Charles Clifton, Jr.
My colleagues are I are currently supported by NIH Research Grant HD-18708, "Language Comprehsion." In the recent past, we have been supported by NIH Training Grant HD-07327, "Training in Psycholinguistics," and NSF BCS 0090674, "Prosody in Language Comprehension". The goals of these grants are summarized in the "Specific Aims" or "Summary" sections of the grant proposals, which follow:
NIH "Language Comprehension" (1997)
The ability to comprehend sentences is among the most remarkable and distinctive human skills, and its impairment or loss following (e.g.) brain damage is among the most devastating of human illnesses. Our research program has, for many years, endeavored to understand the basic mechanisms of the human ability to turn a stream of words into a message, focusing on how readers and listeners impose structure on the incoming stream. This process is termed "parsing," and we assume that it is a precondition for comprehension. Readers and listeners obtain meaning from a structured stream of words, not an unstructured collection of words and phrases. We further assume that the ability to impose structure on a stream of words is at the heart of the human ability to understand language, and that its loss is at the heart of certain aphasias.
Through the 1980s, the psychological study of parsing was dominated by the "garden-path model," proposed by Frazier (Frazier, 1979; Frazier, 1987a; cf. Frazier & Clifton, 1996, Chapter 1) as a refinement of Kimball's (1973) proposals. This model contrasted with the "detective-style" models that existed previously (Fodor, Bever & Garrett, 1974), which were little more than lists of "strategies" or "cues" that were claimed to be relevant to parsing. It made strong and testable claims about the process of parsing, and made clear predictions about what sentence structures would be easy and difficult. The garden-path model was challenged in many ways (cf. Altmann, 1989), but arguably remained the "default parsing model" until very recent years.
Currently, however, parsing theory is becoming far more interesting and exciting. Attractive alternatives to garden-path theory have been proposed. Some retain the basic approach of garden-path theory (in its emphasis on the use of grammar in a symbolic parsing model), but emphasize different aspects of the grammar (Pritchett, 1988; Pritchett, 1992), or add some parallelism or underspecification to parsing representations (Crocker, 1994; Crocker, in press; Gibson, 1991; Gorrell, 1995). Other alternatives challenge the basic assumptions of the garden-path model, proposing more modern and explicit versions of the old "detective models" in the modern guise of constraint satisfaction models (MacDonald, Pearlmutter & Seidenberg, 1994a; MacDonald, Pearlmutter & Seidenberg, 1994b; McClelland, St. John & Taraban, 1989; Tanenhaus & Trueswell, 1995; Trueswell & Tanenhaus, 1994).
These developments have prompted us to re-examine experimentally some of the fundamental assumptions of our garden-path approach. One line of experiments we propose tests our basic assumption that the parser is serial and depth-first (Frazier, 1990; Frazier, 1995; Frazier & Clifton, 1996 submitted), initially creating a single analysis at a time to interpret. Competing constraint-satisfaction models assume that the parser activates a range of different analyses, in parallel, which compete to result in a preferred analysis. We propose experiments to test for the existence of activated analyses other than our presumed primary analysis, and to test other predictions that distinguish between garden-path and parallel constraint-satisfaction theories.
We have also been extending garden-path theory in fundamental ways, to permit it (unlike nearly all existing theories) to deal with what we call "non-primary phrases" (roughly, adjunct phrases and other phrases that cannot be a main predicate or its complement or argument; see below for a precise statement). We term our proposals "construal theory" (Frazier & Clifton, 1996). We have already conducted several tests of these new proposals, and describe several additional tests and extensions in the current grant proposal.
Finally, we continue to examine empirically a topic that has been at the heart of our research program since its inception, the processing of long-distance dependencies. The existence of long-distance dependencies (grammatical dependencies between "moved" phrases such as wh-question phrases and their "underlying" location) poses a difficult problem for any theory of human sentence parsing. At the same time, it gives human language its expressive power (permitting, e.g., quantification and variable binding). We propose experiments that examine the processing of sentences with long-distance dependencies in the context of the questions we have already raised (garden-path theory and its competitors, and its extension, construal theory), as well as experiments that introduce new concepts (such as roles in event structures) into claims about sentence processing.
NIH "Language Comprehension" (2001)
The ability to comprehend sentences is among the most remarkable
and distinctive human skills, and its impairment or loss following (e.g.) brain
damage is among the most devastating of human illnesses. Our research program
has, for many years, endeavored to understand the basic mechanisms of the human
ability to turn a stream of words into a message, focusing on how readers and
listeners impose structure on the incoming stream. We have to date focused on
this process of "parsing," and we assume that it is a precondition
for interpretation and comprehension of language. Readers and listeners obtain
meaning from a structured stream of words, not an unstructured collection of
words and phrases, and our previous research has contributed substantially to
the understanding of, and debate about, how people parse sentences (Frazier,
1978; Frazier, 1987; Frazier & Clifton, 1996).
Substantial disagreements remain about the nature of the parsing process, and we will continue gathering data designed to resolve these disagreements. However, we believe that enough is known about parsing to justify a shift in emphasis to studying the process of interpreting a parsed sentence (cf. Frazier, 1999a). There are many aspects to interpretation, including (1) evaluating the Logical Form of sentences, (2) determining reference and co-reference of referring phrases, (3) determining the relation of sentences to prior discourse and to context including identifying presuppositions, inferences, and discourse structure, and (4) assessing the speaker/author's intention in producing the sentence in its particular context. We propose to conduct experiments on a subset of topics in interpretation, concentrating primarily on the interpretation of elliptical constructions but extending this concentration to include the processing of focus and presupposition, the interpretation of bound pronouns, and the identification of sentence regions or domains that are taken as primarily relevant to the interpretation of new linguistic input.
We propose some principles described below (primarily the Conjunction Principle and a principle of cost-free structure-copying we call "copy ") that impose conceptual coherence on this range of topics. Our experiments are designed to test these principles using both off-line and on-line techniques. The initial experiments focus on how readers and listeners comprehend sentences and short discourses where they must "fill in" material to be able to interpret them. These include, among other forms, VP ellipsis (Keith gave a talk and Sandy did too) and question-answer pairs. Later experiments broaden the range of constructions that we examine and explore interactions between grammatical form and semantic factors such as focus, presupposition, and causality.
We believe that our understanding of the parsing process, while incomplete in major aspects, is adequate to support productive research on how the products of parsing are interpreted. The experiments we describe below, in addition to testing specific proposals about the nature of interpretation, should help support a better understanding of a variety of issues in language comprehension. These issues include the role of focus and presupposition in integrating sentences with discourse, the nature of the syntactic domains within which different forms of processing take place, and the nature of the processes that are distinct and that are shared between sentence comprehension and discourse comprehension (see Frazier, 1999b, for some speculations on this last point). We also we believe that basic work on the relation between sentence comprehension and discourse comprehension will support deeper insights into the nature of aphasia and other language disorders, and note the existence of recent work, described briefly below, discussing whether some of the phenomena that agrammatic aphasics exhibit are essentially syntactic (Friedman & Grodzinsky, 1997) or discourse-based (specifically, reflective of the overuse of ellipsis; de Roo, 2000).
NIH "Language Comprehension" (2007)
We propose research on how semantic structure affects sentence comprehension. We make the working assumption that there are human cognitive specializations for language, and we propose to search for such specializations in a relatively-underexplored domain, that of the processing of semantic structure. We believe that this position has health-relevant implications: language-related impairments such as aphasia seem to involve the loss of specifically-linguistic functions, and these impairments can be fully understood only if the normal operation of these functions is understood.
In particular, we propose to focus on the different aspects of a concept that plays a central role in formal semantics, namely, semantic variables. The variables we will study cover three types. One type is standard quantificational structures, such as Every girl laughed, which are generally analyzed as ‘For each x, girl (x), laughed (x). They thus involve an operator (the universal quantifier every), a restriction on the domain of quantification (girls, plus possible implict restrictors) and an asserted predicate or ‘nuclear scope’ (laughed). Another type of variable we will study is a world variable, where by convention W0 is used to represent the actual world. A world variable is often introduced by a modal, which usually quantifies over possible worlds, or introduced by a non-actuality implicature which may restrict a state of affairs to being non-actual. For example, John would have gone to Paris in May may be understood as implying that John didn’t go to Paris (NOT-John went to Paris in May in W0), implicating another value of the world variable. The third type of variable is a variable bound by a lambda operator, which forms properties. In the sentences of interest here the lambda operator typically forms properties corresponding to a verb phrase: ‘John called his mother and Bill did too’ involves copying the property ‘ λx, called x’s mother’ which may be predicated of John (resulting in an interpretation where John called John’s mother), and which may also be predicated of Bill (resulting in an interpretation where Bill called Bill’s own mother; note, this interpretation is often referred to as ‘sloppy identity’).
We use a variety of behavioral techniques to study the comprehension of written and spoken language. The techniques range from simple (but rigorously designed) questionnaire studies of what people take ambiguous sentences to mean when their form or prosody or context varies, through speeded acceptability judgment, stops-making-sense, and self-paced reading techniques, to measuring eye movements during reading. We propose simple techniques when they answer the questions we pose, and the more refined, analytic, techniques when they are appropriate.
The research proposal is organized as follows. Section 1 looks
at questions about variable binding. Many of the studies investigate structures
where the reader arrives at the same interpretation through distinct routes.
The idea is that contrasting different routes to the same interpretation may
reveal the mechanisms underlying the interpretation. Bound variables (as can
be seen, for instance, in quantificational structures) provide one route to
a ‘covariation’ interpretation
which may operate regardless of whether the covariation is strongly suggested
by extralinguistic knowledge (every man has a nose) or not (every man has a
hose). By contrast, covariation may be introduced simply because it is the
most plausible relation. The boys’ noses probably yields a 1:1 interpretation
by default, but the boys’ hoses probably does not. In general we expect
that when a covariation interpretation results from compositional interpretation
of the semantic structure (involving a bound variable) it will be in principle
available even with implausible relations (though if the structure is ambiguous,
or the implausibility is great enough, then real world likelihood may influence
the selection of a particular interpretation). In the absence of a quantificational
structure, we expect that a covariation interpretation will be arrived at primarily
in cases where it corresponds to expectations about the world, and will constitute
what we might term ‘fake variable binding. ‘(See Shapiro, Hestvik,
Lesan, & Garcia,
2003, for a forerunner of this reasoning.)
Section 2 introduces a topic that has been studied very little
by psycholinguists interested in sentence processing, the role of world variables
in processing modals and non-actuality implicatures. (though see Bucciarelli & Johnson-Laird,
2005; Dwivedi et al.,2006) It looks at one aspect of processing modals, and it
investigates the role of non-actuality implicatures. It uses ellipsis sentences
as a test ground. In a sense, it returns us to the primary focus of our research
during the previous grant period, namely, the interpretation of elliptical sentences.
One of our findings was that, in a two-sentence context that highlighted discourse
relations, the preferred antecedent of an ellipsis was a linguistic object that
is central to the discourse. For instance, readers most frequently interpret
John said that Joe went to Paris. Bill did too as meaning that Bill said something.
But if this is changed to John said that Joe needed to go to Paris. Bill did
too we suspect that the preferred interpretation is changed to “Bill needed
to go to Paris.” We propose that this is because the modal needed to implicates
that the event did not actually happen, which creates a contrast between the
actual world and some “other world” in which the event did happen,
focusing the event and thus increasing its discourse centrality.
Section 3 looks at questions of “quantificational variability.” It
examines sentences where readers have to quantify over some domain (e.g., predicate
something of a variable that varies over some specific domain). It examines
how readers and listeners choose a particular type of domain (e.g., a set of
degrees, a set of entities, a set of propositional alternatives) and it attempts
to identify various preference principles for choosing a domain. It also studies
how readers restrict or expand the domain that they choose. To take a homely
example of selecting a domain, a sentence like The reviewers are mostly exhausted
can quantify over different domains: entities (most of the reviewers) or times
(most of the time) or degrees (nearly exhausted). This section also looks at
the role of semantic ambiguity in determining processing complexity, asking
whether semantic ambiguity increases, or decreases, or does not affect, the
speed of understanding a sentence. It focuses largely, though not exclusively,
on ambiguities concerning domain selection.
Our research examines processes of sentence interpretation in normal, intact humans, but its results should significantly enhance understanding of disorders of language. The contemporary study of aphasia is heavily influenced by basic research findings (e.g., Caplan, 2000; Stromswold et al., 1996), and we believe that the particular topics that we are studying are likely to apply directly to the analysis of processing deficits in aphasia. We note the existence of promising studies of the interpretation of pronouns in ellipsis (Vasic et al., 2006) and of deficits in semantic coercion in Broca’s and Wernicke’s aphasics (Piñango & Zurif, 2001) and submit that the work we propose to do has the potential of stimulating and guiding advances in the understanding of language disorders.
NSF Prosody in Language Comprehension
In order to understand a sentence or discourse, readers and listeners
must use their implicit knowledge of the grammar of their language to put the
meanings of its words together. A great deal has been learned about how this
process takes place during reading. Less is known about the aspects of the process
that are specific to listening. The proposed research will explore how a listener
uses the prosody of a sentence (informally speaking, its rhythmic and melodic
structure) to determine its message. The specific goals of the research are
to work toward identifying precisely what aspects of the prosody of a sentence
affect a listener's comprehension and to examine the cognitive processes that
enable a listener to use prosody. Broader goals include working toward the development
of a theory of spoken language comprehension that specifies how prosodic descriptions
are arrived at and how they interact with other sources of information about
language and potentially enabling technological improvements in communication,
for instance in human-computer spoken interaction.
The proposed research will explore just which aspects of a sentence's prosody listeners use in identifying its structure and meaning. It will evaluate the thesis that a listener constructs a global prosodic representation of a sentence rather than simply relying on local prosodic cues. This thesis makes it crucial to address the question of what in the global prosodic representation is actually effective in guiding sentence comprehension.
Some of the proposed experiments examine how ambiguous phrases can be attached to different points in the phrases that precede them. For instance, in a sentence like Sammy learned that Bill telephoned after John visited, the adjunct phrase after John visited could modify either learn or telephone. Such a phrase should be more likely to attach high, modifying the main clause verb learn, if it is preceded by an intonational boundary, which is signalled by pitch movement and temporal changes. The proposed experiments investigate whether different types of intonational boundaries have different effects and evaluate the claim that only an "informative" boundary will affect interpretation. They evaluate a definition of "informative" that claims a boundary is informative if it is phonologically larger than certain structurally-defined earlier boundaries. The experiments provide several different tests, using several different grammatical constructions, of whether prosodic boundaries that are informative under this definition encourage high attachment of ambiguous phrases
Other experiments addresses the hypothesis that listeners take unusual ("marked") prosody to indicate an unusual interpretation of a sentence. This "markedness strategy" has been proposed a number of times, and several pieces of experimental evidence seem to be consistent with it. However, the markedness strategy may be only a crude approximation to reality. The proposed experiments, which study the interpretation of pronouns, reflexives, and quantified NPs, attempt to determine whether the phenomena described by the markedness hypothesis actually reflect simpler, more general prosodic principles of focus and accent.