Everybody knows what proposition is. It's a sort of statement of fact that is hidden in most sentences. Statements like "cake is delicious", "government collects taxes", "vegans don't eat honey", etc.
Logicians love to take propositions and convert them into fancy looking notation that make look something like delicious(cake), collect(government, taxes) or even ∀x (Vegan(x) → ¬Eats(x, honey)). But most of the time we talk about propositions, we do it in a more mundane way - propositions are statements in which we propose which things are or should be true. Like "proposition G)" which proposes that San Francisco should offer Algebra 1 to students by eighth grade.
Propositions are all the boring statements we make such as "it's getting hot in here" or "she stubbed her toe when she kicked the cat", etc. Propositions are even hidden in all the questions and commands like "do you want me to open the window" or "please, don't kick the cat!".
So, propositions are the sort of sentences we see around us. But what are schemas?
Simply put, schemas are what is actually in our brain to make it possible to make and understand propositions. Propositions are so normal and boring that we don't think much about how we know what they mean. When they think about propositions at all, most people assume that understanding them is just a matter of knowing the meaning of all the words. If they thing a bit more, they will admit that there are also probably some rules for combining them and discovering those rules may be kind of fun.
Those people often become logicians or philosophers of language and if they think propositions are just a matter of "words and rules", they are wrong. We've known that they are wrong at least since the 1970s when all the attempts to build thinking machines out of words and rules failed.
That's when people in AI started to talk about things like schemas or frames and people in psychology started to talk models or gestalts and people in philosophy started to talk about things like language games and family resemblances. (Philosophers and psychologists have a bit of a head start on AI because their view of the world started to fail a few decades earlier.)
So, to illustrate what a schema is, we can take an example from of those people who started to think about this in the 1970s - Terry Winnograd. I'm going to use the two sentences most often cited as examples in the now famous Winnograd Schema Challenge.
Here's the first sentence:
The trophy did not fit in the suitcase because it was too big.
This sentence contains two propositions:
The trophy did not fit in the suitcase.
The trophy was too big.
This may lull us into thinking that the second proposition is a very straightforward encoding of a very simple logical relationship A > B because we can rephrase 2 as "the trophy was bigger than the space in the suitcase"
But the fact that a schema is necessary to understand the sentence is only revealed when we consider a different version of the same sentence:
The trophy did not fit in the suitcase because it was too small.
This sentence also contains two propositions:
The trophy did not fit in the suitcase.
The suitcase was too small.
This sentence looks very similar to the first one because it describing the same state of the world. But yet, it contains too very different propositions. And all it took to change the second proposition into one with a completely different subject was changing the word "big" to "small".
What gave these words the power? All those words together add up to an image in our mind that looks exactly the same. A suitcase and a trophy that doesn't fit in it. But within this image, we know that the word big refers to the trophy and the word small refers to the suitcase. And we know it without any calculus, it's just kind of obvious. How?
All those things in our mind that make this understanding possible are what I'm calling schemas. Rather than give a definition, I'll invite you to think about what all the things that are in your head and are relevant to this look and feel like.
The schemas are the rough shapes of the objects and their relationships that make it possible to understand sentences like this one. They are not easily visible to us, we have to derive them from how we know we think.
Here are some of the schemas that are necessary to make sense of the sentence about trophies and suitcases:
Containment <= triggered by the word fit
Relative size = Container > Content of the container <= Inference of the containment schema
Container size = suitcase (vs. locket or a shipping container)
Contents of a comparable relative size to container <= trophy (vs. e.g. a pebble or a stuffed elephant)
Type of containment = complete loose containment <= suitcase (as opposed to incomplete containment with e.g. a basket or tight containment e.g cement)
Much of the early history of AI (and computational linguistics) involved trying to compile a list of propositions that could be specified as a set of combinatorial rules from which language could be parsed and generated. They all failed utterly. One such attempt was the CYC project which ran from 1984 until early 2000s before it was abandoned.
If it is so obvious that we cannot list all the propositions to capture all aspects of the schemas we use to produce and understand language, why did CYC even try? The simple answer is that this was not obvious, at all. The entire history of thinking about language was based on the assumption that the real meaning is in the propositions and all the rest is just fluff nobody really needs to worry about. The art of reducing language to logical formulas was the sort of thing that got you accolades and a reputation as a highly intelligent person.
Early computational linguists and theorists of AI were focusing on solving rules of inference from logical propositions and believed all the simple things like trophies fitting into suitcases would be mopped up once all the "hard" stuff was solved.
Many people still believe this today. You can see why. When we look at this particular set of propositions and their schemas, it would seem relatively easy to list them all. All we need is a dictionary that includes things like size and shape of objects (like suitcase and trophy) and schemas like "containment" attached to verbs like "fit" and a few inference rules of how they fit together.
But this is sadly not the case. Even the containment schema is more complicated than container > contained. What does the word "not"actually negate in that sentence? It does not flip the containment schema around. It does not actually say that the trophy was bigger than the suitcase. Just that it was too big to fit in it. And there could be many reasons for this which our schemas of the whole situation will readily supply.
Our mental image of "the trophy does not fit in the suitcase" is different than the mental image evoked by the joke "How do you know, there's an elephant in the refrigerator? The door doesn't close." The joke is funny because the schema of containment only applies to things that could possibly be contained in each other. The schema of a container when combined with the schema or a refrigerator and the schema of failure of containment will include things like turkey or 4-gallon jars but not elephants or cars. Because the things that don't fit in a fridge in any normal scenario are not actually bigger than the fridge. So, we don't really just have schemas and their negations, we have two separate schemas - one for containment and one for failure of containment. And each of those have many many subschemas for different ways this could happen.
That's another example of a why listing all relevant properties of schemas is so hard. The nature of the schemas only emerges in interaction and it is almost impossible for an individual or even a team to come up with all the possible statements of the schemas just by thinking about them. The CYC team worked on this for 10 years, ended up with millions of propositions specifying relevant schemas but still had to reduce its scope to a small subset of knowledge.
And the containment schema is a relatively simple one. We can see why this is so hard when we look at the second sentence often used to illustrate the Winnograd schema challenge.
The council members denied the protesters a permit to hold a demonstration because they feared violence.
This sentence encodes two surface propositions:
The council members denied the protesters a permit.
The council members feared violence.
But if we change the sentence slightly, we get two very different propositions:
The council members denied the protesters a permit to hold a demonstration because they advocated violence.
This sentence encodes two surface propositions:
The council members denied the protesters a permit.
The protesters advocated violence.
To fully decode the schemas involved here, we cannot just simply list one general schema that we could draw a picture of like with containment. We need something much more complex and subtle. We need a scenario or script that includes the following:
Protesters hold protests in public spaces (streets)
Protests can be peaceful or violent
Council members decide about permits on the basis of whether a protest is likely to be peaceful or violent
If council member fear that the protest may become violent, they will deny a permit
If the protesters advocate violence, they are giving the council members a reason to deny them a permit.
But this is a very anemic list of facts compared to what is actually involved. It's not at all clear how we could specify the list of logical inference rule in a dictionary that would make it clear which word "they" refers to in the two sentences. The key to making that decision is not in any of the words, it's in the picture of the scenario that they describe when they are put together.
But also, that's just the word "they". There are many more things we "instinctually" know when we hear a sentence like that. They include things like the likely political motivations, context of what the protest may be about at the time we're hearing the sentence, etc. And we can easily imagine that they would be used in interpreting the context. Imagine, some follow up sentences and what knowledge of the scenario is important to understand them:
This sort of behavior should not be tolerated in a well ordered society.
This sort of behavior should not be tolerated in a society that values freedom of expression.
The arcane bureaucratic procedures are too complex to navigate for young idealists.
They need to learn that rules apply to everyone.
Local businesses breath a sigh of relief.
Local businesses mourn the loss of income.
All of those sentences rely on a different aspect of the scenario that is part of any human's understanding of the world but is impossible to list in a dictionary or a some sort of a semantic grammar book. We could write pages after pages of stories and contingencies just to lay out all the things we think matter and we would still miss some. Famously, large dictionaries take teams of people decades to complete. But still dictionaries even if we generously throw in thesauri and encyclopedias are hopelessly behind what it would it take to capture all the schemas involved in understanding all aspects of just the sort of propositions we encounter in the course of a single day.
So, this is what schemas are. Impossible to capture but absolutely at the heart of our ability to make sense of the world and the propositions we put forth as part of that process.
But unfortunately, things don't stop there. Schemas show up in places where we least expect them. The rest of my time on this Substack will be spent jumping out from behind innocent scenery like a bunch of cardinals in a Monty Python sketch and yelling "nobody ever expects schemas." Stick around, you'll never see it coming.
