Release Date: November 10, 1995
BUFFALO, N.Y. -- A software package that translates Chinese to English and comprehends what it's translating has been developed by a graduate student in the University at Buffalo Center for Cognitive Science.
The Chinese-to-English system translates and comprehends simple and complex sentences in seconds.
Besides significantly improving Chinese-to-English translations, the software, which communicates as humans do, using natural language, is helping cognitive scientists to explore in greater detail how humans use language to communicate.
It may have potential for use in conjunction with the WorldWideWeb, making Web-site information instantly available in numerous languages.
By understanding and responding in natural language, the system provides key advantages over machine translators that are commercially available.
"In the same way that a human translator has to understand what he or she is translating in order to do the best job, an electronic system also should understand what it is translating," said creator Min-Hung Liao, UB doctoral student in the departments of computer science and linguistics.
"If you asked me to translate something and you just gave me a dictionary and the synctactic rules of a foreign language, the results would not be satisfactory. That is how machine translators work now."
With Liao's system, the user types in the Chinese characters. The Chinese text then appears on-screen while the software package performs the translation, using artificial-intelligence techniques previously developed at UB.
The correct English translation then appears on-screen.
"The system can perform like an expert system as well," said Liao. "The following statement could be typed in Chinese, 'Mary ate bread and cake and apples.' Later on, I can query it in English: 'What did Mary eat?' It will respond in English that Mary ate three things: bread, cake and apples."
The "brains" behind the software package is SNePS, Semantic Network Processing System, a knowledge-representation system developed over the past several decades by Stuart Shapiro, Ph.D., professor of computer science at UB, and William Rapaport, Ph.D., UB associate professor of computer science. They incorporated SNePS into CASSIE (Cognitive Agent of the SNePS System -- an Intelligent Entity), a computerized cognitive agent that understands information and communicates using natural language.
Liao, who learned about SNePS in the "Natural Language Understanding" course taught by Rapaport, decided to apply this system to electronic translation.
"I'm making CASSIE multilingual," he said.
The system utilizes SNePS as an interlingua, the intermediate representation of meaning that both human and electronic "minds" experience when performing translation from one language into another.
While it is translating, the system's knowledge base grows and changes, so the translations are not always fixed, Liao explained.
"Because the system uses SNePS, it doesn't translate one sentence at a time the way other machine translators do," Rapaport noted.
Instead, it translates all of the material on-screen, transforms it into this interlingua and then translates it from the interlingua into English.
According to Liao, SNePS is much more powerful than other knowledge-representation systems.
With SNePS, CASSIE is able to perform reasoning tasks, make inferences and do belief revision, where it is told something and makes inferences based on that fact, and then corrects itself later if it obtains additional information that indicates that it was misled.
Because SNePS allows the programmer to build world knowledge into the translation system, it has the capacity to reason about what it is translating and the power with which to discern, based on context, the proper meaning of a word or phrase that may have multiple meanings.
This turned out to be particularly important with Chinese because Chinese has a freer word order than English.
While in English a sentence reads "Mary sent the book," the Chinese translation might read "Mary the book sent" and it would have the same meaning.
"Because it's a semantic network, SNePS can deduce the proper meaning," Liao said.