’ve been thinking about it for a while if I have this much to say about it. It’s not difficult to imagine when it comes to qualities of entities like users (i.e. people), identities that aren’t agreed upon, or don’t have equal statuses or cultural interpretations in all places – gender comes to mind – placing a single interpretation on a thing might have the potential to cause harm to individuals and communities?