91av

When tasking AI with writing a story, more coherent plots can be created by prompting it to write a story backwards, starting from an ending.

In previous work, AI has been capable of taking an opening paragraph and use it to generate a whole story, sentence by sentence, by applying patterns gleaned from vast amounts of training data. But this approach can quickly wander into incoherence as the length of the story increases. Rather than building a plot with an outline or ending in mind, it simply decides what sentence is statistically most likely to come next based on the literature it has previously read.

Now, a team led by at the Georgia Institute of Technology has flipped this. They used thousands of film and television plot summaries to train a neural network, then gave it a human-composed ending to work backwards from.

The software analyses the last sentence and creates a question about the characters in it, which addresses why they are doing what they are doing. It then creates 15 possible preceding sentences that could explain that behaviour, and finally chooses one as the most statistically likely or appropriate based on a reading by another AI language model. This process repeats as the story builds in reverse.

Stories generated by the system were found to be 15 per cent more coherent in tests with human readers than those created by another AI tool that developed a story forwards, says Riedl. But he admits that there is still progress to be made.

“It’s still rather linear,” he says. “So every chunk of story kind of sets up the next chunk of story, because we’re going backwards. We’re still not able to do some of the more complicated things you see in the movies and books where you foreshadow really far into the future.” This means that twists, turns and surprises won’t crop up in plots written by the AI.

The story length is undetermined, and the AI will create sentences endlessly until it is stopped. “We actually just kind of run it until we get tired of running it,” says Riedl. “So often you feel like you’re entering a story in the middle.” A story is complete when researchers spot a good point to stop the model.

Reference: