Artificial intelligence can be used to draw insights out of big data, but also to create new content based on existing texts and images. The Finnish public broadcasting company Yle and the largest AI lab in the Nordics Silo.AI collaborated in an extraordinary pilot project to generate horror stories and fake images of ghosts to support these stories.
Yle and Silo.AI joined forces to see how novel technologies such as general adversarial networks (GANs) and groundbreaking text generation solutions could be used in content creation for the multi-media house. The idea was to use AI to generate personalized horror content that would add on to the existing material on Yle’s website.
Silo.AI’s natural language processing (NLP) and computer vision (CV) experts used text and image generation to create previously non-existent material. By leveraging cutting-edge open-sourced solutions, the pilot project was able to give birth to hundreds of ready-to-use horror stories and ghost images.
Wesa Aapro, who led the project at Yle, comments:
“Once in a while we have a chance to try something completely new and try to see to the future of media. The content generated by Silo.AI was very insightful and it showed us the need to make more similar experiments in order to develop the AI solution further together. “
“Leveraging novel technologies requires adapting them to your usage and you can’t make that happen unless you know what they are capable of.”
Life story for a ghost – written by an AI solution
Our NLP expert Luiza Sayfullina used generative models that are able to produce text based on the source material, so called “seed text”. These generative models form the basis of the creative AI solution. In the following, she describes what was needed from the AI generated horror story:
“To generate a life story for an other-worldly person, you would need meta information, such as name, birth and death dates and something to make it personal, such as the person’s story with his or her hopes and regrets. In addition, since we are talking about a ghost we wanted to add a cause-of-death story.”
We used the following datasets as the basis data to teach our NLP solution:
– Dataset of Finnish names and their frequencies
– Dataset of Finnish cities and the number of citizens for localization
– Dataset for hopes, regrets and “about story” generation
– Open sourced news dataset for cause-of-death generation
– Book Corpus: Charles Dickens “A Christmas Carol”
The two models used in the AI solution were OpenAI’s GPT-2 and another model called BERT. Both are cutting-edge NLP models that had a lot of progress in 2019 in terms of content generation. Nowadays the texts created are getting better in grammar and the topic of the generated text can be controlled by teaching the solution with a longer seed text. In this project, we also used some of the data sources to fine-tune the tone-of-voice towards the desired direction.
Here are some of the AI generated ghost stories:
All my life, I have tried to live a ghost story. I have tried my best to be a normal person, to live in a way that will bring me joy and comfort. But now, in my darkest days, I know I have lost all hope.AI generated ghost story #1
That’s all it takes to be a ghost. I know I will be reborn in my own way again. I will be the first who will have a chance on the crosshairs that lie in our hands. Do not be afraid to be yourself.AI generated ghost story #2
Our AI Scientist Luiza Sayfullina behind the text generating part describes the project:
“It was a super-creative project where I could see how with a few hundred typical sentences one can fine-tune existing pretrained models to generate new sentences that meet the defined topic and subject.”
As mentioned, the two models she decided to use had quite different characteristics:
“GPT-2 model is intrinsically more suitable for generating nearly grammatically correct passages of new text from the seed text, compared to the other model BERT. This latter one is able to predict or alternatively make new suggestions for words in the existing sentence. So, when it comes to creative content generation, GPT-2 is definitely the way to go.”
“However, I noticed that the seed text affects the result of a random generation and can be used to control the text outcome. For example, if your seed text contains informal language, the continuation will likely follow the same style.“ Luiza comments.
An image of a chimpanzee produced the best ghost images
For the image generation Silo.AI leveraged generative adversarial networks, often shortened GANs. GANs are famous for being behind AI generated image phenomena such as Deepfakes, and the automated face generation website This Person Does Not Exist. GANs excel at creating “authentic looking” content because the neural networks in the machine learning system compete with each other to get closer and closer to the wanted outcome. As a result, the generated content is completely new, but looks real, as if it existed before.
We used Nvidia’s state-of-the-art face generation model StyleGAN, which is model used in the above website This Person Does Not Exist. Although the model normally produces random images, with intensive computation, it can be guided to produce images that resemble a target image. In other words, the images can be tweaked to adapt a certain style.
Ari Heljakka, our AI Scientist behind the image generation part comments:
“What is special about these kinds of models is that we can start from a generated real-looking face and push it gradually towards another face type that we are interested in. In this case, we wanted the face to look like a real person, but to be somehow uncanny, like a ghost.
Eventually, what worked best came as a surprise to Ari: “I tried to merge many real photos of people from different ages and ethnicities and peculiar features, but eventually came up with the chimpanzee which to my amazement worked out the best. Although the results were good for this purpose, we are barely scratching the surface of what is possible with these models.”
Chat behind the grave – AI has storyteller potential
As a result of the pilot project, Yle learned more about what kind of stories and visuals can be created with AI solutions. The challenge with AI continues to be the training data, which affects the content the solution is able to create. The teaching phase determines what can come out of the solution.
While it is fascinating to create completely new stories with AI, getting useful additions to the existing Yle material requires a significant amount of work. The AI solution needs to produce the right kind of content, but also Yle then needs to find out ways to use these images and texts in the right context. Finding this forum, be it on the website or in a special marketing campaign, will be the next task to solve. For Yle these images and stories generated are available to use through API, which lets programmers to take them into products and services.
For both Yle and Silo.AI the project offered fascinating ways to test AI’s capabilities in the real world from the content generation perspective. This kind of exploration is crucial for the future of the media, which as an industry needs to be able to provide personalized content at an increasingly fast speed. As the ground-breaking advancements of the technologies presented in this article are very recent (2019), it is expected that we will see a growing number of advanced AI solutions helping media companies produce future-proof content.