Truth-conditional semantics

↵ Back to class homepage

To really understand what pragmatics is all about and why we need a model of pragmatics at all, we need to look at what ideas the theory of pragmatics emerged as a response to; we need to look at what semanticists believed about language before the study of pragmatics emerged.

Understanding the details of semantic theory is not strictly necessary for this class (this is, after all, a pragmatics class, not a semantics class), but at least understanding the general spirit of the ideas—and, most importantly, the problems with them—will help you understand what pragmatics is needed for. Pragmatics also uses a lot of semantics concepts, so some of these same ideas will reappear throughout the class even though it's not really a semantics class.

Because this is a pretty tricky topic, we're going to have to divide it across two modules: first we'll see what the main idea of semantics is, and then in the next module we'll see where it breaks down and why we need pragmatics to fill the gap. Things will start to get a bit abstract and crazy here, so buckle up.

Meaning as reference?

Semantics is the study of meaning, and its job is to figure out how we understand the meanings of expressions: how we understand the meanings of individual words, how we put those meanings together to make more complicated propositions, etc. So one of the central questions of semantics is: what even is meaning. What does it mean to "mean" something?

One of the early ideas (going back at least to the early 20th century) was that words refer to things in the world. E.g., you understand the meaning of the word "cat" by picturing a particular cat in your mind. In some way, the meaning of the word "cat" is a cat. In other words, the meaning of a word (or multi-word expression) is the thing it points to in the real world.

It turns out, though, that there are tons of problems with this idea. Here are a few of them:

These all highlight that meaning is not just the thing that some expression refers to in the real world. These problems led semanticists and philosophers of language to recognize a difference between reference (what an expression refers to) and sense (what an expression actually means). But that still brings us back to our original question: we can see now that sense is not the same as reference, but what is sense?

Meaning as truth conditions

Going deep into that question is beyond the scope of this class (but take a class in semantics or philosophy of language if you really want to make your head spin!), but let me just summarize the main idea that lots of semanticists have more or less settled on.

Consider a dog. You probably have some mental definition of what a dog can be. There are probably some features (it has four legs, fur, it wags its tail, etc.), although some of these may be more crucial to the definition of "dog" than others (you can probably imagine something that doesn't have four legs but is still a dog, something that doesn't have fur but is still a dog, something that doesn't wag its tail but is still a dog, etc.). The details of this definition are not crucial here (it's often difficult to explicitly express what your mental definition of some concept is). The important thing is, any time you see some thing in the real world, you could decide if it is or is not a dog.

We can think of the meaning of the word "dog" as sort of a black box inside your head (semanticists, philosophers, and logicians call a black box like this a function). You can put anything into the black box, and inside that box some magic that we don't fully understand happens, and then the black box spits out one of two things: TRUE if what you put in was a dog, and FALSE if what you put in was not a dog.

An image containing a picture of a dog, then an arrow pointing from the dog to
			a black box which says \ An image containing a picture of a clownfish, then an arrow pointing from the fish to
			a black box which says \

This may seem a little stupid and circular. (The meaning of "dog" is knowing what a dog is? Duh!) But there is an important insight here. If you know what "dog" means, then you know what could be a dog—even if you see something you've never seen before, you can decide if it's a dog or not a dog. If you can't take any new thing and decide whether or not it's a dog, then it seems like you don't really know what the word "dog" means.

These black-box functions can be combined together. For instance, say you have one function representing the meaning of the word "pie" (if you see something in the world and insert it into this mental function, the function will spit out TRUE if that thing was a pie, and FALSE otherwise), and you have another function representing the meaning of the word "eat" (if you see some event happening in the world and you stick it into this function, the function will spit out TRUE if that event was an instance of eating, and will spit out FALSE if it was anything else). These can be combined together to make a new function which spits out TRUE if the event you're seeing is an instance of pie-eating, and FALSE in any other situation. (The details of how this combination happens are beyond the scope of this class, but are studied in the field of compositional semantics.)

Another name for these black-box functions is truth conditions. If you know what some word (or bigger expression) means, you know the conditions it needs to meet in order to be true. That is the "mental definition" we discussed above. Importantly, you can know what an expression means without knowing if it's true or not. If you hear the expression "The last person who came to my office was wearing a purple T-shirt", you probably don't know if that sentence is true or not (unless you saw who came to my office last). But you know what it would take to be true, and if my office had a security camera (it does not!) and I showed you a video of the last person to come then you would be able to decide if the sentence was true or not.

The idea, then, is that meaning is truth conditions. If you know the truth conditions for an expression, you know what it means. This can handle all the problems raised in the previous section. Take, for an example, a sentence like "The current king of France is bald." Since there is no current king of France, some people say the sentence is neither true nor false (or cannot be judged true or false). But we can all understand what it would take for this sentence to be true. Likewise, we can understand the different truth conditions of "Lee Lai Shan" and "the first Hong Kong athlete to ever win an Olympic gold medal"; we could imagine a world in which Lee Lai Shan did not win a gold medal (or in which some other Hong Kong athlete won a gold medal before her), but we could still understand what each of these expressions means because we know their truth conditions.

Truth conditions are closely related to entailment. Entailment is a certain relationship between expressions (or propositions) like phrases or sentences. For example, the sentence "I ate a red apple" entails the sentence "I ate an apple", because if it's true that I ate a red apple then it must also be true that I ate an apple. More generally, some proposition A entails some proposition B if A being true guarantees that B must also be true. It should be clear that an entailment is a truth condition: for the sentence "I ate a red apple" to be true, one of the things that must be true (i.e., one of the truth conditions) must be that I ate an apple. For this reason, throughout this class, I will sometimes use the terms "truth-conditional meaning", "entailment", "semantic meaning", and "literal meaning" interchangeably.


What we've seen so far is that semantic meaning is about truth conditions. But in the next module we'll see lots of ways that truth conditions don't totally explaining meaning.

Video summary

In-class activities

Just a few random discussion questions you can have students consider.

⟵ Three kinds of meaning
The limits of truth-conditional meaning ⟶

by Stephen Politzer-Ahles. Last modified on 2021-10-05. CC-BY-4.0.