Generating Essay (Set of correlated sentenses) from a set of keywords

Question

I originally belong to Computer Vision world. Recently I worked on Scene Understanding. There, I accept an image and generate a sentence explaining the contents of the scene. Now, my professor wants me to generate an essay for a scene.

As per my literature survey, this is not a very easy task to do. So I am completely confused with what should be the approach.

I came across Automatic Essay Grading systems, which break down essay using Latent Semantic Analysis. I am thinking of generating reverse approach:

1. From a set of keywords (bag-of-words), form a "random" Entropy Matrix
2. Generate a training set to have a Entropy_Matrix: Essay
3. Train a system (may be Neural_Net)

But I am not sure what would be representation of essay? I mean how would I generate a training set's output entity?

So please help me find the answers to following questions:

a. Is my approach correct in targeting the problem? If yes, how may I further create a concrete plan?

b. What other parallel approaches could be taken into consideration?

My professor thinks this can be achieved in couple of weeks time, for a constrained "context".(Essay is supposed to have a context, let's say "nature" and "beach" are only two contexts to start with) I strongly disagree with him. If you could share written evidences of your view, it would further help me in coming up with a time-frame.

Your help is already appreciated

Answer 1

At first let's review your method: You are going to generate an essay based on semantically related words, that's not enough for an essay to be coherent. It would be context-free grammar. Within the scope of CFG: Yes, you can and as your professor mentioned it only takes several weeks (if you need to prepare noun list, programming issues etc.). A well known example of this method is SCIgen By Jeremy Stribling et al.

As you specifically mentioned ...for a constrained "context" which I presume you want an essay to be meaningful not just gibberish generated by algorithms then you should consider other methods like N-gram based text generation and Neural network

In conclusion, to answer question (a) I must say what is your definition of "correct"? by all means it's a working method but is it the best-working method? or the most efficient? You can't say, currently there are some better methods for generating meaningful textual data; For CFG this is an approach that works.

For a general plan regarding CFG you must first choose your subject, Nature or beach; what are specific vocabulary related to these essays? Simple analyze those specific essays by programs like AntConc and Wordsmith add them to a list (By the way you can use the data in SCIgen's Github page at here). write the code in a language you're familiar with. There you go.

For [deep] Neural-network or N-gram models you need a large corpus of essays (specially for DNN and NN you need very large corpora) and analyze them by code like this which is python.