It's obvious on brief reflection why this should be so --- language, more than any other human ability, distinguishes our species from the rest of the animal world, by enabling us to share and record our thoughts, plans, and feelings. It is no surprise, then, that our vision of truly intelligent computers invariably endows them with the basic linguistic competence that all human beings share.

What perhaps is surprising is that this vision remains such an elusive goal. Unlike our other intellectual abilities -- proficiency in calculus, or the mastery of chess, for example -- which require our conscious study to attain, the acquisition of our own native language is automatic and nearly effortless, occurring at an age that precedes most of our mathematical and reasoning capabilities. But although children learn to speak a language with no formal education, and adults use language with little conscious difficulty, how this is accomplished remains an enigma. One of the goals of natural language processing research is to create precise computational theories of language that can answer this question and move us toward our vision of a fluent computer system.

Suzanne Stevenson, one of our Computer Science faculty members, who has a joint appointment in the Rutgers Center for Cognitive Science, is working on building computer systems that can attain human-like performance in understanding English sentences. Motivated by her background in the cognitive sciences, Suzanne has explored the capabilities of connectionist models -- also known as artificial neural networks -- for capturing fundamental aspects of language understanding. Because of their inspiration from brain-like processing mechanisms, connectionist models have been touted as the key to the realization of intelligent behavior in computers, but their deficiencies in representing complex symbolic information have limited their effective use in language processing. Suzanne's research on connectionist models has primarily focused on two challenges raised by the goal of a fully conversant computational system.

First of all, we must determine how to represent within a computer system what a person unconsciously knows about her language. Modern linguistic theories posit rich hierarchical representations of sentences, which define an intricate mapping between the string of words that comprise a sentence, and its underlying meaning. Computational linguists must adapt these linguistic formalisms to data structures that can capture the relevant information, and algorithms that can efficiently compute the structures and determine the possible meanings of a sentence. Current connectionist models have been inadequate to these representational tasks. Suzanne's research has addressed these deficiencies through the development of distributed processing techniques for incorporating structured linguistic knowledge into a connectionist network model. The second challenge for a language understanding system is to decide which of the potential meanings of a sentence is the one intended by the writer or speaker. Communication in human language is replete with ambiguous utterances -- that is, sentences with more than one meaning. Yet people efficiently focus on a single interpretation of what they hear or read, and are rarely conscious of the alternative meanings. For example, on hearing the sentence "Spock condoned illegal alien arrests," most people will understand that Mr. Spock condoned the arrest of illegal aliens, not the illegal arrest of little green men. An effective computer model of language understanding must also choose a single preferred meaning of an utterance, and to be accurate, the one it chooses must match human preferences. Within the same connectionist framework mentioned above, Suzanne has devised simple computational techniques for choosing preferred interpretations of sentences, and has demonstrated the accuracy of the model in mimicking human preferences that have been observed through psychological experimentation.

In responding to these two challenges, the goal is to build a computer system that can adequately represent what people know about the grammatical structures of their language, and encode the principles that allow people to resolve ambiguities effectively. One of the exciting aspects of this research is the opportunity for substantive multidisciplinary collaboration. Suzanne has worked with both linguists and psychologists to explore novel computational representations and algorithms that can bring us one step closer to our vision of natural communication with a computer.

Open House

Conversing

Data Mining

Mobile Wireless Computing

Women in CS

Remote Learning

PhD Recipients

Recent Awards

DCS Web Page