Computer tells convincing story

By Chhavi Sachdev, Technology Research News

Although text generators have been around for decades, there’s a good reason why a machine has never won a Pulitzer. The text tends to be choppy and simplistic.

A pair of researchers at North Carolina State University has developed software that produces more sophisticated prose by combining the rules of natural language generation with artificial intelligence research on story generation. The result is output that comes closer to the free-flowing speech of HAL than the subject-verb-object utterances of E.T.

Given a “logic-based representation of characters, props, actions, and descriptions, [it can] convert them into prose that looks just like the text you would get if you bought a book off the shelf at a bookstore,” said Charles Callaway, now a research scientist at the center for scientific and technological research at the Cultural Institute of Trentino in Italy.

The Storybook software does this by using a narrative plan, which is a logical representation of the characters and actions in a story, Callaway said.

The researchers borrowed the concept of a narrative plan from the Russian Formalist school of literary criticism, which explains a story in terms of fabula and syuzhet. The fabula is the sum of events. It is the synopsis of a film or a book you would tell a friend. The syuzhet is the plot or the order in which the fabula unfolds. The narrative plan charts the who, what, when, where, why, and how of a story along with the order of events.

Storybook’s prose-generation architecture first constructs short sentences and paragraphs. It then looks for prior references, choosing synonyms for previously used words, and matches actors and events in the story with semantics and syntax. Last, it reorders the text to eliminate short choppy sentences and formats it.

The result is text composed of average-sized sentences with a wide vocabulary that includes pronouns, adverbs, and dialogue. It is longer, more variable, higher quality prose than previous systems have made, Callaway said. “Only one other computer system has produced text longer than two paragraphs, and it was in a scientific domain,” he said.

The system could generate conversation in role-playing games, said Callaway. It could also be used to create animated teaching agents in tutoring systems. “If … a young child could create the characters and plot of a story by dragging around icons on the screen with a mouse, then the system could generate a story corresponding to their selections,” he said. The child could then alter the elements to read a subtly different story each time. “Researchers in tutoring systems believe that this type of self-motivation will result in children recreating the story [many] times,” Callaway said.

The work is a nice piece of research that “shows how natural language generation (NLG) technology can be used to enhance the quality of story generators,” said Ehud Reiter, a lecturer in computing science at the University of Aberdeen in Scotland. “This is the sort of thing that people in the past have vaguely thought about, but I believe this is the first serious attempt” to integrate natural language generation and story generation, he said.

“Story generation and natural-language generation … have in fact been quite separate strands of research, with people in [the former] focusing on high-level content issues … and ignoring how well it reads at the sentence level; and people in [the latter] focusing on how well it reads at a low level, but in non-fiction applications such as generating customer-service letters and weather reports,” said Reiter.

“I was also pleased that the authors … made an attempt to evaluate their stories experimentally, which, as far as I know, has not been done in previous research on story generation,” said Reiter.

One barrier to practical use, however, is that the system requires a considerable amount of knowledge to be encoded in knowledge bases and a finite-state narrative model, Reiter said. In other words, “setting up the system to produce a story [requires] a lot of time specifying in computer-friendly form information about the structure of the story and the world the story takes place in,” he said.

In addition, the work may be “more interesting scientifically than practically, since there is no shortage of human authors who are willing to write stories,” he added. One possible application is incorporating messages like ‘smoking is bad for you’ in the stories, he said.

The system could be in practical use within three years, Callaway said. “There still is a lot of work to be done on the front end … but I'm confident that I won't be the only person working on it. There are a few topics in [artificial intelligence research] that draw everybody's attention, like robots and speech recognition, and I think story generation is one of those areas,” he said.

The next step is to speed up behind-the-scenes processes. “It took me four months to write the input logic for the two-page story shown in the presented paper, although it takes the system 45 seconds to turn that into text,” said Callaway. He is also looking into prose generation in Spanish or Italian, he said.

Callaway’s research colleague was James C. Lester at North Carolina State University. The research was funded by the university. They presented the paper at the Seventeenth International Joint Conference on Artificial Intelligence held from August 6 to August 10 in Seattle, Washington.

Timeline:  Now
Funding:  University
TRN Categories:  Natural Language Processing; Artificial Intelligence
Story Type:   News
Related Elements:  Technical paper, "Narrative Prose Generation," Physical Review E, presented at the Seventeenth International Joint Conference on Artificial Intelligence, in Seattle, Washington, August 7, 2001.


October 10, 2001

Page One

Tiny tubes make logic circuits

Mobile radios make intranet

Quantum code splits secrets

Computer tells convincing story

Virtual beings boost evolutionary theory


Research News Roundup
Research Watch blog

View from the High Ground Q&A
How It Works

RSS Feeds:
News  | Blog  | Books 

Ad links:
Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Offline Publications     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds
Forum    Comments    Feedback     About TRN

© Copyright Technology Research News, LLC 2000-2006. All rights reserved.