Highlight
Evidence for knowledge of syntactic rules without linguistic experience
Smolensky-IGERT Figure 1
Achievement/Results
Recent research conducted at Johns Hopkins University has produced some of the first evidence from controlled experiments that confirms a controversial hypothesis made famous by Noam Chomsky a half-century ago: that the human brain holds unconscious knowledge of syntactic rules that have never been experienced, knowledge which is exploited to facilitate language learning. The breakthrough was made possible by bringing together recent advances in experimental methods for studying learning of controlled languages, new theoretical methods for modeling human learning, and classic empirical findings in linguistics.
This powerful interdisciplinary approach was developed by Jennifer Culbertson (Figure 1) under the auspices of a major National Science Foundation (NSF) training grant, “Unifying the Science of Language”, housed in Hopkins’ Cognitive Science Department, home to the grant’s Principal Investigator, Paul Smolensky. This training grant, under NSF’s Integrated Graduate Education and Research Training (IGERT) program, made possible the development of an integrated interdisciplinary curriculum training all Ph.D. students in mathematical theories of language and learning, the latest techniques from experimental psychology for studying the mind/brain through behavioral and neural research, and the methods of linguistic analysis needed to precisely characterize the general patterns in the world’s 6000 human languages. This training program initiated a new paradigm for graduate training, “problem-centered training”, which ignores the boundaries imposed by traditional discipline-centered training, and constructs a special program training scientists to address a single problem area—in this case, human language—by integrating methods from all the disciplines that have traditionally studied the problem, largely in isolation from one another.
The program has demonstrated that it is possible to train a new type of scientist, one that is defined not by a discipline but by a multidisciplinary toolbox for solving a single problem. Jennifer Culbertson is herself proof: her research results derive from a new behavioral paradigm that she developed, building on current work in experimental psychology; a new mathematical model for the mental systems driving the observed behavior; and a broad cross-linguistic perspective on the key patterns that come from linguistic analysis of hundreds of languages. Prior to entering the program, Ms. Culbertson, like most of her peers, had been trained in only one of the cognitive disciplines, linguistics; the experimental and mathematical components that are central to her research are made possible entirely by the problem-centered training fostered by the IGERT program.
In Ms. Culbertson’s, experiments, participants learn a carefully-designed artificial nano-language called Verblog. They learn from an alien informant Glermi who they interact with through a video game interface. In one experiment, for example, they might see a display like that of Figure 2, and describe it by saying “geej slergena”, an appropriate Verblog phrase; (“geej” means blue, and “slergena” names a type of alien object).
Ms. Culbertson knows, as her experimental participants do not, that many human languages use this word order, Noun preceded by Adjective, and many use the order Noun followed by Numeral (“slergena glawb” in Verblog means ‘three slergena’)—but very rarely are these two orders combined in one and the same human language. Is this because the human brain has encoded within it grammatical knowledge that bans this combination of orders? If so, there is something unnatural about Verblog for human language learners, and they should learn it with some difficulty—relative to other participants whose task is to learn other languages matched to Verblog in every way except that their word order rules match patterns commonly found in human languages.
And Ms. Culbertson’s experiments show that indeed, human learners do have a special difficulty learning Verblog. These adult English-speaking learners have little or no exposure to languages with orders different from English (Adjective-Noun: ‘large deer’; Numeral-Noun: ‘three deer’). Yet they learn alternative, but cross-linguistically common, word orders without difficulty. It is only on Verblog that learners fail.
This is exactly what is predicted by the Chomskian hypothesis that, despite experience limited to English orders, the learner’s brain knows that the Verblog order-combination is extremely unlikely. This hypothesis explains both the experimental results and the exceptional rarity of actual human languages with Verblog-like word orders.
Extracting this unconscious knowledge hidden within learners required development of a new experimental paradigm where learners are exposed, in the training they get from Glermi, to a particular mixture of word orders. When exposed to such input, learners improve the consistency of the language by shifting the mixture to more strongly favor the dominant word orders—except when the mixture contains mostly Verblog orders. Ms. Culbertson’s Mixture-Shift Paradigm has proved sensitive enough to reveal the brain’s knowledge of the alienness of Verblog.
Formally, this means that human learners have a bias against the Verblog orders (as well as a bias against inconsistent use of orders). Ms. Culbertson developed a mathematical model in which learners deploy Bayesian probabilistic inference to learn a probabilistic model of the artificial language, a model which they then use to generate their own utterances. Because of the Bayesian prior that encodes the biases, learners exposed to Verblog will not acquire a model of their language that corresponds to the models acquired by learners of the other languages.
The source of this knowledge—the bias—is at this point mysterious: it is the subject of Ms. Culbertson’s future research.
Address Goals
It is widely recognized that interdisciplinary research is playing an increasingly important role in science. Cognitive Science—the study of the mind/brain—has been at the forefront of development of interdisciplinary research, and an area in which the US has played a leading role. However, as other countries have built up strength in this area, they have often been able to catch up quickly, as the US seems to be hampered by disciplinary-boundedness to a greater degree than many other countries. The strong emphasis on specialization in US graduate education has perhaps been important to American post-war success in building a world-class research establishment, but as we move into an era dominated by interdisciplinary research, this same focus on specialization is a clear impediment to training the next generation of scientists. Problem-Centered Training—pioneered by the Johns Hopkins Interdisciplinary Graduate and Research Training (IGERT) program in the cognitive science of language—is, we believe, potentially a more appropriate structure for the future of American research, allowing the US to maintain its position as a world leader in fundamental science. Certainly we feel that the success of the trainees in our IGERT program is proof of the effectiveness of the concept in this important area of science.
For solving key societal problems, it is even more evident than in basic research that Problem-Centered Training is the appropriate model. Most of these problems are multifaceted and demand a highly multidisciplinary approach to solutions. Two problems of special importance in this information age are understanding human learning and understanding human language, in order to better serve the societal needs of retraining during the lifespan and of extracting information from the enormous quantities of written and spoken language that are now available. The Hopkins IGERT program, ‘Unifying the Science of Language’ (Paul Smolensky, Principal Investigator), is training a new breed of scientist in which interdisciplinary approaches are foundational. The payoff in advancing our fundamental understanding of human learning and language is already evident in the research advances made by trainees during their time in graduate school. Trainee Jennifer Culbertson’s work, part of which is discussed above, is an excellent example.
