Current
Projects in a Nutshell
Dynamical Models of Linguistic
Generalization. This project is
concerned with understanding how speakers of a language generalize beyond their
input. It is focused on discerning the line between syntactic and semantic
constraints, and thus on understanding better what this distinction is all
about.
Dynamical Models of Complex Structure. One piece of my postdoctoral work indicated a way of
using dynamical structures like basins and attractors to organize constituent
structure representation in a neurally plausible mechanism. The central
interest of this approach is that it portrays a set of relationships between
structured systems which are hard to discern in traditional symbolic
formalisms. Now I'm looking at what those relationships are and how they help
with the hard problem of learning complex structures.
Ungrammatical Influences. Another piece of my postdoctoral work identified a
case of the natural language phenomenon I call "Ungrammatical
Influences": sometimes the parser seems to temporarily construct
representations that are not grammatically consistent. The case I studied there
involved English compound nouns. To find out if the phenomenon is widespread
enough to be considered a significant issue for the theory of language
processing, I am now investigating, in collaboration with Bruno Galantucci,
other types of ungrammatical influences, involving English idioms, English
subcategorization and Italian anaphora. In conjunction with the experimental
investigations, we are studying neural network models which predict the effects
in order to better understand if and how the phenomenon urges us to revise our
current notions of linguistic representation.
The
Thinking that Led to these Concerns
My research is centered around the question,
How can structure undergo graceful change? I am interested in the relationship
between systems (like generative grammars) that are highly structured, that
make absolute claims, and that achieve great mileage by using a calculus of
symbols, and systems (like physical dynamical systems) that are flexible, that
make only relative claims, and that lend themselves to analysis using
continuous-variable math. My hunch is that these two types of systems can
correspond to two different viewpoints of a single system, much as Smolensky
(1988) has suggested by casting the corresponding interpretations of
connectionist networks as different ``levels of description". I believe
that a deeper understanding of this relationship can shed light on a variety of
general questions like how institutions can escape from states of entrenchment,
how evolution generates new, highly-structured entities that function, how
human beings can strike a balance between their drive for control and their
longing for creative loss-of-control.
In graduate school, I examined the flexiblity
question by studying historical grammar change. It is often possible to
establish that, at a certain point in time, a language has grammar A and at a
later point it has grammar B, which is distinct from A. Moreover, the
transition is usually ``graceful'' in the sense that only small changes are
made over small intervals of time. But the grammars of the transition period
are often puzzling hybrids of A and B. My thesis involved a text-based analysis
of several such representationally challenging episodes. One study traced the development
of the degree modifier use of sort of and kind of in English
(e.g., The wheelbarrow sort of slipped off the track) from the
historically-prior Noun-Preposition use (e.g., What sort of garden do you
envision?). I chose connectionism (or ``neural networks'') as a framework
because it provides a way of expressing structure in a system that is
underlyingly flexible, and because certain connectionist networks develop
representations which can be explicitly compared to standard linguistic
constructs. The models I worked with made an interesting prediction, which
several case-studies in my thesis supported: categorical grammar changes are
preceded by anticipatory frequency changes. For example, prior to the emergence
of the first unambiguous degree modifier uses of sort/kind of, the
relative frequency of the collocation in now-ambiguous examples like a sort
of orange brick rose disproportionately to other changes in the
distribution of the collocation. I called this phenomenon
"Q-divergence" for "Quantitative Divergence". The model
predicts the effect by fitting a continuous function to a set of categorical
data: most of the data points can be accurately modeled, but some smoothing is
inevitable near points of discontinuity. This analysis implies the existence of
ties between seemingly independent structural components which can serve as
pathways for revision of the structural relationships. See the papers, `Lexical
Change as Nonlinear Interpolation' and `Continuity in Language Change'. (See papers page.)
While at Rochester, I realized that there is
an appealing similarity between the issues raised by my language change work
and issues raised by psycholinguists who study the role of lexical influences
in sentence-processing: while I have found that incremental quantitative
changes in the behaviors of words can lead to their grammatical
reclassification over time, a growing body of psycholinguistic research
indicates that small quantitative differences in the behaviors of words are correlated
with the strengths of people's tendencies to temporarily misclassify them (i.e.
to be temporarily ``garden pathed"). This phenomenon has been the subject
of a lively debate in the field of psycholinguistics over the past two decades:
it is clear that syntactic information can dictate the expectations of the
parser in many circumstances, but it is also clear that lexical biases can
hedge syntactic bets, sometimes even hedging them in the direction of
ungrammatical structures.
What kind of representation system could
express this surprising interaction of levels in a parsimonious fashion? In
collaboration with Cornell Juliano and Michael Tanenhaus I extended the
connectionist work of my thesis to develop a theoretical framework which
answers this question. The framework is based on dynamical systems theory. An
important notion is that of an attractor---a point toward which a dynamical
system gravitates. When a system is near an attractor, it behaves, for all
practical purposes, as though there were no other attractors around. But when
it is intermediate between several attractors, it reflects their simultaneous
influences. By modeling syntactic parse structures as attractors, we were able
to account not only for their predominant ability to ``capture'' linguistic
behavior but also their marginal ability to influence it. We used a
connectionst network to define the attractor topography. The surprising ability
of ungrammatical constructions to ``hedge the bets'' of the processor is a
reflection of the same ties between grammatically independent components that
allow the historical models to make appropriate predictions about language
change. This work is described in the papers, `A Dynamical System for Language
Processing' and `Parsing in a Dynamical System' (See papers page).
On another postdoc at MIT I worked on the
question of how to represent complex constituent structures in flexible,
learning systems like neural networks. The papers 'Dynamical Automata' and
'Fractal Encoding of Context Free Grammars in Connectionist Networks' describe
a solution to this problem and a useful insight into the nature of the
difference between symbolic and neural computation.
On a third postdoc at Cornell, I began
investigating a phenomenon that is predicted by self-organizing models of
syntax and presents a particularly interesting challenge to current
representational theories: readers sometimes get distracted by a familiar word
sequence whose familarity should only be recognizable if they are forming a
structure that violates the grammar of their language. I call this phenomenon
"Ungrammatical Influences". It is described in the poster called
'Ungrammatical Influences in Sentence Processing'. (See papers page.)
Now I have a job in the Department of
Psychology at the University of Connecticut. There is a good synergy here
between people working on language and people who work on dynamical systems
from the perspective of haptic perception and motor movement (See the Center
for the Ecological Study of Perception and Action). I teach courses in Sentence
Processing, Cognitive Psychology, and a laboratory in Psycholinguistics. I am
pursuing three major projects:
Dynamical Models of Linguistic
Generalization. This project is
concerned with understanding how speakers of a language generalize beyond their
input. It is focused on discerning the line between syntactic and semantic
constraints, and thus on understanding better what this distinction is all
about.
Dynamical Models of Constituent Structure. My work at MIT indicated a way of using dynamical
structures like basins and attractors to organize constituent structure
computation. In the papers described above, I showed how such models provide a
way of implementing canonical symbolic grammars in neural hardware. The central
interest of this approach is that it portrays a set of relationships between
structured systems which are hard to discern in traditional symbolic
formalisms. Now I'm looking at what those relationships are and how they help
with the hard problem of learning to use complex structures.
Ungrammatical Influences. My work at Cornell identified one case of the
existence of ungrammatical influences. To find out if the phenomenon is
widespread enough to be considered a significant issue for the theory of
language processing, I am now investigating, in collaboration with Bruno
Galantucci, other types of ungrammatical influences. In conjunction with the
experimental investigations, we are studying neural network models which
predict the effects in order to better understand if and how the phenomenon
urges us to revise our current conceptions of linguistic representation.
In sum, I am currently in the trenches of a
large project whose aim is to understand how human beings strike a balance
between control and flexibility, a challenge upon which much seems to hinge.
Return to Whit Tabor's homepage.
Go to Whit Tabor's papers.
This page was last updated 99-08-26.