We use topic modelling usually on a collection of documents - which makes the input. But what if I only have a single document where I want to see the underlying topics in it? I have heard that you can break them by paragraphs in cases like that, but what is the need for that? Does that mean I can't use latent dirichlet allocation (LDA) or it is not supposed to use with a single document as the input?
You can use a sentence splitter and split your document into sentences. I have never used the approach myself, but the tool is available with the open.nlp package in R, Python and Rapidminer.
What you could also do is to train a topicmodel on corpus with clearly defined topics. Next you use the same model on your one document and you see how the topic structure turn out.
In my own experience, using the gensim lda model, it worked out just fine with only one document. Gensim's lda requires a list of "corpus". My list only contains one element. The results, tested on a number of documents, looked pretty good.
It is not the most sophisticated strategy, as it is solely count-based, however, it does the job and you cannot really get more out of it if there's no further input for the model.