In our Culture Crossover series we pick up examples of projects that delightfully bridge the worlds of technology and culture. We'll be reviewing exhibitions, giving you a heads up on cultural events or talks coming up in the UK and highlighting the best techy art.

To read more instalments of Culture Crossover click here.

© iStock/metamorworks
© iStock/metamorworks

Those who joke that Hollywood films are so formulaic they could be written by an algorithm may soon find there is more truth in their smears than they expected, after researchers discovered that they could predict the success of a movie with AI.

The team from Sungkyunkwan University in Seoul, South Korea, used the ratings on review aggregation website Rotten Tomatoes as the measure of a film's success and the CMU Movie Summary Corpus of crowd-sourced plot summaries as the synopsis of its plot. They then used deep learning models to analyse the sentiment of each sentence in the summary.

If the combination of the audience and critic consensuses on Rotten Tomatoes was above 75 percent positive, the film was deemed to be a success. Any less than 65 percent and the film was considered not successful.  

Successful films such as Alice in Wonderland and Das Boot tended to have frequent sentiment fluctuations, whereas the fluctuations were less frequent in unsuccessful films, such as The Limits of Control and The Lost Bladesman.

Yun-Gyung Cheong, an associate professor in the college of computing at Sungkyunkwan University, admitted that the Rotten Tomatoes aggregator was not the most accurate measure of a movie’s success.

"The correct measure would be the ROI, which requires the total revenue and the investment cost, [for] which we couldn’t find reliable sources," she told Techworld in an email. "To make it worse, we need to consider the inflation when the movies were released decades ago. Therefore, we used the review scores instead.

"While there are several movies review websites are available, we thought that the Rotten Tomato score system is best since it offers two types of scores; one by the audience and one by the professional critics. Unlike other 5-scale score systems, the Rotten Tomato scores suggests 60 percent as the threshold to classify fresh (successful in our words) and rotten (not successful). This was another reason we took the score system."

Algorithmic filmmaking

The Sungkyunkwan University project is one of a growing number of attempts to algorithmically predict the success of a film. One of the most prominent exponents of the method is Belgian startup Scriptbook, which predicts a screenplay’s success by comparing the characteristics of the storyline to a dataset of thousands of scripts that have already been released and had their box office business measured.

At the 2018 Karlovy Vary International Film Festival in the Czech Republic, ScriptBook founder Nadira Azermai said that the system retroactively predicted 22 of the 32 Sony movies that had lost money between 2015 and 2017.

Los Angeles-based Cinelytic adds talent analytics to the mix. Its platform allows producers to swap one actor for another and find out how the casting change could affect the box office results.

Not all the algorithmic predictions have proven to be accurate. Relativity Media’s attempts to use a Monte Carlo model to run thousands of simulations that could estimate the odds of a given film couldn’t save the media company from filing for bankruptcy.

Whatever the benefits of these systems, they are attracting growing interest from studio executives. In 2018, 20th Century Fox revealed that it had been using machine learning to predict the films that people would want to see by analysing the content of trailers to identify patterns of success.

Dr Polina Zioga, who founded the Interactive Filmmaking Lab at Staffordshire University to explore interactive filmmaking, told Techworld that these tools could help producers but would struggle to take account of all the subtleties of human experiences of films.

"When we go and watch a film, we watch it at a specific period of time in a specific place. So I think the responses are quite complex," she said. "These tools can probably help the industry, but that doesn't necessarily mean that they will be the tools that will provide the ultimate solution and prediction for all cases, and at all times."

Critics of these systems worry that they will lead studios to only produce films that are highly likely to succeed, reducing the opportunities for more unusual scripts and new screenwriters. Cheong argues that they could in fact help discover and develop writing talent.

"When the prediction result of successful stories gets enhanced, our techniques can help them discover new talented screenwriters," she said. "Moreover, it can also help the writers to evaluate their own works."