How to mathematically/logically represent the sense of sentences like "The cat drinks milk"?

Question

I am absolutely new in the AI area.

I would like to know how to mathematically/logically represent the sense of sentences like:

1. The cat drinks milk.
2. Sun is yellow.
3. I was at work yesterday.

So, that it could be converted to computer understandable form and analysed algorithmically.

Any clue?

Answer 1

Let start by classify the phrases you propose:

1. The cat drinks milk. => action
2. Sun is yellow. => descriptive/declarative, immutable
3. I was at work yesterday. => descriptive, time related

1) The easiest ones are always the descriptive and immutable (in the context) phrases as "Sun is yellow.". Some usual representations:

• prolog:

color('Sun',yellow).

or simply:

yellow('Sun').

• object oriented:

Sun.color=yellow

2) When the fact is time related as in "I was at work yesterday", we divide the description in a time indicator and a immutable fact:

• prolog:

when(yesterday,at(I,workplace)).

note how when has two parts, the time identification and the immutable fact.

Another prolog variant is:

at(I,workplace,[when(yesterday)]).

where the content in the list (brackets) means "optional related facts".

• object oriented:

I.at = {

position = workplace;

when = yesterday

}

3) Actions as "The cat drinks milk." are a few more difficult:

• prolog:

drinks(cat,milk).

or

action(cat,drinks,milk).

• object oriented:

cat.drinks=[milk]

or

cat.action = {

action=drinks

object=milk

}

Obviously, these are only the main ideas, there are as many representations as different programs, but most of them handles same kind of structures.

( note: the term "computer understandable" is ambiguous. Current computer doesn't understand anything. We say these expression are understandable in the sense that its compiler/interpreter accepts them, and describes the content of the phrase, and the program can transform them to other results).

Answer 2

People normally represent sentences like this as vectors of a specific length, normally about 2500 in length. The algorithm that can do this is sentence2vec. It is basically a derivative of word2vec. It allows you to train a model that can transform sentences into vectors that you can then feed into a neural network or another algorithm. You can check out the paper, which you should be able to find on google scholar. If you need the link, I can get it. Another possibility is word embeddings, which I have not found a good paper on, but cortical.io has a free API that allows you to mess around with their implementation. The word embeddings mimic the real human brain much better based on our current research, but sentence2vec/word2vec is used much more often in practice.