Can an AI be trained to generate the outline of a story?

Question

I know that one of the recent fads right now is to train a neural network to generate screenplays and new episodes of e.g. the Friends or The Simpsons, and that's fine: it's interesting and might be the necessary first steps toward making programs that can actually generate sensible/understandable stories.

In this context, can neural networks be trained specifically to study the structures of stories, or screenplays, and perhaps generate plot points, or steps in the Hero's Journey, etc., effectively writing an outline for a story?

To me, this differs from the many myriad plot-point generators online, although I have to admit the similarities. I'm just curious if the tech or the implementation is even there yet and, if it is, how one might go about doing it.

Answer 1

As far as I am aware, this has not been done yet.

I see several problems with this. A neural network is basically a classifier, which matches an input to an output. Both input and output are usually numerical values, though they could be matched to concepts or words.

To train a NN you provide an appropriately encoded input, and the corresponding output. The NN learns the associations between the two, and can then classify unseen input accordingly. This has recently been used to transform images in a particular style etc.

What would the input and output be for generating screenplays? You could use previous scripts as inputs, but what would the output be? It could be narrative 'moves' of some sort, perhaps. So you could train an NN to recognise narrative elements from screenplays.

However, you are still not creating anything, but just recognising stuff. You would need some other input. I guess you could train an NN on "The Simpsons", get a narrative structure, and then present it with an Episode of "Friends" and see what happens. It won't be a new episode of a screenplay, though.

The other way round might work: you feed it narrative moves (a kind of story skeleton), and get a script out. But it would need a lot of (human) post-editing to be at all useful.

I think an NN is the wrong tool to use here. There has been work done with generating stories and screenplays, even way back in the early days of AI. But that was all based on symbolic AI, not on the kind of ML which seems to currently be en vogue. Have a look at James Ryan's website; he has recently written an overview over historic approaches to story (and screenplay) generation.

Answer 2

Ideally, yes. Ideally, because the network should be fed with the words of an entire book (wich vary around 100k words). With an hypotetical amount of processing power, you should be able to just train the NN with like thousands of books. It might be possible to be trained with quantum computers.... who knows...

For smaller stories, I think that the major problem is to know in what "shape" should the story be generated. Because if it simply outputs some words, then the first thing the network should be able to do is speaking, that means the model should evolve from a pretrained NLP model, and (from what I know) we still have some problems with that.

So.... I really think that to do such kind of things, the approach we take to make NNs learn should be changed. The fact that humans exist proves that genetic algorithms would work 100%. But we obviously don't have 3+ billion years to evolve a "brain" from scratch, that's why we use training algorithms: we force them to learn from something.

But back to the question: humans do a lot of work by thinking about what outcome to chose. To just make a netork generate an outcome, without imitating humans, it would be easy to just choose randomly some aspects of this outcome. For example, a randomly chosen outcome might be "outcome: Dennis dies, Morty kills Eminem, sad sciene, happiness sciene, the end". That means that the NN or any ML model doesn't actually produce an outcome to the story. In fact, what it does is to connect some generated "checkpoints" about that story. Actually, you might train a model to generate checkpoints to but this is just a random idea from a newbie, so I've got no clue about how to actually implement that.

I'm italian b.t.w., sorry about my english :)

Answer 3

The answer is yes, an AI can be trained to write even a whole story. I just want to tell you right off the bat that an AI already did something even more difficult than generating a story. I'm talking about that thing at the end of my explanation.

All the links in my explanations are leading to external sources that I found, you can go check them. Without any further do, here are the main reasons why I think AIs can generate the outline of a story:

1. AIs are really good at recognizing patterns, and generating things that are similar to others. Surprisingly, there's a lot of patterns in stories. Stories are always structured, so this part isn't the real problem. There's a great Wiki about the seven basic plots.
2. But even if an AI can generate a good story structure, can it make a story appealing? Well, it depends on how big the "brain" of the AI is. Because it turns out that the more neurons and synapses an AI have, the more it can "understand" human language or emotions. So, if an AI has a big enough brain, it can generate stuff that make sense. Here's the best example of an AI being able to generate human-like stuff : https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html.

For the how, I think that the training data isn't insignificant. So, to be able to train an AI like that, we need a lot of examples. This is possible, because movies' screenplays are public, and can be downloaded by anyone. So, an AI can easily learn from this huge amount of screenplays. Here are some examples of websites where we can get screenplay of movies: https://stephenfollows.com/resource/sites-to-find-movie-scripts/, https://www.simplyscripts.com/movie-screenplays.html.

After that, we just need to format the data, so we can give it to our AI. In my opinion, it's completely possible to make a good AI that writes good stories, because Google already did something similar. I think that the chatbot Meena, created by Google, is the proof that an AI can learn way more than just pattern recognition.

Answer 4

As far as I know, there isn't any system like you describe yet. However, there are some interesting approaches to narrative intelligence that can be found at the University of New Orleans Narrative Intelligence Lab site: https://nil.cs.uno.edu/

Hopefully those can be helpful in guiding a deep-learning approach to narrative generation problems.