Imagine asking a group of people in a room to each write a poem titled “A Computer Scientist’s Meeting.” The possibilities for different combinations of verses, words and imagery are endless, and the poem could even end up looking like this:
A Computer Scientist’s Meeting
-Upon arrival, morning meeting start,
-Attendees telling jokes and people laugh,
-Intense like fire, minds exchanged to heart,
-My peers and colleagues shaking hands with staff.
-Unreal like dreams but it’s reality,
-A blessing goddess made the wise and good,
-Machines had flourished with humanity,
-Computers wondered how a creature should.
-Diligence was my terminology,
-As long as several breakthroughs got around,
-Our dream advanced with new technology,
-With best solution we have ever found.
-A meeting moment was a moment’s spent,
-Scientists outpour research so that they invent.
However, this poem wasn’t written by a human being. In fact, it was created entirely by an artificial intelligence algorithm developed at the UCLA Samueli School of Engineering by Nanyun (Violet) Peng, an assistant professor of computer science, and her Ph.D. student Yufei Tian.
Given only the title of the poem, the AI was able to string together a sonnet that could have plausibly been written by a living, breathing person.
“I have genuine interests in language and wordplays,” Peng said. “I also like to work on fun problems. I choose topics like sonnets, hyperbole and puns because they are all fun problems and can be hard for humans, and I think it would be cool if computers can do them too. It also brings a lot of joy when I look at the pieces that our model generated.”
The goal of natural language generation is to train AI to produce written content that humans can comprehend. Peng also does work on natural language understanding, which aims to give AI the ability to understand human language.
Peng’s journey began with natural language processing early in her undergraduate career. While completing a double major in economics and computational linguistics at Peking University in China, she developed a passion for computation and statistics and ended up staying on at the university to get her master’s degree in computer science with a specialization in natural language processing.
She then traveled all the way to Maryland to get her Ph.D. at Johns Hopkins University, working at its Center for Language and Speech Processing. Peng’s doctoral thesis focused on information extraction as she found ways to make it easier for humans to get the information they needed out of a given text, which she describes as being close to the opposite of natural language generation.
Thanks to a Defense Advanced Research Projects Agency (DARPA) program called Communication with Computers, Peng pivoted her research focus to natural language generation.
“The goal of the DARPA program is to have the computer communicate with humans naturally and to accomplish some tasks that either party alone (human or computer) would do much worse on,” Peng said. “We decided to focus on creative composition which is even challenging for humans as it is easy to hit ‘writer’s block.’ The composition tasks particularly interested me because they are highly interactive and intuitive.”
Given the highly subjective nature of writing, natural language generation poses unique challenges because there is not any one answer or output an AI can create that is necessarily correct or incorrect. Because of this, Peng and her team must consider not only the mechanics of writing — grammar, coherence, sentence structure, etc. — but also creativity, fairness and whether a piece is actually interesting or compelling overall.
To make her work even more of a challenge, Peng said she has chosen extremely creative formats for experimenting with natural language generation, including sonnets, hyperboles and puns.
Peng’s technologies could possibly be used in a variety of fields that require writing skills in the future.
“It has great potential for the entertainment industry — stories, movies, advertisement — especially as a writing assistance system to be used in collaboration with human writers to break through writer’s block,” she said. “It could also potentially be used for news reports.”
Looking ahead, Peng says she hopes to work on natural language generation technology that can write longer pieces, such as stories and papers, while maintaining long-term coherence. She’s also working on improving common sense in the algorithms and addressing offensive or biased content, which can be difficult to automatically detect and prevent.
“My hope is that natural language generation can ultimately increase creativity, especially in human-computer collaboration settings,” Peng said.
Sara Hubbard contributed to this story.