As one of the leading experts behind the THEaiTRE project, you are developing a generative system for theater plays with the help of artificial intelligence. What does such a creative process look like?
It is based on the neural language model GPT-2 from the OpenAI consortium, which uses about 8 million English¨language articles scraped from the Internet. When we enter the beginning of the text into this model, it tries to generate a continuation which corresponds to the part that is already written. During our project, we insert the beginning of the scene into the model, which usually means we supply a description of the initial situation and the first replica of the individual characters. The model then generates additional replicas. However, the human operator does have the opportunity to intervene in the process. They can discard the generated text in case they do not like it and prompt the model to create an alternate sequel. As the model only works in English, the text is then automatically translated into Czech by our CUBBITT translator.
Artificial intelligence encounters is still not perfect at processing text: it is often described as indecent and potentially discriminatory. Have you experienced similar problems?
There are many unresolved issues in this arena. Language models are created on the basis of all kinds of texts, which usually cannot be manually browsed, checked and filtered due to the sheer amount of text required. Therefore, models indirectly learn all sorts of human ills – racism, sexism and so on – from texts written by people. Such shortcomings are rightly pointed out even in our first play. In addition, researchers can select AI’s better attempts to be performed, as opposed to the more abrasive ones. However, our goal was not to embellish the outputs of the model too much: we aimed to show what the automatically generated outputs really look like today. That is why we have preserved various imperfections in the text.
What procedures besides the intervention of human operators can be used to eliminate errors?
Many researchers are actively working on eliminating prejudices or biases in language models. We usually talk about so-called debiasing, as in getting rid of the model of various biases or prejudices. Let’s say the model exhibits bias when it comes to women and men – in the AI generated text, the boss could automatically be a man and the assistant always a woman. In such a case, for example, training data can be pre-processed and men and women randomly shuffled to improve gender balance. In general, however, it can be said that the model only reproduces pre-existing patterns in training data. Therefore, the most common way to correct outputs is data modification. But that remains difficult to do – partly because we can’t properly define and separate exactly what’s right and what’s not. For example, both a mother and a father can prepare breakfast for schoolchildren, but a mother will be more likely to breastfeed infants. In short, information about the difference between men and women cannot simply be omitted, since that would only cause other problems.
What makes AI better than humans?
Conversely, if all is done well, AI can be more fair than humans often are. This means making decisions based on relevant information only and not being influenced by irrelevant external factors..
Where did the idea of employing artificial intelligence on stage originate? Do any similar projects exist abroad?
The main impetus was the 100th anniversary of the first performance of R.U.R. by Karel Čapek, the play which coined the word robot. Our project was the first to present a play that lasts almost an hour, with 90% of the script written by artificial intelligence. Previous projects have either created much shorter works, such as the film Sunspring, or the percentage of the work generated by artificial intelligence is significantly smaller. This category includes, for example, the British musical Beyond the Fence from 2016 or the Australian play Lifestyle of Richard and Family, which premiered two years later. The Improbotics group has a different approach to working with AI, since they use replicas generation live and directly on stage. One character gets replicas fed into their headphones and the other character improvises accordingly. However, all previous projects known to us were in English. Our creation is therefore also the first automatically generated theater play using the Czech language. However, the collection of poems Poetry of the Artificial World has worked with artificial intelligence in our country before.
What was the audience reaction to the first play written by a robot, presented by the Švanda Theater? Last year’s online premiere was watched by 13,500 spectators, if I am not mistaken?
The reactions were mixed. Some people were thrilled with the play, but some also dismissed it as complete nonsense. In general, according to the reviews, our play is comparable to a worse play written by a relatively weak human playwright. We consider the fact that an automatically generated play can even be judged against plays written by real people to be a great success and proof of the significant progress AI has made in recent years. Spectators often note that the play seems “absurd” or “difficult to understand”. At the same time, when performed, the audience often laughs at certain scenes, which we take as proof that the model has managed to generate various funny replicas – although it will probably be more by chance than on purpose.
How did you, as one of the principal creators, evaluate the final play?
If we return to the question that our project asks itself, whether artificial intelligence can write a play – we haven’t quite proven that yet. However, before watching the play, many were convinced that such a possibility was very far away or that it would never happen. Our project has convinced many people that plays authored by AI are a very real possibility. However, there are fundamental problems for which we do not have the answer. In particular, it should be noted that the machine, unlike a human, has no inner need to express something, it only generates text. In this sense, it resembles a human author who creates soulless works solely for profit, without higher artistic ambitions. It is quite possible that a machine simply cannot create deeply meaningful art. It can master technique, practice craftsmanship, and deliver formally perfect results. But the works will have no pivotal message, no higher artistic value.
So should we not expect a flood of plays generated by artificial intelligence in the future?
In general, artificial intelligence is good for tedious, repetitive activities: after all, robotic automation is primarily deployed for industrial tasks. In art, on the other hand, we look for novelty and originality. In a sense, a work of art has value only if it is new and not just an imitation of something we have seen a hundred times. And that’s not exactly what artificial intelligence is well-suited for. I believe that in a few years, AI will be able to generate a theatrical play that will not suffer from the same shortcomings that our play suffers from. Such a future work will be formally high quality, but not very artistically valuable. I can imagine the use of automatic generation, for example, for writing the scripts of endless television series, where it is apparent that the same thing happens over and over again and there are no high artistic demands. We would enjoy trying this, but we have not yet found a suitable collaboration to make it happen.
“In general, artificial intelligence is good for tedious, repetitive activities: after all, robotic automation is primarily deployed for industrial tasks. In art, on the other hand, we strive for novelty and originality.”
Are you planning other related projects?
We have recently prepared a project proposal together with the ASCR, proposing neural modeling of Czech poetry with the goal to not only generate poems, but mainly to learn about existing poetry. The model can observe regularities, irregularities, and patterns. It is difficult for it to create a new interesting poem, but we believe that it could well generate a new poem typical of a given author or epoch – it could somehow average and distill the essence of Czech romanticism, for example. In connection with current topics, other possibilities open up as well. We could generate, for example, a text Mácha would write about global warming.
The idea of a project with the Prague City Library and the Czech writer Julia Nováková is in its infancy. We would like to create a tool to help beginning writers. These authors are still learning the craft, so it might help if AI alerts them to various formal shortcomings in their work. Artificial intelligence could suggest how to improve, for example, a short story in progress – in short, it could recommend what to add so that the story resembles other successful texts. Of course, there are limits to the extent this should be used so that there is still enough space for the author’s own creativity and individuality.
The script of your theater play was created in English and automatically translated into Czech using the CUBBITT tool, in the development of which you participate. How is CUBBITT different from other automatic translators?
All translators available today function similarly. They utilise a deep neural network, usually of the Transformer type, trained on a large number of bilingual texts. Both DeepL and Google speak a lot of languages, which is their strength. CUBBITT only speaks a few languages, but we were able to focus on these languages in more detail during development. According to our measurements, it is still true that CUBBITT achieves higher quality in translation between English and Czech than competing translators. How exactly CUBBITT differs is difficult to say, because while CUBBITT is open-source and its operation is described in detail in scientific articles, such information is not available when it comes to commercial translators. We assume that the neural network architecture is the same for all models, but in our case Martin Popel, the main developer of CUBBITT, could simply afford to put more work into tuning individual parameters specifically for Czech.
Among other things, the CUBBITT translator helps people fleeing the war in Ukraine to overcome the language barrier. What was the most difficult part in the development of the Ukrainian-Czech version?
It was crucial for Ukrainian to get enough bilingual texts, but fortunately we found a way to create the translator from scratch in a few weeks. It differs from the Czech-English translator mainly in the training data, the architecture of the model is the same and the parameters are also set similarly. The data properties do result in some imperfections; for example, it is difficult to translate the names of smaller Czech or Ukrainian cities, because they occur little or not at all in the training texts. Therefore, the translator sometimes “translates” a Czech city as a Ukrainian city – for example, Jihlava has long been translated as Zhytomyr for us. But colleagues are working hard on this, and the model is constantly improving. As with English, CUBBITT has a higher translation quality than the competition for translation between Czech and Ukrainian.
In addition to robopsychology and research into artificial neural networks, you are also involved in popularizing science. How do you manage to motivate young people to study artificial intelligence?
I’ve been trying to popularize science for a long time, but the outcomes of these efforts are hard to measure. What I try to convey to people is the enthusiasm we devote to science. For example, I often participate in open days, we give various lectures for high schoolers and the like. We hope that this will perhaps attract someone to science and our field – or at least not discourage anyone.
Can you think of other ways to arouse interest in artificial intelligence among the general public?
My colleague Tereza Hannemann and I founded the AI In Context group, where we try to look at artificial intelligence in various fields of human activity. As part of this project, we organize lectures, prepare an interactive exhibit for the Hybernská Campus, and also take care of the Elements of AI course at Charles University, which is an online artificial intelligence course for everyone. We have other plans and ideas for the future, too. Together with the Svijany brewery, for example, we are trying to come up with an AI pub where guests could try out various new technologies and learn something about how they work. In this way, we would like to target the part of the public that most of our communication tends to miss.
Next up from prg.ai
Typical Prague AI firm is young, self-sufficient, and export oriented, shows our new comprehensive study
130 companies, 11 interviews, 9 business topics. Explore all that and more in the unique study authored by prg.ai, which contains an overview of last year's most notable events on the local AI scene or articles on the future of AI or gender equality in research.
prg.ai newsletter #39
What did the first month of 2024 bring, and what can you look forward to in February? Find out in the next prg.ai newsletter. Check out what's new on the artificial intelligence scene (not only) in Prague.
prg.ai newsletter #38
Last year was dominated by the AI boom. The latest issue of prg.ai newsletter provides a summary of key developments in the artificial intelligence scene, not only in Prague, as the year came to a close. Discover hot news and interesting facts as we step into the new year together on the right (🦵/ 🦿) foot.
We are hiring: Program Manager!
We are strengthening our team and looking for someone to take charge of our educational activities.