AIs, writing, and computational literature

bender-penA few months back, I stumbled upon Inkitt. Well, more like they stumbled upon me – they were looking for someone with a background in analytics and in writing to build models around the stories in their community. The goal was to build analytic models that would understand what a good story was to basically create an AI submissions editor, an AI slush-pile reader.

On the one hand, Iniktt is building a community of writers (much like Wattpad). On the other (the business model), they are selecting the top novels to offer to book publishers. Should the book publisher not take the manuscript, Inkitt, because they already think the novel is good enough for a publisher, will publish the book. If the book sells well, Inkitt will return to the publisher once more. If the book doesn’t sell well, the rights revert to the author.

Of course, the key is to find the good novels (isn’t success in publishing always about good stories?). The community will bubble some of this up, but perhaps having a model that learns from the community what is good could accelerate the discover of new novels.

Building models of literature
I found this intriguing and started looking into computational literary analysis, also known as Digital Humanities (there’s even a journal). I uncovered a long history of work to make sense of different forms of writing, being able to analyze writing as a scholar would (here’s a recent article from Berkeley).

IBM has championed the concept of “cognitive computing“, a third wave of computing after the first two waves of tabulation and programmatic computing. In cognitive computing, systems are no longer programmed by human-generated rules, but are taught through machine learning and models trained from real data (and plenty of nudging from human specialists).

We do this at work – we feed a corpus of text into our system, along with what ontologies our experts have to give some semblance of meaning to the text (that’s the hard work some people gloss over), and the system builds a model of understanding, pulling together the relevant topics (you can see it in action here). This is how organizations are getting better at understanding sentiment, tracking leading topics, going beyond keywords and rules to build a responsive system that no human alone can build (though, don’t get swayed by the hype, as this very good article warns).

So how are folks teaching systems about story?  By giving them something to read. Facebook is teaching its system by feeding it children’s books (see reading list here). Google has been feeding a system with thousands of romance novels. Alas, these two companies are not necessarily trying to build a model for what a great children’s book or romance novel is. They are trying to tech their systems how humans converse, to better provide conversational services (bots!). Though, as many parents of early readers know, what goes in is what goes out, and young conversationalists are quite impressionable (read about the Microsoft bot). But these systems will end up being a smart as a puppy. Here’s Google’s system with some exercises that look like beatnik poetry.

Folks have also been going beyond conversation and having such systems actually write novels. For example, for NaNoWriMo (National Novel Writing Month) writers spend the month of November writing a 50,000-word story (quantity over quality). NaNoGenMo (National Novel Generation Month) is a riff off of NaNoWriMo – participants build programs that create 50,000-word stories, using computers (hm, I wonder if something mechanical would count). The exercise generated some quite fun results (not to mention the call-outs to @hugovk, whom I know, and not surprised he dove into this). I am not sure how many of these were programmatic rather than cognitive-like, being more human programming cleverness rather than machine-original.

Does it matter who writes it?
I think the distinction between all-machine or all-human or a hybrid writer is irrelevant. Already financial and sports news reports are written by machines. I received spam that reminded me of Burroughs’ cut-up fiction. A machine-generated novel recently made it through the first round of a literary contest. To me, if the story is good, does it matter who wrote it? Rather than ponder _if_ an AI can write a novel, we should be thinking of how do we live in a world where AIs write or help write novels.

I have just spent the past many years feeding and encouraging a writer (human, that is). It’s a joy to share books, writing, discuss plot and style, and practice, practice, practice. AIs will be the same – we, the humans, will give them the tools to learn and grow and find their voice. What’s wrong with that?

I, for one, welcome my new novelist overlords.

Now, excuse me, as I point my AI to go play on Wattpad.

Image by Tony Delgrosso