Intermediate Vectors for Linguists

Tags: physics intro units math

In my previous article, I start introducing the idea of a vector from the perspective of translating words between two languages. Now that we have a way of representing a translation, let’s see what we can do with it.

Since I’ve declared that this is intermediate vectors, I am going to allow myself the indulgence of using a couple of numbers. Specifically, I’m going to use 0 and 1, since they’re quite useful and come up a lot.

Since I have there numbers, I have a nice chance to expand a bit on a the inner product that I discussed in my last post. I had mentioned that the inner product of two vectors represented how well they matched up. <animus|mind> was really big and <animus|pogo stick> was really small. As you might guess from the words “big” and “small”, we’re going to use a number to represent this inner product. This isn’t strictly necessary - you could use anything that has a big and small. You could use loaves of bread, with <animus|mind> being a giant baguette and <animus|pogo stick> being a tiny little crumb. Unfortunately, I don’t have any bread, so we’ll have to use numbers. I say unfortunately mostly because I haven’t had lunch yet and could go for a sandwich right about now.

Out of the pair of numbers that we’re using, lets go ahead and start with zero. In my last post, I said that two vectors were orthogonal if their inner product was the smallest possible value. Well, zero is the smallest number we have, so we’ll go with it. If two vectors are orthogonal, then they have an inner product of zero.

<pogo stick|animus> = 0

In the above line, I’ve made a fairly bold declaration for someone who isn’t a linguist. That line says that there will never, ever be a Latin text where, when translated into English, the correct way of representing what the original author meant when she wrote “animus” is to use the English phrase “pogo stick”. I’ve known enough linguists to know that one will now construct such a sentence, but I hope that you’ll accept it as a general truth that no Roman author ever meant “pogo stick” when writing “animus”.

We’ve used one number, so it’s time to use the other number. I’m going to declare a convention here that will make our lives easier.

<animus|animus> = 1

The above line says that, if the original Latin text had the word “animus” and you’re translating the text into Latin, it’s always appropriate to translate the word as “animus”.

Or, to put it more succinctly, that it’s always appropriate to translate a word as itself.

If the inner product of a vector with itself is equal to 1, we say that the vector is normal. Now, it might seem that every vector should be normal, but remember that we can add vectors together to make new vectors. The vectors may not be normal. For example, we will look at the vector (|animus> + |pogo stick>):

(<animus|+<pogo stick|)(|animius>+|pogo stick>)
= <animus|animius>+<animus|pogo stick>+<pogo stick|animius>+<pogo stick|pogo stick>
= 1 + 0 + 0 + 1
= 1+1

So the vector (|animus>+|pogo stick>) is NOT normal, since it has an inner of 1+1, instead of one¹.

Orthonormal Basis

In the previous post, I introduced an identity matrix that turns any word back into itself.

ID = ∑ |word><word|

However, I’m afraid to say that I cheated slightly. If we take the whole of the English language and use that to make the matrix, we won’t get something that turns every word back into itself. To take an example, let’s take a subset of the English language that just contains the words “green” and “verdant”.

ID = |green><green| + |verdant><verdant|

If we use this matrix on the |green> vector, we get the following

ID|green> = (|green><green| + |verdant><verdant|)
= |green><green|green> + |verdant><verdant|green>
= |green>*1 + |verdant>*<verdant|green>
= 1*|green> + <verdant|green>*|verdant>
= |green> + <verdant|green>|verdant>

Now, we wanted to get back the vector |green>, but we got that plus the extra <verdant|green>|verdant> term that we don’t want. We could get rid of this term by saying that <verdant|green> = 0. Except, we’ve already declared that the inner product is zero when the two words have nothing in common, so we know that the result is not zero. This ID vector is a total fraud!

Now, there is a way of fixing the identity matrix. We look at a subset of English again, but we insist that all of the words we look at are orthogonal. When you have a collection of vectors which are each individually normal and are all orthogonal to each other, the group is said to be orthonormal. If we restrict our identity matrix to only working over orthonormal words, then it will work again.

You might be disappointed that we had to drop the word |verdant> from our vocabulary in order to keep this matrix representation. However, there’s a trick that allows us to keep it. This trick comes from the fact that, while |green> and |verdant> are similar, they aren’t exactly the same. We said that <verdant|green> wasn’t zero, but it isn’t one, either.

We’re going to make up a new word, verdont, which means something that’s verdant, without all of the bits of verdant that are green.

|verdont> = |verdant> - <verdant|green>|green>

This new word is orthogonal with green:

<green|verdont> = <green|verdant> - <green|<verdant|green>|green>
= <green|verdant> - <green|green><verdant|green>
= <green|verdant> - 1*<verdant|green>
= <green|verdant> - <verdant|green>
= <verdant|green> - <verdant|green>
= 0

However, this new word also let’s us keep the concept of verdant separate from the word green.

|verdant> = |verdont> + <verdant|green>|green>

So, our ID vector can work, as long as we have a subset of our dictionary where a) all the words in the dictionary are orthonormal and b) every word in our language can be formed from a combination of words in our dictionary. This special dictionary is called an orthonormal basis.

Eigenvectors

Now that we know how to make that ID matrix, we’ll go head and make two different vectors:

• ID_e :: The ID matrix for the English language
• ID_l :: The ID matrix for Latin.

ID_e = ∑ |English word><English word|
ID_l = ∑ |Latin word><Latin word|

We can then use these matrices to translate between the two languages.

ID_l |English word> = |Latin Translation>
ID_e |Latin word> = |English Translation>

We can extend this idea to translate back and forth.

ID_e ID_l |English word> = ID_e |Latin Translation> = |English Translation of Translation>

This bring me to the fun little site translationparty.com. The site uses a little script to translate a phrase back and forth between two languages until it gets the same result back.

Can we find something similar with our vectors? What we want to do is find some vector |x> and scaling factor a such that

ID_e ID_l |x> = a |x>

In other words, given the matrix (ID_e ID_l), we want to find a vector that is merely scaled up or down after being multiplied by this matrix. We perform the translations and get the same value back.

This vector is called an eigenvector for the matrix and a is it’s corresponding eigenvalue. Eigenvalues and eigenvectors are quite important in quantum mechanics, but that will probably need to wait for a later post.

footnotes =========