A Black List Neural Network for Screenwriters

I trained a Neural Network for a screenwriting service that compares the style features of a screenplay to the style features found on a number of scripts on the Hollywood Black List published every year.

The service is becoming more and more popular and the screenwriting service receives many questions about the neural network. So I’m writing this article to explain to clients, but also to anyone that’s interested!

The service tries to make it clear that it only analyses style not structure. Nor does it analyse the layers of story or character. It cannot detect if a writer has created and answers a compelling narrative question. Nor can it detect if the writing drives readers to turn over after page.

The closest it comes to such an analysis is a sentiment score across the screenplay. Other features it calculates include: percentage of words that are adverbs, or which are adjectives. How long the average action or dialogue paragraph is. What percentage of dialogue is spoken by the character that speaks the most dialogue. A large number of features like this where developed in Python and the neural network was trained and retrained trying out different feature sets as input.

The actual screenplays used to train were the Black List screenplays of the last few years, plus low placing screenplays in the competition that sells access to this neural network. There are hundreds of examples available in both of these categories. The logic was: a low placing script on a small/medium sized competition is unlikely to be a Black List script. So the network was trained to score 0% likelihood of such scripts being on the Black List, and 100% likelihood of Black List scripts being on the Black List.

In a recent test, some fascinating results emerged. The network has not yet been updated with this year’s Black List screenplay features. But when the top 5 2023 screenplays were tested on the network, they score 75% or above! Similarly, most screenplays submitted to the competition that sponsored the network scored below 10% on the network, consistently.

Now this doesn’t mean that screenplays that score low are not great stories and vice versa. Perhaps it points more to a subtle uniformity of style in the upper reaches of the Black List.

Clients of the competition have sometimes been disappointed with their neural network score, when a human has judged their screenplay to be a finalist, for example. However the neural network is unable to experience the script. It has no sense of the cumulative effect of character development and events. It is not a good judge of a good screenplay.

I am a screenwriter myself, and have experienced the hurley burley of being judged and given “a number” as a rating for my screenplays. I will share something which may be of interest: I had a couple of screenplays – one of which scored a 9 on the Black List website (totally separate from the annual Black List the neural network is trained on), and one of which scored an 8. When I ran these through the neural network, they both scored way above average in probability of being on the Black List. This may be a coincidence, but interested me nonetheless.

So given all these provisos, what is the purpose of the neural network? Well, I do believe that style matters. Style matters far less than story and character- a stylish but boring screenplay is useless. But style added to a great story can improve a screenplay’s reading experience, and the reader’s emotional reaction. So I continue to run my own scripts through the neural network. They do seem to do well on average. Perhaps because I read a lot of Black List scripts every year, and let them influence my style.

One final point: the neural network report also attempts to provide strategies to increase the probability for a writer’s script of scoring high. It will give one to five features that a writer could adjust to increase their neural network score. However these strategies are calculated without taking into account story and character. For example – the report may say to reduce sentiment by 30% (by making the script more sad). Or reduce the average length of dialogue by 50%. And as a result, the report might say, you can increase your neural network output to 85%. The report presents a selection of such strategies, together with what the estimated neural network output will become.

The aim of this is to concretise the whole process. Give some “practicals” that show how the information from the network might be usefully used.

However, we all know that to write a script that has an 85% chance of being in the Black List is far more about story and character than style. It is relatively easy to improve style over 100 pages, but extremely hard to improve storytelling over that length. Also some of the advice may contradict the raison d’etre of the screenplay. Implementing the strategies could destroy what makes the screenplay unique or meaningful.

But the fact still remains that many newer writers could improve their style, and why not use the Black List as a benchmark?