How can we use the same word in multiple different ways and distinguish the senses so easily?

Question

Say for example some plant names. We have an orange which we easily know is a fruit, but is also a color. We have green which is a color, and greens which is plants, or money, or I could imagine it being a sports team like the Red Sox, but even if we didn't capitalize it we would know the red sox is a team, not red [x].

For plants we have poison oak, which is a type of plant, not "poisoned" oak. But you could have poison greens, which are actually poisoned. I don't see how we tell the difference. Poison is also a band, and a bunch of other stuff.

There are other plant names like sage. But a "sage" is a type of person as well. It's also a seasoning, and incense. Sage has a ton of disparate meanings.

Wondering what is the process I am going through to realize the meaning of these words which have multiple totally different meanings.

They went to the forest and walked by some sage, holding some sage next to the sage [their friend]. They saw some poison oak and poison greens in the greens, but not Poison or green.

I understand that "context" is an important factor. But I don't quite see what context exactly is coming into play when I read these sentences or see some words. There is punctuation as in Poison vs. poison. There is adding extra description as in sage [their friend]. But other than that, there is not much literally there.

When I see coffeeberry, I don't think "coffee", even if I don't know what it is. It has "berry" in the name so maybe it's a fruit like boysenberry. Turns out it's a plant. So perhaps I couldn't guess what it is without learning it's meaning explicitly. When I read dogwood I don't think "dog", maybe because I would probably see that related to some botany event. I kind of think of a band when I hear that name, like Fleetwood Mac and Three Dog Night, so somehow my brain is trying to make sense of the new input. But I don't see how it's doing it.

But we can get more complex than just using nouns, and start converting these words in verbs too! Dual formed words.

So we can say "They greened the landscape." to mean they planted lots of plants covering the landscape. Then we can say:

I want to use my green to green the landscape with some greens.

I want to use my money to plant some plants on the landscape.

There it is again, plant a plant. "I want to plant a plant in the plant". (Indoor plant/factory).

So I am wondering how we are able to determine the meaning and form of these words whose meaning is totally different in different places in the sentence. As an extreme example, say we converted the word "I" to a verb. And then we have 10 different things we've called "I" (like how "Sage" has like 50 things called sage). Say we even created 10 or so verbs called "I", similar to how "plant" as a verb has multiple meanings (plant a plant, lay something on the ground). So we have this sentence.

I I I I I I I I I I I I.

Obviously this doesn't make any sense because we created too many same-named words. So there is a balance of some sort. But I would like to know what the balance is, when we are allowed to map new meanings to existing words (that is, create new words using existing words as their name). When too much is too much.

Basically I don't understand how it's possible for us to understand these sentences. I would've assumed that every word has to have exactly one meaning in order for a thing to be understood.

So the question is, if there is some sort of way to tell if adding new meanings to a word (i.e. repeating a word's name with different meaning) will make it easier or harder to comprehend a sentence. Related to this but probably a bit too lengthy is how we go about understanding the meaning of these words.

I am also familiar with the ideas of sentence structure like having SVO and how we can use that to help figure out the meaning of words. So no need to explain any of that :)

Answer 1

The number of words humans produce are 'just enough' to allow you (and hopefully another party) to disambiguate the content. We don't stop speaking, or writing until we believe we've disambiguated the thought. If we communicate something that we believe is ambiguous... we just keep talking/writing. Sense disambiguation uses a ton of little clues that logically reduce a set of a polysemous word to a single sense.

That said,...

1. We use the 'topic area' as our initial clues; they reduce the set of senses to a limited, specialized lexicon. If we're watching baseball, we know what they mean by 'strike' or 'ball'.

2. We use words in the surrounding clauses (before and after). If I say "I got an extension but was surprised to see how much my taxes were.", you still know that 'extension' was a 'tax extension'. To my computational linguist friends, note the window extends beyond the traditional colocation boundaries.

3. We use grammar as clues. Part of speech, compound nouns, coordinating conjunctions, etc. If I see "He ate apples and bananas", "and bananas" helps me to group it with apples (fruit) not computers.

4. We use common sense as clues. If I see, "the man ate an orange.", I know that "ate" as a verb, and "orange" as a direct object, is a common pattern for 'consumption of edibles'. Better yet, don't think of it as "common" sense, think of it as "frequent" sense, (as in a frequency count).

5. Everything else. Morphology, possession, type of determiner preceding a noun, type of argument complementing a verb, quantifiers preceding partitives, etc.

The problem of 'Word Sense Disambiguation' is an old problem; it's easy for humans and (historically) difficult for computers.

New senses are added by either defining them inline (appositives, etc.) or by using a pattern. If I say, "The Yankees winked the Red Sox 10-1." You'd recognize the pattern of {Team-1 verbs Team-2 Team-1-score Team-2-score} enough times to recognize that verbs means "beats". Suddenly, you're wondering if "winks" has another sense, meaning "beats". You might be skeptical, but you see it a few more times. Later, you start using "winks" as a synonym for "beats", and alas, we have a new word sense.

Answer 2

The simplest answer is that context is stronger and contains more clues than you think.

Have you seen IBM's Watson play Jeopardy? Check out, for example, this video around 45 seconds in. The prompt is this:

Kathleen Kenyon's excavation of this city mentioned in Joshua showed the walls had been repaired 17 times

Watson correctly answers "Jericho", but interestingly, we get to see the other words it had in mind with their percentage of probability:

Jericho 97%
Jerusalem 42%
Lachish 7%

Even though this supercomputer generally models language quite differently from how we think humans parse it, this particular strategy is similar to priming. In us, too, the presence of related words triggers associations — but not just one. Instead, it seems that, like a good search filter system, cues from the context cumulatively narrow and weaken the likelihood of some candidates to let others float to the surface, close to the docks where you stand with your fishing line...

A classic finding, for example, tests the interpretation of the word "port". Quick: define that word!

If you're like most people, you probably read that word and gave a definition like "harbour" because of the last sentence of the previous paragraph. If instead I had used a metaphor involving words like "sweet" and "intoxicating", you would instead have said it was a kind of wine. That's the effect of priming. However, it isn't black and white, as we've noted. You can easily make the Gestalt shift between the two if new information comes in, re-ranking the relative percentages. (Thank goodness this process is automatic!)

The reason A.M. brings up double articulation is to help distinguish between the building blocks (meaningless but distinctive) and the significant elements. On the most basic level, this is a model of how phonemes like /k/ and /b/ have no meaning in themselves but permit us to distinguish between the meaningful items "cat" and "bat". On another level, the one we're interested in, there are lots of cues in the overall discourse that are not themselves the topic but permit us to distinguish between topics. Or syntactically, the places of the words in the syntactic structure is not the meaning but permits us to distinguish the meaning. For example, in "plant a plant" it's unambiguous that the same phonemes must first be the verbal lexeme and then the nominal lexeme. And the first one is already working, from the moment you hear it, to narrow down the second one so that you probably imagine a green, growing thing instead of a factory, since we don't "plant" that kind of "plant".

The examples closer to the edge, like your "sage" sentence, have a different answer. How do we distinguish them so easily? We don't; they are indeed harder than usual contexts! And the extreme example with the homonymous "I"s is far worse. In theory it's not impossible, since people can eventually parse "Buffalo buffalo buffalo Buffalo buffalo", but the relationships between the items are so disguised and the range of meanings so unfamiliar that it's a chore. Another type of case where the terms are syntactically undistinguishable and require special knowledge is slang, like the compliment "Your teacher is bad!" Even then, tone, gestures, and discursive context can help clue in an observant person to the suspicion that "bad" doesn't have its usual sense, but when these cues are missing, as in text-only communication, the system comes up with the wrong answer.*

Incidentally, we also prime based on phonology, and this is one of the proofs that there is some reality to non-atomic blocks such as the syllable: syllables shared between words have a stronger priming effect than arbitrary sequences of phonemes shared between words. And by the way, "coffeeberry" and "dogwood" do prime words related to "coffee" and "dog", but the head of the compound has the stronger effect.

---

_{* Of course, even Watson gets it wrong sometimes.}