While your question has some ambiguities, I try to answer.
From my understanding you want your model to predict “topic” of a sentence or a description. It’s just a classification problem with huge possible number of output classes.
The first initial issue is very short length of documents (sentences). Most of topic modelling algorithms such as LDA have statistical approach and do not work very well with very short documents (less than 50 words could be a good definition of very short document).
The second issue is how do you want to collect enough data to train your model that supposed to predict target out of extremely large number of output classes? Dictionaries are not enough because they offer a single definition for each word. Examples of words in dictionaries don’t help much and they will probably affect your model adversely. How can your model be generalised by a single (or few) example(s) for each class?
So, it’s not possible, but maybe having some innovations can help.
Here is definition of “apple” in oxford dictionary: ”a round fruit with shiny red or green skin that is fairly hard and white inside”. There are just two nouns in the definition: "fruit" and "Skin", if we just read the definition without considering these two words, even we, as human, struggle to guess.
Consider nouns in input data and use them to build up a natural graph. You just consider main classes such as "fruit". If you’re getting some good results, consider other words, adj, adv, ...