The Computer is Spinning Tales
The notion that one day in the future computers will be able to comprehend and even make up stories has inspired and guided curious humans, take scientific researchers, programmers and sci-fi writers as the best examples of that, for many years. We may be tempted to ask, could artificial intelligence one day be able to read and understand the merits of a good story?
Now an answer might be within reach. This very complex problem is being tackled head-on by scientists from the University of Massachusetts - Boston and Disney Research. They have created neural networks that can evaluate short narratives. These artificial intelligence units do not try to mimic the judgment of literary critics, but instead attempt to predict which stories appeal to large audiences. The team presented its findings on August 23 at the International Joint Conference on Artificial Intelligence in Melbourne, Australia.
Prof. Boyang “Albert” Li from Disney Research said that their neural networks have had some success in predicting the popularity of stories. While they can guide future research, according to him, they cannot yet be used to choose winners for local writing competitions.
“The ability to predict narrative quality impacts on both story creation and story understanding,” said Markus Gross, vice president at Disney Research. “To evaluate quality, the AI needs some level of understanding of the text. And if AIs are to create narratives, they need to be able to judge the quality of what they are producing.”
In order to develop models for automated evaluation of story quality, we will need large databases of stories that have already been evaluated by humans, so that those can be used to train artificial intelligence. That is also the opinion of Tong Wang, a computer science Ph.D. student at UMass - Boston and an associate at Disney Research. The researchers found that Quora (the famous questions-and-answers search engine) is a good data source, since many of the answers provided are in the form of stories. The number of upvotes and downvotes are also useful in measuring how popular a story is and can also serve as a standard to measure narrative quality. His work was also assisted by Ping Chen, PhD adviser and associate professor of computer science and engineering at UMass Boston.
So the researchers went ahead and gathered roughly 55,000 answers and developed an algorithm to classify them either as stories or non-stories. That first filter yielded around 28,000 stories with an average of 369 words each. Then, in order to understand the subjectivities and complexity of the stories, the team looked for manners to represent semantically the influence of story structures in the neural networks, since a sequence of events can interact to reveal character intentions, and therefore, outcomes.
Then the researchers created an artificial intelligence model that evaluated separate regions of each story, including the questions that prompted each. A network was created to look at the regions interdependently; another network still was created to take a holistic view, in order to establish how the meaning of the events and story regions emerged from the entire story. In each case, the artificial intelligence models made predictions as to which texts would be considered the most popular with readers.
So far, the experiments showed that the neural networks experienced an improvement over a baseline text evaluation system, while the holistic network registered an improvement of 18 percent.
And here’s the best part of the story. The part where you are both the author and the hero. With our experience in artificial intelligence, data gathering and analysis, we at Tenex are ready and available to boost your business objectives. Give us a call or send us an email and we will be glad to demonstrate how we can improve your outcomes with our elegant, scalable and thorough solutions.