Research Statement
The mind can perform a myriad of amazingly complex tasks effortlessly and flawlessly. This is perhaps the most important starting point for any psychological research. My own research interests are in what underlies adults' abilities to perform highly complex tasks and how these abilities develop in infants and children. Complex abilities are so interesting because they depend on and arise out of an interaction between sets of simple and general cognitive tools - for example, tracking the frequency of events, forming abstract categories, comparing expectations and actual outcomes. These same basic tools are combined in various ways to produce the behaviors of several higher-level systems (e.g. navigation, language, social interactions). On this view, domain-specific systems such as language are the outcome of the interactions of a number of domain-general tools. My interest is in discovering the nature of these simpler underlying tools (or basic capacities) in the mind and brain, and how they interact to produce such amazing emergent abilities. These capacities do not necessarily come entirely built in through genetic endowment; years of experience shape these fundamental systems as well as how they relate to each other in various domains. Development of complex mental capacities depends on a dynamic integration of innately specified biases and environmental nurturing. Understanding the nature of these biases, as well as the information that can be acquired through learning, are critical components of a comprehensive account of complex psychological skills.
My graduate research has been focused on describing the cognitive subsystems underlying just one central complex ability, that of hierarchical representations. Hierarchies are used in cognitive systems across a wide variety of domains. Some of the most significant early work on hierarchical organization comes from Lashley (1951), who detailed how hierarchies are necessary to properly characterize planned action sequences. These include typing the letters of a word in the proper order, speaking the words in a sentence in order, and also performing most complex motor actions (e.g. shaking hands) and goal-directed activities (e.g. making a sandwich). Hierarchies are also critical for our understanding of how the world works. Causal relationships are often best understood as components of larger hierarchically organized relationships. For instance, in billiards, one ball will strike another, causing the second to begin moving and changing the direction of the first. These individual actions make little sense on their own, but put within the larger hierarchical context of trying to sink a ball in a pocket and set up a good next shot, they become natural parts of a larger whole. Hierarchical structures also play a central role in models of memory (e.g. schemas) and categorization (e.g. a greyhound is a dog, dogs are mammals, mammals are living things). Clearly, hierarchical organization is a central characteristic of many cognitive domains.
While many mental systems seem to involve hierarchically organized representations, some of these systems must produce linearly sequenced behaviors as their outputs. The interaction between a hierarchical structure and a linear string is often quite complex, as linear sequences can be thought of as having a smaller bandwidth than hierarchical structures. This is the problem of pushing a two-dimensional signal (hierarchical structure) through a one-dimensional channel (linear sequence). The result is a mapping problem: how can a person learn to map a linear sequence, with only simple linear relationships between its elements, to a hierarchical structure with multiple levels of relationships among its elements? Perhaps the most significant system in which this problem regularly arises is language.
Language is used to convey ideas that have quite complex hierarchical relationships. In the sentence "John walked to the large park by the river" there are simple relationships such as between "the" and "river", larger relationships such as that between "the large park" and "by the river", and very large relationships stretching across the whole sentence between "John", "walked", and "to the park by the river". In linguistics this is called phrase structure - a representation of the hierarchical relationships between words in a sentence. When a person hears such a sentence, they understand its meaning through understanding these hierarchical relationships, and when they speak such a sentence, they order their words to represent this hierarchical organization. This raises a number of questions. First, how do people develop the ability to map from a linear sequence of words to a hierarchical phrase structure? Second, what biases and limits do people come to these tasks with? Third, what sorts of cues, either from within the linear sequence or accompanying it, do people use to learn the hierarchical structures?
My dissertation work has sought to answer all three of these questions through a series of integrated miniature language learning studies. Upon completion of my dissertation this winter, I plan to submit this entire set of interlocking studies for publication as a pair of journal articles. Because of the nature of these studies, publishing any one independently would have far less impact, as it is the differences between the studies and their results that is most interesting. My work so far is both narrow in scope and broad in implication. Its conclusions are not just important for psycholinguistics, but also for theoretical linguistics (which makes use of phrase structures), for language development (which involves hierarchical learning), for general cognitive theories (which involve other types of hierarchical structures that are likely to be similarly learned), for general developmental theories (which likely rest on the same learning and interactions), for clinical populaces (with language deficits), and for evolutionary theories (which seek to explain what has evolved in humans that allows us to control complex language while our other primate cousins are unable).
This work, done with Professor Elissa L. Newport, utilizes miniature artificial languages to study how learners acquire hierarchical mental representations from linear sequences. These languages allow the experimenter to have complete control over every aspect of the input to the learner, a situation not achievable through using natural languages. By carefully contrasting different types of miniature languages, it is possible to uncover the cues and mechanisms which learners utilize to acquire a phrase structure. We present a few hundred sentences from these languages to adults and watch their learning over a period of approximately 45 minutes. Accompanying the sentences are visual scenes that represent what the sentences describe. Periodically during the experiment we test participants to measure what aspects of the language they have learned, focusing mostly on finding out which groups of words participants are starting to treat as phrases and what knowledge the participants have developed of the correct linear word order of the language. Under some specific circumstances, participants show incredibly rapid learning of some rather complex features of these languages.
One potential cue that is always present in linear sequences is the degree of statistical co-occurrence between neighboring items (that is, their mutual predictiveness). In previous work in the Newport lab it has been shown that statistical regularities between syllables are used by adults and infants to segment word-like units from a continuous speech stream (Saffran, Aslin & Newport. 1996). Looking at sequences of words, rather than syllables, it appears that similar statistical regularities hold in natural languages. Most importantly, many words frequently occur next to other words that are in the same phrase as them. For instance, "the" is often followed by a noun, and together these two words form a noun phrase. My first studies set out to discover whether mutual predictiveness between words in a miniature language would induce learners to impose a hierarchical phrase structure. These studies revealed that whenever such statistical information was available to learners, they made use of it to acquire a basic phrase structure. A series of studies was performed to determine if this use of statistical information was governed by any internal biases in the learner to prefer certain kinds of hierarchical structures over others. According to many current linguistic theories, only certain kinds of hierarchical structures appear in natural languages, but it is an open question whether people actually have an innate preference for such structures, or whether these limits are simply theory-internal constructs. My studies have found that miniature language learners show surprising flexibility in the kinds of phrase structure they can learn. As long as mutual predictiveness relationships were present to support a particular phrase structure, the learners would acquire it. This included phrase structures involving pairs of words and sets of three words combining at once to form a phrase. When statistical relationships were more ambiguous or absent, learners did not impose a phrase structure on these sequences or form phrase preferences. Thus it appears that learners come to the task of language acquisition with very flexible skills and rely on the information available to build simple phrase structures. These studies indicate that while phrase structures are a very specific instance of hierarchically organized mental structures, learners rely on a universally available kind of information to build these structures. The results suggest that a domain-general learning system is being applied to a domain-specific problem.
Although statistical information has proven quite powerful for learning structure from miniature languages, it has severe limitations as well. One of the greatest weaknesses of local mutual predictiveness is that it cannot handle sequences in which related items are not immediately adjacent. However, complex hierarchical structures in natural languages create just such nonlinearities on a regular basis. The sentence "The man in the hat ate the sandwich" has a hierarchical relationship between "man" and "ate", but they are separated by the phrase "in the hat" which modifies "man". Fortunately for natural language learners, statistical relationships are not the only cue to phrase structure. Semantic dependencies -- the fact that the meaning of a word depends on the presence of other words ("eat" requires someone to do the eating) -- are another potential cue to hierarchical structure.
To explore the importance of semantic dependencies and how they interact with statistical information, a new set of miniature languages was constructed. These languages varied on two dimensions: the amount of local statistical information present among the words (large or small) and the type of visual world that accompanied the sentences. There were two forms of visual worlds, one involving clear semantic dependencies, and one where each word picked out an individual object, but without a meaningful dependency on any other word. This produced four different languages; individual learners were exposed to one of these four languages. The conclusions from this study are straightforward: When either local statistical regularities or strong semantic dependencies were present, learning of phrase structure was quite good. Without statistics or semantics, however, learning was extremely poor. Finally, when both cues were available and correlated on the structures they cued, learning was at its highest levels. This indicates that language learners do not simply rely on a single cue to build hierarchical structures from linear sequences of words, but instead rely on what information is available, balancing different cues and combining them when possible. Learning phrase structures (or any hierarchical representation) is a complex and difficult task, so it should come as no surprise that the mind uses a complicated solution to unravel this problem.
This work has looked at one aspect of hierarchical representation and explored the mental components underlying it, the role learning plays, and the interaction of different computations. Language learning is one of the most remarkable accomplishments of children (and sometimes adults), and this work may help to explain a key component of it. The domain of language has historically been a significant battleground for questions of mental organization and domain specificity, because of its unique properties and also its appearance in only one species. My dissertation work has explores a central component of language and reveals its true underlying nature as a set of highly interactive domain-general abilities and cue sensitivities, which together produce a domain-specific behavioral pattern. Phrase structure was a central focus of Chomsky's original theories of syntax and his critique of Skinner's Verbal Behavior. An understanding of phrase structure and hierarchical representation is therefore an important component of understanding language in general, and to understanding the nature of the human mind.
In the future there are a number of follow-ups to this work that I would like to pursue, as well as a number of other similar complexity issues I would like to study. First, the account here is incomplete on a number of fronts, and more work is necessary to understand all of the cues and their interactions that are components of hierarchical language learning. One component that has been left out is lexical variation in statistical and semantic information. The mental lexicon is of great interest because it forms direct connection between language, concepts and categorization. Second, it is crucial to measure how these cues are used and combined by children and infants for whom language learning is a daily task. Furthermore, it is important to understand how universal these types of cues are in other hierarchical domains, such as motor planning and the comprehension of causal relationships, in adults and children. Finally, it is necessary to ask where these simple sensitivities to cues came from evolutionarily and to explore what it is about these systems that sets humans apart from other animals. To achieve this, I hope to pursue collaborative work to explore capacities for hierarchical structure in other species.
My interests are certainly not limited to hierarchical structures, as this is just one instance of highly complex and domain-general mental abilities. I plan to pursue studies of how complex abilities of all kinds emerge from the interactions of the fundamental subsystems that underlie them. Sensation, perception, memory, categorization, hierarchical and linear organization, and learning all combine in a variety of fascinating ways to produce the behaviors that are so familiar in daily life. It is these complex interactions and combinations that I seek to analyze and explain. In all aspects of this work, I anticipate involving undergraduates as collaborators and colleagues. I have learned quite a bit from the process I chose for my dissertation and recognize the need for long-term research programs to answer large questions, as well as simpler individual studies that can stand on their own and produce individual publications. I have also had experience mentoring and advising students in research through a Laboratory in Development and Learning class that I taught last year. In this course I led a combined seminar on developmental research methods and worked with groups of 3-4 students on developing, organizing, running, analyzing, writing up, and presenting studies of cognitive and language development in preschool children. Each group had varied interests, from bilingualism, to autism and theory of mind, to concept formation, and even music cognition. Overseeing each of these research projects, which needed to be created and completed in a rather brief 10 week period, taught me how varied the student-professor relationship must be to best educate each student; I have learned to guide some scholars in pursuing their own research interests in thoughtful and scientifically cogent ways, and to involve less self-directed pupils in existing research programs to excite their enthusiasm in the field. While the brevity of this course was limiting, I think it serves as a good model of how I hope to involve a variety of undergraduates in my own research, while also supporting more senior students in pursuing their own research goals. My experience in research, combined with my love of teaching, will allow me to further mentor undergraduates and help them develop into successful researchers, thinkers, and contributors to both the discipline of psychology and the greater society.