How to validate absurd judgments/suggestions?

Question

Text Processing Automation

Example:

The air elephant basked on a green cloud during a glowing round cube where square trees descended from the left wheel stepping over a spherical horse in a vacuum.

Are there any methods (theory) and developments (code) in this direction (most of all - the theory is of interest)?
In advance, thanks to all who responded.

Answer 1

The truth table of all valid natural properties of objects, taking into account the correctness of the logical relations of each of the properties among themselves.
Example: Air elephant
Reasoning logic (code operation algorithm):
1. Elephant -> animal, moves, walking on a dense surface, breathes air (partially filled with it)
2. Air -> Air -> leans on any surface (outside), fills any free space (from the inside).
3. Air -> Fills from the inside or surrounds the object from all sides.
Conclusions:
1. An elephant is not 100% air.
2. The elephant is not surrounded by air on all sides.
Outcome: Air elephant - impossible, the judgment is absurd.

Answer 2

only by statistics, when recognizing, we are forced to assume that the text makes sense for the speaker, and to draw a conclusion about absurdity - only through the accumulation of typical properties used,
i.e. after training for 10,000 texts, typical properties will be highlighted for objects,
and if we say to teach on avant-garde poems, then the original example will be meaningful, this is normal
absurdity = majority opinion, i.e. statistics

Answer 3

I'm just wondering how you are going to do this without a dictionary? It is clear that almost all adjectives can be applied to an object (an elephant can be wounded, in love, sick, Indian, trained, etc.).
The first option is pairing. There are 3 tables - nouns, adjectives and their relations, in which the status of the pair (true / false) is also written, such as "Indian elephant" - true, "Liquid elephant" - false. At the same time, it is worth considering that some groups of adjectives, for example, comparative (or whatever?) are applicable to 99% of nouns - BIG bush, BIG nose, BIG problems, SMALL growth, SMALL opportunities, etc.
The second option is to teach the AI ​​the language so that it understands the context, word forms, because this same context must also be added to the first option, for example, "square trees descended from the left wheel." Or split into pairs "noun - adjective" and if at least one pair is false, then the whole sentence is false.
PS Usually such nonsense is said / written by patients who have problems with their heads, I don’t remember the name of the disease, specifically, but maybe the TC just wants to do it for this, I xs))
PSS I don’t rummage in AI, I understand that dictionaries should be just huge, I'm an ordinary couch theorist running past))

Answer 4

Word embeddings on a large corpus of texts reveals the distances between occurring words, incl. specifies that "puppy" is to "dog" what "kitten" is to "cat". A phrase can be given a "meaningfulness weight" as a function of the proximity of its constituent words. Less weight is more likely nonsense.
The result depends on what texts are used for learning. If you feed scientific publications and if the children's library of fairy tales and fantasy, estimates of the distance "air" and "elephant" will be different.
Another weakness is that the model does not distinguish between the multiple meanings of the words: “looked at the cloud” and “filled into the cloud”. Ps and "looked at the bay")

Answer 5

It's you probably want to comb the crooked translation from one of the services?
It is unlikely that it will be possible to understand where the absurd judgments / suggestions are. Alas

Answer 6

BERT has a mode for checking the logic of two phrases. I think it can be used.

Answer 7

I would do it easier and more accurately.
We take the last characters from all the words that are in the sentence, but up to the characters a, e, and, y.
If in a sentence all words from the end to say three characters forward, well, that is, substr -3, and there is one of them, and there are many of them in the general sentence, say 3 or 5, then spam is unambiguous.
And no AI is needed and dictionaries too.