First Year Report
The Role of Prosody in Pronoun Resolution:
Evidence from Psycholinguistics and
Neuroimaging
Ryan C. Taylor
Daily Supervisor: John C. J. Hoeks
Promotor: Gisela Redeker
Information & Communication Sciences
July 17, 2009
Contents
1 Theoretical Introduction
1.1 Defining the Problem: Accenting, Pronouns, and Accented Pronouns
1.2 Overview of Existing Work . . . . . . . . . . . . . . . . . . . . . . .
1.3 Semantic accounts of pronoun resolution . . . . . . . . . . . . . . .
1.3.1 Person, gender, and number marking . . . . . . . . . . . . .
1.3.2 Implicit causality . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Syntactic accounts of pronoun resolution . . . . . . . . . . . . . . .
1.4.1 Subject assignment/Parallel function . . . . . . . . . . . . .
1.4.2 The extended feature match hypothesis . . . . . . . . . . . .
1.5 Plausibility/General knowledge . . . . . . . . . . . . . . . . . . . .
1.6 Working memory . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7 Discourse structure . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.1 Information structure . . . . . . . . . . . . . . . . . . . . . .
1.7.2 Coherence relations . . . . . . . . . . . . . . . . . . . . . . .
1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.8.1 Interim Summary . . . . . . . . . . . . . . . . . . . . . . . .
1.8.2 Future directions . . . . . . . . . . . . . . . . . . . . . . . .
1
1
2
3
3
4
6
6
6
8
10
11
12
17
19
19
20
2 Experiment 1
2.0.3 The Resolution of Ambiguous Pronouns
2.1 Method . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Participants . . . . . . . . . . . . . . . .
2.1.2 Materials and design . . . . . . . . . . .
2.1.3 Procedure . . . . . . . . . . . . . . . . .
2.2 Results . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Prosodic naturalness . . . . . . . . . . .
2.2.2 Referent preference . . . . . . . . . . . .
2.2.3 NP1 and NP2 verbs . . . . . . . . . . . .
2.2.4 Grammatical role parallelism . . . . . .
2.3 Discussion . . . . . . . . . . . . . . . . . . . . .
21
21
25
25
25
26
27
27
28
29
30
31
i
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
ii
2.3.1
An account that incorporates the conceptual level, coherence
and information structure . . . . . . . . . . . . . . . . . . . 32
Information structure and intonation . . . . . . . . . . . . . 32
2.3.2
3 Experiment 2
3.0.3 Sentence processing models . .
3.0.4 Discourse structure ambiguities
3.1 Method . . . . . . . . . . . . . . . . .
3.1.1 Participants . . . . . . . . . . .
3.1.2 Materials and design . . . . . .
3.1.3 Procedure . . . . . . . . . . . .
3.2 Interim Results . . . . . . . . . . . . .
3.3 Discussion . . . . . . . . . . . . . . . .
4 Future directions
4.1 Context-poor environments
4.2 Context-poor environments
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
35
35
37
38
38
38
39
39
40
41
. . . . . . . . . . . . . . . . . . . . . . 41
. . . . . . . . . . . . . . . . . . . . . . 43
A Schedule
A.1 Year 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.1 Year 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.2 Year 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
44
45
45
B Quick and Dirty Guide to Optimality theory
46
References
49
Chapter 1
Theoretical introduction to
accented pronoun comprehension
1.1
Defining the Problem: Accenting, Pronouns,
and Accented Pronouns
In spite of decades of work creatively exploiting different levels of linguistic analysis and different linguistic theories, there is not a model of unaccented pronoun
resolution that is completely effective. Furthermore, recent research shows the
importance of intonation in Dutch and English information structure and even
syntactic parsing. For this reason, research on prosody and pronouns has been
avoided. However, theoretical advances in the study of discourse structure, and
technical advances have given us new techniques and new theories unavailable to
our forerunners, and of course the opportunity to learn from previous attempts.
Advances in the understanding of prosody and technical advances in ERP now
make the on-line investigation of prosody possible—a notable advantage over past
attempts to understand accented pronouns. But improvement of technique without
improvement of theory is like having a faster car with nowhere to go.
Attempts to explain pronoun reference run the gamut of linguistics. This heterogeneity makes artful cataloguing of earlier work difficult, although the diversity
of the approaches also means that we have much to draw on in creating a working
model.
That being said, the proposals are a product of their times. Early theories
of pronoun resolution tend to draw their ideas from single sources and make use
of just a single heuristic. By comparison, more recent theories are eclectic in
their approach, attempting to draw on multiple sub-fields. This eclectic approach
appears sensible; many of these approaches merit experimentation. We have also
tried to cast the net wider than previous theories by incorporating theories of
1
CHAPTER 1. THEORETICAL INTRODUCTION
2
information structure, activation, and plausibility. No one has yet made a proper
go of extending these theories to pronoun resolution. We present past and current
theories/heuristics/models that apply to pronominalisation.
In this chapter, we have grouped the theories according similarities in the
way that they determine pronoun reference. The sub-fields of semantics, syntax
and discourse structure are all represented. Additionally, we examine the roles of
plausibility and working memory. We end the chapter by discussing which theories
we feel are the best for future experimentation. By grouping the theories together
it is possible to see pros and cons of each and thereby which theory is best for
future experimentation.
1.2
Overview of Existing Work
Within the research we have reviewed, we can distinguish four main approaches
of pronoun disambiguation: semantics, syntax, working memory, plausibility, and
discourse. Semantics are always acknowledged to be important—probably all theories agree that gender and number agreement are necessary. Implicit causality
also used semantics to describe pronoun disambiguation by looking at the influence of verbal semantics on pronoun interpretation (Caramazza, Grober, Garvey,
& Yates, 1977). Two syntactic theories of pronoun attribution have proven particularly influential: parallel function (Smyth, 1994) and subject assignment (Crawley,
Stevenson, & Kleinman, 1990). These theories have taken syntactic properties of
pronouns as important features in the disambiguation process.
Plausibility or general knowledge likely also plays a role in pronoun resolution;
to this end, we examine recent proposals like plausibility (Connell & Keane, 2004,
2006).
The notion that working memory contributes to language originally proposed
in Chafe (1973) finds a recent parallel in the working memory-model of sentence
processing (Lewis & Vasishth, 2005; Lewis, Vasishth, & Van Dyke, 2006). Lewis
and Vasishth propose a set of heuristics that together are supposed to explain the
impact of working memory on sentence processing.
Discourse structure has frequently been used in theories of pronoun resolution.
Information structure, in particular has been used a great deal. Centering theory uses information structure as a primary feature. It predicts that the most
prominent concept in discourse will be the pronoun (Gordon & Hendrick, 1998,
1999).
Accent has figured more prominently in recent work than before. Recent work
sees accent as a component of information structure (Steedman, 2007; Schwarzschild,
1999; Vallduvı́ & Vilkuna, 1998; Vallduvı́ & Engdahl, 1996). Schwarzschild (1999)
formalized the definition of givenness and uses it to predict accenting. Beaver
CHAPTER 1. THEORETICAL INTRODUCTION
3
(2004) extends centering theory to accenting. De Hoop (2004) describes accenting
in terms of continuing topic and contrastive stress.
Like proponents of information structure, proponents of coherence relations
use discourse structure as an explanatory tool in pronoun disambiguation (Kehler,
Kertz, Rohde, & Elman, 2008; Kehler, 2005; Kertz, Kehler, & Elman, 2006b,
2006a; Rohde, Kehler, & Elman, 2006). Coherence relations show great promise,
although likely other factors are additionally important in a working model of
pronoun resolution. Hendriks (2004) combines the notions of coherence relations
with contrastive topic and suggests that some of the coherence relations result from
constraints that govern the arrangement of information. We further investigate
notions of contrast in order to fully exploit these observations (Steedman, 2007;
Vallduvı́ & Vilkuna, 1998). The role of coherence relations in pronoun resolution
seems a promising area for further investigation.
To facilitate understanding how exactly each theory works, when possible we
will discuss the theories’ predictions for (1):
(1)
1.3
1.3.1
John called Geoff, and Steven emailed him.
NP1 Verb NP2 and NP3 verb
Pro.
Semantic accounts of pronoun resolution
Person, gender, and number marking
In the study of pronoun resolution, one of the few points of universal agreement
is that at some stage person, gender, and number marking are taken into account
when finding the referent of a pronoun. In European languages, pronouns agree
with their referents in person, gender, and number (Comrie, 1999). The person of
the pronoun is whether it refers to the speaker, the hearer, or some third party. The
number of a pronoun may be singular or plural in languages like English, although
in other languages, for example Arabic and Hebrew, there is the additional category
dual, referring to two things.
While all languages have a person and number system, grammatical gender is
not present in all languages (Comrie, 1999; Aikhenvald & Dixon, 1998). Crosslinguistically, grammatical gender (or noun class) is sometimes not present at all.
In Turkish for example there is no grammatical gender. This is not to say that in
Turkish biological gender is not used whatsoever; the king is pregnant is no more
sensible in that language than in any other, but Turkish has just one pronoun
referring to a man or a woman, o (Comrie, 1999).
Gender often matches with a semantic feature, such as in English where a per-
CHAPTER 1. THEORETICAL INTRODUCTION
4
son (or sometimes animal’s) sex is important1 (Comrie, 1999). In other languages,
gender is a morphosyntactic feature. Pronouns and determiners agree with the
pronoun in gender. For example in Dutch where a computer is hij ‘he’, and the
universiteit is zij ‘she’ (Mel’čuk, 1988; Comrie, 1999). Most often systems with
gender are somewhat mixed—some nouns are assigned to classes based on semantic
features, and others have arbitrary gender (Comrie, 1999).
While person, gender and number are cross-linguistically widespread, this does
not imply that they contribute to faster processing. Hobbs (1977) describes a
“naı̈ve algorithm” that matches pronouns based on essentially taking the first
referent matching in gender and number within the same clause, moving from
left to right. He finds that this algorithm works surprisingly well (although not
perfectly) implying that number and gender could at least help to reduce the field
of possible referents. This suggests that gender and number should at least be
helpful in finding the referent. But does gender constrain the parser? Whereas in
(1) the referent was ambiguous with respect to gender, in (2) the referent can be
chosen based on gender alone:
(2)
Sarah called Geoff, and Steven emailed him.
In (2) the sentence is only grammatical if the referent is singular, male and animate. Van Berkum (1999) cites three possibilities with regard to on-line gender
comprehension: gender is not used; gender is used, but late in processing; and
gender is used immediately after it is encountered. His ERP evidence suggests
that gender information is used immediately after being encountered. So, gender
can restrict the field of available referents. But sentences like (1) can be resolved,
suggesting that person, gender, and number are not sufficient to explain pronoun
resolution.
1.3.2
Implicit causality
If gender and number marking has been the most universally accepted feature in
pronoun resolution, then implicit causality is the most studied. Garvey, Caramazza, and Yates (1975); Caramazza et al. (1977) propose that implicit causality
is a feature of a verbal semantics that selects an NP as the antecedent of a pronoun. Moreover, they state that the semantics of the verb are “perhaps dominant”
in the selection of an NP, outside of suprasentential context (Garvey et al., 1975;
Caramazza et al., 1977). To determine the implicit causality bias of a sentence,
participants were presented with sentence fragments followed by a blank. Participants wrote continuations to sentences that had the format shown in (3), where
1
although sometimes ships are female, and children of unknown gender are it
CHAPTER 1. THEORETICAL INTRODUCTION
5
both the NPs and the pronoun had the same gender, making the pronoun ambiguous.
(3)
John admired Bill because he
NP1 VP[past] NP2 because Pro
They found that the VP biased participants towards completing the sentence such
that either the first NP (NP1 ) or the second NP (NP2 ) was referred to by the
pronoun, a property they refer to as implicit causality2 .
Garvey et al. (1975); Caramazza et al. (1977) suggest that many, although not
all NP2 verbs refer to the emotions of the antecedent as one possible explanation,
and they also suggest “pragmatic factors”. Applied to (3) above, implicit causality
predicts that the referent of the pronoun is determined by the implicit causality of
the verb in the first sentence. The opacity of the concept of implicit causality is
admitted by Garvey et al. (1975); Caramazza et al. (1977) to be problematic (see
§1.7.2 for a coherence relations interpretation of the results).
Koornneef and Van Berkum (2006) look at implicit causality verbs on-line in
a self-paced reading task to see at what point the processor uses this information.
They compared sentences where the implicit causality bias was consistent with the
pronoun to sentences where the implicit causality bias was inconsistent with the
pronoun (4):
(4)
a.
NP1-biased verb, bias-consistent pronoun
David and Linda were both driving pretty fast. At a busy intersection
they crashed hard into each other. David apologized to Linda because
he according to the witness was to blame.
b.
NP1-biased verb, bias-inconsistent pronoun
David and Linda were both driving pretty fast. At a busy intersection
they crashed hard into each other. Linda apologized to David because
he according to the witnesses was not the one to blame.
There are signs that the information is integrated at the pronoun (although this
fails to reach significance in both the F1 and F2 analyses). The results are clearer
still in Van Berkum, Koornneef, and Otten (2007). They found that when a
pronoun was inconsistent with the implicit causality bias (as in (4b), a deflection of
the ERP signal was detected 400 ms after the onset of the pronoun, corroborating
the evidence of the previous study. Implicit causality, then, is used from an early
stage in processing.
2
Our NP subscript 1, 2, 3 is essentially the same as Garvey et al.’s (1975) notation, except
that in this paper we are using a subscript on the pronoun to denote its referent.
CHAPTER 1. THEORETICAL INTRODUCTION
6
More recently, implicit consequentiality has been proposed (Crinean & Garnham, 2006; Pickering & Majid, 2007). In implicit consequentiality, inferencing
is the reverse of implicit causality: cause precedes effect in implicit causality and
follows it in implicit consequentiality, (5):
(5)
Because Bill admired John, he
Implicit consequentiality creates bias much like implicit causality (Crinean & Garnham, 2006; Pickering & Majid, 2007).
In the section on coherence relations we discuss the recent proposal that coherence relations are in fact the “semantics” underlying choice of referent in implicit
causality and consequentiality verbs.
1.4
Syntactic accounts of pronoun resolution
Others have sought a syntactic rather than semantic solution to the pronoun reference problem, an obvious route within the syntactocentric atmosphere in the late
twentieth century.
1.4.1
Subject assignment/Parallel function
Crawley et al. (1990) focused on the subject assignment strategy and the parallel
function strategy. The subject assignment strategy proposes that when interpreting
a pronoun, participants choose the subject of the previous phrase as antecedent
(NP1 ). The parallel function strategy states that participants will choose the NP
of the previous phrase that matches the grammatical role of the pronoun.
Note that in a sentence such as example (1), p.3, the two hypotheses would be
in direct contradiction. The subject assignment strategy predicts that the pronoun
will refer to the subject of the sentence (the first constituent, NP1 ) and parallel
function predicts that the pronoun will refer to the NP with the same grammatical
role (usually the second NP, NP2 ) or will fail. Crawley et al. (1990) found that
there was a bias toward the subject of the preceding noun phrase, thus favouring
the subject assignment strategy over parallel function
1.4.2
The extended feature match hypothesis
Smyth (1994) took issue with Crawley et al.’s results, citing the form of the stimuli
and the manner in which the two heuristics were examined in particular. He
contends that the majority of the stimuli that had been used in the previous study
were not in fact parallel, given that they were not controlled for thematic role
CHAPTER 1. THEORETICAL INTRODUCTION
7
or grammatical role parallelism. Moreover, he takes issue with Crawley et al.’s
analysis, suggesting that the two strategies were examined as separate entities.
With regards to parallelism, there are actually two kinds of parallelism here.
Crawley et al.’s (1990) version refers to sentences where NPs have the same linear
order, whereas Smyth’s (1994) definition requires grammatical role parallelism
in addition to linear order. If the grammatical roles are not the same in both
sentences, then the referent of the pronoun is uncertain. In (6a) reference fails
because the grammatical roles are not the same:
(6)
a.
b.
The carpenter gave the plumber an invoice, and the electrician gave to
him a cheque.
The carpenter invoiced the plumber, and the electrician gave him a
cheque.
(Smyth, 1994, p.202)
In (6a), the first clause is ditransitive, and the second clause is transitive. In (6b),
the grammatical roles match. The information is essentially the same but the
grammatical roles do not match. Restating the hypothesis, when the grammatical
role of one clause is not fully syntactically parallel with that of the previous clause,
or does not match morphologically, participants resort to the subject assignment
strategy.
Smyth, like Crawley et al., essentially describes pronoun resolution through the
lens of syntax, notwithstanding the inclusion of gender and number agreement as
factors. Smyth follows his forerunners in his use of sentences that have “pragmatically possible antecedents”(p.197), but are “free” of pragmatic biases. We discuss
later how these pragmatic biases may be the key to the entire puzzle in the form
of plausibility p.9 (or at least one tool in a lock-picking kit).
The experiments Smyth performed indicated that in fully parallel sentences,
the referent of a subject pronoun is also a subject. Smyth’s hypothesis derives
in part from binding theory, which states that pronouns are matched with an
antecedent that shares their grammatical role, gender, or number. According
to this extended feature match hypothesis, the pronoun is matched with possible
antecedents and the antecedent that shares the most features is the referent. Some
of these features, such as gender and number cannot be violated, whereas others,
such as grammatical roles, may be violated.
This account attempts to go beyond linguistic explanation, explaining the parallel preference as a form of priming. In psycholinguistics, linguistic features are
primed by previous mention of a word or sentence that has the same structure.
Structural priming occurs when a grammatical structure used in the first clause is
activated, and the items in the first clause are expected to have the same syntactic
role in the second clause. It is worthwhile to point out that Smyth was wise to
CHAPTER 1. THEORETICAL INTRODUCTION
8
include the notion of interclause priming and reactivation3 . These types of priming
have found support in a number or places outside the domain of pronoun resolution, ranging from on-line response time experiments to neuroimaging (Arregui et
al., 2006; Kaan, Wijnen, & Swaab, 2004).
Recent studies have shown that structural priming is strong enough that even
syntactic misanalyses have an effect on the parsing of the next sentence (Van
Gompel et al., 2006). It has also been attested with various constructions, crosslinguistically, in spoken sentence completion and sentence recall (Van Gompel et
al., 2006).
But against Smyth’s (1994) hypothesis is the finding that priming can be either
syntactic or non-syntactic in nature (Van Gompel et al., 2006). That is, similarity
of the linear position of words between two clauses is enough to allow facilitation
in the processing of the second clause. This means that parallel structures do
not need to be syntactically identical, nor do they need to be arranged the same
linearly, although both of these help processing (Van Gompel et al., 2006). This
is further discussed in the section on coherence relations.
The extended feature-matching hypothesis predicts that a pronoun will refer
to the NP in the previous clause that shares its grammatical role. Otherwise, it
will refer to the subject of the previous clause. This means that in (7), if the NP2
is an oblique object, but the pronoun is an object, then the pronoun refers to the
subject, NP1 .
(7)
John called Geoff, and Steven emailed him.
NP1 Verb NP2 and NP3 verb
Pro.
An example of an oblique object is “James” in “Frank gave Steve a parcel for
James” where “Frank” is the subject, “Steve” is the direct object, and “a parcel”
is the indirect object4 .
1.5
Plausibility/General knowledge
Smyth (1994) and Crawley et al. (1990) both refer to pragmatic biases, although
they do not define what precisely these biases might be. Rather, they seem to
3
a.k.a. syntactic frame recycling in Arregui, Clifton Jr., Frazier, & Moulton, 2006, or structural priming Van Gompel, Pickering, Pearson, & Jacob, 2006.
4
“In this view, subject assignment is a default strategy for sentences in which the degree of
nonparallelism exceeds some limit; PF is a specific outcome of the more general principle that the
probability of parallel resolution depends on the number of features shared by the pronoun and
the candidate antecedents. Retaining subject assignment in the model allows us to account for an
otherwise mysterious asymmetry between subject and non-subject pronouns”.(Smyth, 1994)¡pp.
205–205¿
CHAPTER 1. THEORETICAL INTRODUCTION
9
suggest that bias is something above the sentence level and offer no further commentary. Caramazza et al. (1977) at least attempt to define how pragmatics interacts with pronoun resolution, suggesting a “pragmatic principle”, stating that
verbs that describe emotions cause participants to create sentence continuations
describing the reason for the experiencer’s emotions. In this perspective, verbs
that describe internal states (e.g. think, believe, feel) cause participants to supply
sentence completions such that the pronoun refers to the NP1 ; on the other hand,
verbs describing actions trigger responses that refer to NP2 (Caramazza et al.,
1977).
While past research has avoided what are vaguely referred to as pragmatic influences, some research has sought instead to manipulate them. Pragmatics is
often used indiscriminately, without definition (or even reference to a definition)
by a large number of authors. Moreover, when the term is used, is quite often used
as a last resort, to appeal to something outside of the “core” linguistic system (for
an extended discussion, see Levinson, 1983). The term suffers the same fate in
psycholinguistics. Essentially, when authors appeal to pragmatics, it is as an amorphous, functionally opaque entity, above the sentence level, with great explanatory
power. Pragmatics is essentially a code word for context. The context residing
above the sentence level includes discourse structure and general knowledge.
Research on plausibility seeks to better understand the influence of shared/general
knowledge on language. One common way to model the effect of general knowledge on pronoun resolution is to use plausibility ratings (Pickering & Van Gompel,
in press). While these ratings are themselves opaque (Connell & Keane, 2004),
they at least offer some measure of the influence of general knowledge of actions
and actors on pronoun resolution. However, recently there has been an attempt
to qualify plausibility.
The notion of plausibility seems vague at first blush. In spite of this, it has
received attention in psychology and has played an important role in distinguishing
between different models of sentence parsing (Pickering & Van Gompel, in press).
Plausibility ratings endeavour to tap what participants perceive to be the likelihood
of events in real life. Participants generate scores by rating the plausibility of
events in questionnaires, and these plausibility ratings in turn predict participants’
understanding of events at a conceptual, non-linguistic level (Connell & Keane,
2004).
Connell and Keane (2004) found causality and plausibility to be closely related.
They performed a study in which participants rated the plausibility of conjoined
propositions whose relation was causal, attributal, or temporal succession or unrelated. (Note that these map one-to-one to Kehler’s cause–effect, structural and
contiguity relations, discussed in §1.7.2.) Although there was some overlap between the different relations, the plausibility ratings on average were such that
CHAPTER 1. THEORETICAL INTRODUCTION
10
causal > attributal > temporal succession > unrelated.
Although causality tended to be rated highly, Connell and Keane (2004, 2006)
attribute the relation between plausibility and causality to factors underlying plausibility, not causality. Connell and Keane (2004) state strongly that their findings
measured plausibility independent of language. They draw this conclusion because
they found no interaction between word co-occurrences and the plausibility of a
sentence that contained them. This contrasted with their findings on naturalness
judgments, which strongly correlated with co-occurrence. Moreover, there is past
evidence that participants judge naturalness separately from plausibility, a finding
that Connell and Keane (2004) replicated. Participants in their experiment rated
novel scenarios for both linguistic naturalness and plausibility, and the measures
were found not to be correlated.
What is clear is that plausibility has an effect on linguistic processing, and this
effect is early. This has been shown by studies where the plausibility of events temporarily over-rides syntax in language comprehension (Hoeks, Prinsen, & Stowe,
2003; Hoeks, Stowe, & Doedens, 2004). The ERP signal observed for implausible scenarios has also been observed for non-linguistic phenomena (Sitnikova,
Kuperberg, & Holcomb, 2003), suggesting that linguistic plausibility and visual
plausibility are the same (although naturalness and plausibility are not).
1.6
Working memory
The influence of memory on activation in language has been suggested as early as
Chafe (1973), who cited it as the root of the given–new distinction. More recently,
computational models of working memory activation have surfaced (Lewis et al.,
2006; Lewis & Vasishth, 2005). These models draw on general principles of working
memory and apply them to syntactic parsing. Most of the work has focused on
the retrieval of nouns at a verb, although there is no reason why this work should
not be possible to extend to the retrieval of pronoun referents.
What is involved in working memory activation? Lewis and Vasishth (2005)
cites research showing that the short-term memory is limited to roughly three
items. The magic number seven plus or minus two, the idea that most people’s
working memory is between five and nine digits (Miller, 1956), appears to have had
some help from phonological rehearsal (Lewis & Vasishth, 2005). Effects of this
maximum are referred to as storage-load effects (Lewis & Vasishth, 2005; Lewis et
al., 2006).
Furthermore, Lewis and Vasishth (2005) suggest that retrieval becomes harder
when other entities in memory resemble it, a property referred to as interference.
This interference can occur both with entities before the target – Similarity-based
encoding interference – and entities after the target – Similarity-based retrieval
CHAPTER 1. THEORETICAL INTRODUCTION
11
Melissa knew that the toy from her uncle in Bogotá arrived today.
Sarah1 invited Vivian2
over for dinner, and Evan gave her2 the address.
Figure 1.1: Retrieval of a noun at a verb top, and, for comparison, pronoun resolution bottom. Retrieval in both cases is similar, and thus the bottom should be
subject to the same memory effects as top. Top from Lewis, Vasishth and van
Dyke 2006.
interference. Melissa causes encoding interference because she can potentially
arrive, and her uncle causes retrieval interference for the same reason.
Locality effects result from the temporal distance between the target and the
point of retrieval. Evidence of Anti-locality effects also exists. For example, the
distance between the head noun of the subject phrase and the main verb did not
create locality effects. Lewis, Vasishth and Van Dyke (2006) suggest this occurs
because their is a minimal amount of memory load at this point.
Anti-locality effects could perhaps be better described with a non-memory
theory. For example there is plenty of evidence that context will raise the likelihood
of a given referent from an early stage (Van Berkum, in press; Hagoort & Van
Berkum, 2007; Hoeks et al., 2004; Hoeks, Vonk, & Schriefers, 2002). The status of
the other memory effects with regards to pronominalisation certainly seems worthy
of investigation. The question remains, is working memory applicable to pronouns
given their stronger requirements on the identifiability of the antecedent?
The same factors could logically affect pronoun resolution. Assuming that
Vivian is the name that needs to be recovered, encoding interference would occur
because Sarah is a female proper noun, retrieval interference could occur because
Evan is a proper noun. Inclusion of many referents in the discourse model has
been shown to make processing a pronoun difficult, according to ERP results
(Nieuwland & Van Berkum, 2008).
1.7
Discourse structure accounts of pronoun resolution
The referent of a pronoun is often not within the same sentence, but the sentence
before or even several sentences before, so it is extremely important to examine
pronouns from a discourse structure perspective. The types of discourse structure
CHAPTER 1. THEORETICAL INTRODUCTION
12
we examine here are information structure and coherence structure. Centering
theory, centering OT, and the continuing topic/contrastive stress approach all
make use of the notion of topic, an information structure phenomenon, to predict
the referent of a pronoun. Other theories predict the pronoun using the coherence
structure of a text, the way two propositions are related.
Although information structure has had some success in determining pronoun
reference, recent experimentation on coherence relations suggests that coherence
structure is also important. This would mean that it is not the structure of the
sentence that is important, but rather the relationship between the propositions
that two sentences express. There is evidence of this both in the work of Kehler
and colleagues and the experiment in Chapter 2 (Taylor, Stowe, Redeker, & Hoeks,
2009).
1.7.1
Information structure
Information structure is the way that sentences are formed so that information
from a previous sentence is integrated in a future sentence in a comprehensible
way. It arranges the information so that previous information integrates smoothly
with new information. It takes into account what already has a representation in
conversants’ minds, what needs to be represented and ties the two together. It
organises the information to maximize comprehensibility.
Centering theory tries to describe the structure of information with topic structure. The two inculcations of centering theory below attempt to describe how people understand topic structure. Contrast, or kontrast are also potentially important
in pronoun resolution. De Hoop’s discussion of contrastive stress and continuing
topics points to some important future directions, although lacks a clearly defined
definition of topic or contrast. Vallduvı́ and Vilkuna (1998) provides a clear definition of kontrast (their version of contrast), which is further elaborated in a very
specific way by Steedman (2007).
Centering theory
Centering theory, proposed by Gordon and Hendrick (1999, 1998), uses topic structure as the driving force behind pronoun resolution. Their prominence-based approach to pronoun resolution proposes that use of a pronoun is not dependent on
syntax, but rather on whether the referent had been previously mentioned5 .
According to centering theory, each sentence has a backward-looking centre and
forward-looking centres. The list of forward-looking centres comprises all entities
5
This echoes Mel’čuk’s (1988) theoretical treatment of pronouns, which argues that it is not
governor-dependent relationships, but rather linear order in the surface-form which counts in
pronominalisation (pp.80–81).
CHAPTER 1. THEORETICAL INTRODUCTION
13
in the discourse, whereas the backward-looking centre, for example a pronoun,
refers to an entity from the previous sentence. In our sentence (1) (repeated
here as (8)), NP1,2&3 are forward-looking centres, and the pronoun, him, is the
backward-looking centre.
(8)
John called Geoff, and Steven emailed him.
NP1 Verb NP2 and NP3 verb
Pro.
Centering theory states that both the syntactic focusing structure and pronoun
usage are determined by discourse prominence. Gordon and Hendrick (1999) state
that in a normal active sentence the subject is the most prominent NP. It predicts
that the backward-looking centre (most frequently a pronoun) will refer to the
focused NP (Gordon & Hendrick, 1999). For this reason, subject preference is not
necessary as an independent principle, but can be tied to the work on discourse.
Their data indeed suggest that pronouns referring to something that has been
mentionedanaphoric, are easier than pronouns that refer to something that will be
mentioned, cataphoric. In the results reported, (9a) is judged correct more often
than (9b) or (9c):
(9)
a.
b.
c.
John’s roommates met him at the shop. .94
John’s roommates met John at the shop. .37
His roommates met John at the shop. .23
(Gordon & Hendrick, 1999, p. 83)
Sentence (9b) was hypothesized to be more difficult to comprehend because participants interpret the second “John” as a new entity. In other words, the repetition
of a name was interpreted as potentially being a new referent. Gordon and Hendrick (1999) hold that it is discourse prominence, not purely syntax, that controls
pronoun usage.
Discourse prominence theory predicts that in (1) on p.3, NP1 is the referent,
because the subject is the focus of a basic active sentence.
Centering OT and accent marking
Below we present a section on centering OT, an account that attempts to explain
pronoun resolution and accenting. As an initial step we explain some recent work
on givenness and accenting that is used in centering OT.
Schwarzschild’s (1999) explanation of accent marking shares some similarities
with the section above, in that he includes a long discussion of accent—his explanation is in large part a formalization of Chafe’s (1973) notion that accent is
closely tied to givenness. Additional differences between Schwarzschild’s account
and Chafe’s are that Schwarzschild uses entailment as a motivating factor in accent
CHAPTER 1. THEORETICAL INTRODUCTION
14
and semantic OT. Semantic OT has been much used recently, so we have included
a “Quick and dirty guide to OT” as appendix A, p.46–48 for those unfamiliar or
in need of a brush-up.
For Schwarzschild, something is prosodically prominent if it is focus-marked
and focus-marked if it is not given. Importantly, Schwarzschild’s formalization of
givenness states that information is given if and only if it is entailed, or mentioned
explicitly by the text before it. Schwarzschild uses the following constraints:
GIVENness If a constituent is not F-marked, it must be given.
AvoidF Do not focus more than is necessary.
Foc A focus-marked phrase contains an accent. Where a focus- marked node is
an focus-marked node that is not dominated by another focus-marked node.
HeadArg The internal argument of a phrase is more prominent than its head.
Schwarzschild (1999) ranks them:
GIVENness, FOC AvoidF HeadArg
While Beaver’s (2004) account follows Schwarzschild in using OT, it differs by
employing centering, not givenness as the engine behind focusing. As discussed
above, centering suggests that each sentence has forward-looking centres and a
single backward-looking centre. The backward looking centre refers to elements
mentioned in the previous sentences, and forward-looking centres refer to elements
that are being introduced. These principles are encoded as OT constraints.
Agree An anaphor must agree in gender and number with its referent.
Disjoint An anaphor is not dominated by the same verb as its antecedent.
Pro-Top The topic of a sentence is pronominalised.
Fam-Def NPs that are definite are familiar, no new info is provided about them.
Cohere The topic of the current sentence is the same as the previous sentence.
Align The topic is the subject of the sentence.
Agree Disjoint Fam-Def Cohere Align
This hierarchy predicts that the subject of a sentence is the topic, and that the
subject of a preceding sentence is pronominalised in a subsequent sentence. This
gives a parallel reading, if there is a pronoun in subject position. An additional
constraint, is added, *Block.
*Block A sentence must be optimal for both the speaker and the hearer.
CHAPTER 1. THEORETICAL INTRODUCTION
15
This constraint assumes that a pronoun saves the speaker processing, but increases processing for the hearer. This makes sense in light of the suggestion by
Van Berkum et al. (2007) that speakers make use of an easily produced, high
frequency pronoun, because these pronouns are easier to use than more difficult,
low-frequency name. Moreover, pronouns are used more frequently by those with
processing deficits (Hendriks, Englert, Wubs, & Hoeks, 2008).
From the point of view of the hearer, however, the pronoun is more difficult to
retrieve, the greater the number of referents that have been mentioned. Block*
effectively imposes the speaker’s knowledge of what the hearer is thinking. In
order to make the pronoun more salient, it is accented, thus Beaver adds AvoidF.
If a pronoun refers to a more recently mentioned referent, then it will appear in
neutral, unaccented form. If it appeals to a less recently mentioned referent, it
is focused and therefore accented. In this way, accenting a pronoun switches the
referent of a sentence6 .
In this centering OT account, Beaver (2004) predicts that pronouns will be
resolved to the subject, if they are not, they will be stressed. Applying Beaver’s
model to the two-clause conjoined sentences described above without intonation
predicts that when there is a single pronoun in a sentence it refers to the subject of
the previous sentence. Centering OT predicts that an ambiguous (object) pronoun
will refer to the subject of the previous sentence, but if accented will refer to the
next least oblique noun phrase in the previous sentence (notions that are inherited
from centering theory, that are not expressed in the constraints). So if a pronoun
is accented, and the previous sentence is ditransitive, then the pronoun will refer
to the direct object (all other things being equal).
6
It seems slightly strange, however, that Beaver has neither included nor replaced
Schwarzschild’s HeadArg and FOC constraints. Without these two constraints, there is no
motivation to include an accent (AvoidF). It could be the case that Beaver envisions these two
constraints as subsumed under *Block, with the idea that a speaker knows what a hearer needs
focused. While integrating theory of mind with language comprehension is tempting, *Block
is perhaps too powerful.
What is particularly attractive about *Block, and bidirectional OT in general, is that they
manage to include theory of mind in the set of ranked constraints. However, one must be
judicious and restrained in the application of theory of mind, or one risks attributing any detail
of language to it. For example, one normally forms the English plural by adding an /–s/. This is
done because it is a rule of the language, but this is also done because the speaker knows that if
the hearer receives a noun unmarked with an /–s/, he or she will understand that the noun has
singular number. So rather than proposing morphological or semantic constraints on the use of
/–s/ we could simply appeal to *Block
CHAPTER 1. THEORETICAL INTRODUCTION
16
Continuing topic and contrastive stress
Whereas the previous account takes centering, or “discourse prominence” as primary, the proposal by de Hoop (2004) hinges on the use of topic 7 . De Hoop calls
Continuing Topic the interpretive counterpart of Beaver’s Pro-Top. She suggests that there are two constraints governing pronoun resolution, Continuing
Topic requires the pronoun to be the topic of the previous sentence. She also
postulates Contrastive Stress: “Stress on a pronoun indicates a rhetorical
relation of contrast” (p. 36). In de Hoop’s hierarchy, contrastive stress is ranked
above continuing topic. Her notion of topic does not take into account parallelism,
the property that something will be coreferential with something in a similar linear position. She predict that for (1) the pronoun refers to NP1 , and refers to
the same pronoun with accenting. There is a slight semantic difference for the
accented version, in that it would invoke a set of things that it contrasts with.
Nevertheless, her ranking does not make correct predictions for the sentence pairs
used in Chapter 2 (Taylor et al., 2009).
While de Hoop’s analysis initially appears flawed in light of parallel sentences
with object pronouns, her proposal to rank contrast above topic is an important
insight. In sentences where plausibility plays no role, her ranking could be salvaged,
with a slightly more complex notion of topic, or prominence. Contrast does appear
to over-ride other exigencies on pronoun reference.
“Kontrast”
Kontrast was proposed by Vallduvı́ and Vilkuna (1998) and refined by Steedman
(2007). Vallduvı́ and Vilkuna (1998) introduce the term kontrast, which they claim
occurs when an element evokes a set that the element is then compared with. So in
the sentence Jeff went to Marina’s party, Jeff is contrasted with other individuals
who could have potentially attended the party. Steedman’s (2007) contribution is
the idea that the intonation also expresses information about the speaker’s beliefs,
the hearer’s perceived beliefs and the thematicity of the element. We discuss this
further in the conclusion of Chapter 2.
In Dutch and English, kontrast is expressed with prosody, but in other languages syntax or morphology are used (Vallduvı́ & Vilkuna, 1998). For example,
Vallduvı́ and Vilkuna (1998) claim that Catalan uses syntax and Japanese uses
morphology. We would predict that although the surface realization of kontrast in
one language may be different from the surface realization in another, the effect of
kontrast on reference interpretation should be comparable across languages. This
idea that information structure takes a set of packing instructions and creates a
linguistic form is particularly important when looking at coherence relations and
7
De Hoop does not give a definition of topic or refer to one
CHAPTER 1. THEORETICAL INTRODUCTION
17
information structure. Information structure takes aspects of coherence and makes
them linguistically apparent.
Kontrast is a nice illustration of the difference between information structure
and coherence relations. Coherence relations describe the relationship between two
clauses. Contrast or kontrast is a form of coherence relation, because it describes
the relationship between two clauses. The information is then structured with
prosody, syntax, or morphology to express kontrast. Our discussion of coherence
relations in the next section will help to make the distinction between the coherence
relations and discourse structure clearer.
1.7.2
Coherence relations
Coherence structure has also been proposed as an organising principle in discourse
in addition to information structure. Whereas information structure acts to make
the relationship between topics clear and the transitions smooth, coherence structure is the relation between propositions. Taken together, the relations between
different ideas form a structure. Kehler and colleagues develop coherence relations
from the notion that there are essentially three ways to relate ideas together: resemblance, cause–effect, and contiguity (Kehler et al., 2008; Kehler, 2005; Kertz
et al., 2006b, 2006a; Rohde et al., 2006). They propose that cause-effect relations
require for there to be some causality between two sentences (10)
Cause–effect:
(10)
Jeff assaulted Frank, and Steve arrested him.
Conventionally, the person who commits the assault is arrested. The first clause is
thus the cause, and the second is the effect. One further type of relation between
sentences is that of contiguity. In contiguity-related sentences, the one sentence
posits a set of circumstances, and the other an action within that set of circumstances (11):
Occasion:
(11)
Jeff went to the market. He bought vegetables.
Kehler et al. (2008) suggest that the coherence relation between two propositions
has important implications. Kehler et al. (2008)shows that pronouns linked by
cause–effect (in particular the sub-category result) may refer to an antecedent
which is not in the same position in the sentence. In (12) below, in the first
sentence, the pronoun refers to the first noun phrase (Nicole). In (13) it refers to
the second noun phrase (Vivian).
(12)
Nicole1 yelled at Charlotte2 , and Isaiah pleaded with her1 to relax.
CHAPTER 1. THEORETICAL INTRODUCTION
(13)
18
Sarah1 invited Vivian2 over for dinner, and Evan gave her2 the address.
This is not possible with resemblance-related pronouns, where the pronoun must
be in the same place as its antecedent.
In the resemblance relation parallel, the two utterances are related to each
other such that common properties of each are either shown to be similar (14) or
dissimilar (15):
Resemblance:
(14)
Jeff called Frank, and Steve wrote him.
(Parallel)
(15)
Jeff called Frank, but Steve wrote him.
(Contrast)
In (14), the speaker is likely stating that both events involved communication,
whereas in (15), the speaker is likely stating that while both events involved communication, the exact type of communication is different. According to Kehler
it is not necessary for two paralleled expressions to have the same grammatical
structure, but rather the same linear ordering.
Stating parallel information in a parallel way as a rhetorical device can be
traced back at least as far as Cicero. It has been described more recently in Kuno
(1974) and Prince (1981). Notably, to Prince and Kuno this is not a coherence
relation, but rather a principle governing the arrangement of discourse entities.
This arrangement is implicitly understood between speaker and hearer. This argument is strengthened by a considerable work in psycholinguistics showing that
the argument structure of one verb primes the argument structure of subsequent
verbs (Van Gompel et al., 2006, discussed above in the section on syntax).
Furthermore, two propositions can be at once related by cause and effect,
occasion and parallel (16):
(16)
a.
b.
Frank wanted to buy Sarah an ice cream cone.
He took her to the ice-cream parlour.
In this example, (16a) causes (16b), they follow each other temporally, and they
are arranged in parallel. It seems that there could be a strategy of arranging things
in a parallel way as a memory aid. It seems sensible to arrange entities parallel
when possible, irrespective of the coherence relation.
Hendriks (2004) proposes that adding the idea of contrastive topic into the
coherence relation account is beneficial. Hendriks also suggests that in fact the
different coherence relations constitute an epiphenomena, resulting from multiple
language-general processing constraints. Hendriks hypothesizes that put in the
form of OT, cause-effect relations and occasion relations are constraints, whereas
parallel results from discourse conventions. In particular, the topic is at the beginning of the sentence and remains constant from sentence to sentence.
CHAPTER 1. THEORETICAL INTRODUCTION
19
The difference observed between cause–effect and resemblance stimuli in object
pronouns described above could also come from differences in plausibility. In
particular, cause–effect sentence pairs are stronger than contiguity pairs because
the plausibility connecting them is greater. Moreover, there is a very close relation
between the causality of a sentence and its perceived plausibility (Connell & Keane,
2004, 2006). In our experiment we observed that it is common for one referent
to be very plausible, and the other very implausible for cause–effect sentences
where the pronoun has two possible antecedents. Thus the choice of antecedent
is determined by general knowledge, and general knowledge is not just knowledge
of causality but also of what constitutes a plausible set of events. Plainly stated,
plausibility and causality appear to be strongly linked.
Plausibility may also help to describe the difference between the parallel relation and occasion relation. Occasion relations describe scenarios with strong
plausibility and no actual causality. In sentence (11) above, it is quite plausible
that someone at the market should buy vegetables, but going to the market does
not cause one to buy vegetables. So which factor is important in pronoun resolution? Is it actually causation, or simply plausibility, or are both important for
pronoun resolution?
The plausibility interpretation discussed above has further ramifications. Kehler
et al. (2008) suggest that implicit causality/consequentiality effects are due to coherence relations. On seeing a conjunction like because, the probability of the
referents is computed given the verb and the coherence relation. It seems possible
that the coherence relations act as a bridge to the scenario. Information about the
coherence relation allows more specific inferencing about the scenario. It is unclear
whether it is simply the interaction of the verb and an indication of coherence relation that allows this prediction, or whether there is a necessary involvement of
knowledge about the scenario.
1.8
1.8.1
Conclusion
Interim Summary
We have attempted to show the breadth and depth of the pronoun reference literature. It is possible – from our perspective even likely – that there are several
factors at work in pronoun resolution; likely several accounts together will need to
be taken together to create a fully functional model.
Plausibility in particular should be researched because of its close relationship with coherence relations. Understanding plausibility is not just trying to
eliminate a lurking variable; it may be at the centre of pronoun resolution. Plausibility may be responsible for the predictive abilities of coherence, or vice-versa—
CHAPTER 1. THEORETICAL INTRODUCTION
20
experimentation is necessary.
The working memory model has not yet been tested in relation to pronouns.
The predictions it makes regarding the distance and/or time between the referent
and the pronoun are interesting from a processing perspective. Working memory likely delimits the available referents for reference, in concert with situational
knowledge.
Centering theory too makes predictions with a concept of activation, although
centering theory cites discourse prominence as an important factor. Either of
these may be an important factor, although both could not, because they make
diametrically opposing predictions. And of course information structure is not just
restricted to topicalisation, kontrast merits more in depth study
Coherence relations need to be more thoroughly investigated, especially to
distinguish them from information structure and plausibility, or to understand the
mutual influence.
1.8.2
Future directions
With so many possibilities, it is necessary to carve up the territory; we do this
by looking at pronouns in context–rich and context–poor environments. In the
chapters to come, we look at the effect of context on pronouns: we look at pronouns in a context-rich scenario and in scenarios where there is a lack of context.
In Chapter 2 we look at the roles of the contextual factor coherence relation; because this chapter is submitted as a stand-alone paper the introduction overlaps
with the theoretical exposé presented above. We then examine the time course of
pronoun resolutions in sentences where there is an initial lack of information in
Chapter 3. Finally we propose a series of experiments in Chapter 4. The results
from Chapter 3 suggest further research into the resolution of pronouns in initially
context-impoverished environments, and the results from Chapter 2 propose further experimentation on context-rich sentences that manipulate accent, coherence
relations, and plausibility. Taken together these two approaches will enrich our
knowledge of prosody in pronoun resolution.
Chapter 2
Experiment 1
Information structure,
plausibility, coherence and accent
as determiners of pronoun
reference
Pronoun resolution has not yet been accounted for, in spite of the fact that wellstudied languages like English, Dutch, German, and French are amok with them.
Prosody, the tune of a sentence, has also remained mysterious until late. These two
factors conspire as accented pronouns, and findings about them have wide reaching consequences. Understanding accented pronouns informs studies in ambiguity
resolution in language, sentence processing, discourse structure, information structure, and the interface of general knowledge and language, amongst others.
Although accented pronouns interact with a wide variety of linguistic phenomena, many theories of pronoun resolution have suggested pronoun-specific heuristics, which do not generalize well to the understanding of ambiguity at large. Furthermore, previous psycholinguistic models are of limited predictive value when it
comes to determining what a pronoun refers to. In response to the limitations of
previous models, and the results of our study, we elaborate a model that draws on
general principals of language structure and correctly accounts for the data.
2.0.3
The Resolution of Ambiguous Pronouns
Pronoun research in psycholinguistics has focused on parallel sentences with an ambiguous pronoun (Garvey et al., 1975; Caramazza et al., 1977; Grober, Beardsley,
& Caramazza, 1978; Crawley et al., 1990; Smyth, 1994; Gordon & Hendrick, 1998;
21
CHAPTER 2. EXPERIMENT 1
22
Venditti, Stone, Nanda, & Tepper, 2002; Wolf, Gibson, & Desmet, 2004; Smyth,
1994; Kehler et al., 2008; Venditti et al., 2002). These are typically sentences with
two clauses as shown in (1):
(1)
John1 saw Jeff2 and Stephen called him.
NP1 VP NP2 and NP3
VP Pro
In order to be a parallel sentence, each of the clauses must have a subject (NP1 ,
NP3 ), a verb verb (VP ) and be conjoined by and. Many of the initial attempts to
explain coreference in these sentences creatively exploited syntactic theory. Early
syntactic accounts suggested that the grammatical role was most important to
pronoun reference. Subject Assignment states that all pronouns refer to the subject
of the sentence (Crawley et al., 1990). Parallel preference mandates that a pronoun
will refer to whatever entity has the same grammatical role in the previous sentence
(Smyth, 1994; Crawley et al., 1990); a subject pronoun refers to the subject of the
sentence, an object pronoun to the object.
Yet another theory, the extended feature match hypothesis proposes that the
degree of match between the grammatical roles in the first and second sentences
determines what the referent is, not the grammatical role. For example, in sentences (2a) and (2b) taken from Smyth (1994, p.202) the same idea is expressed
and the same entities are involved, but him has a different grammatical role in
each sentence.
(2)
a.
b.
The carpenter gave the plumber an invoice, and the electrician gave
him a cheque.
The carpenter invoiced the plumber, and the electrician gave him a
cheque.
According to Smyth (2004), him is an object in the first sentence, and the subject
of a subordinate clause in the second sentence. The differences in grammatical role
parallelism between (2a) and (2b) have important consequences. If the two clauses
have matching grammatical roles, a pronoun refers to the NP in the previous clause
which shares the same grammatical role, as in (2a). However, when two clauses do
not match in terms of grammatical roles, the pronoun will refer to the subject of
the previous clause, as in (2b). Notably, the authors excluded sentences that were
felt to be “pragmatically biasing”. Whereas it is apparent that extra-linguistic
factors play a role in pronoun attribution, the term pragmatic factors has been
consistently been applied to those sentences where pronoun reference results from
world knowledge, and is beyond the scope of whatever theory is being proposed
(Smyth, 1994; Caramazza et al., 1977; Garvey et al., 1975).
One advantage of discourse/coherence based theories is that they account for
pragmatic factors. Recent work by Kehler et al. (2008) has shown that coherence
CHAPTER 2. EXPERIMENT 1
23
relations are much stronger predictors of pronoun referent than syntactic factors.
Indeed, coherence relations predicted the referent of a pronoun upwards of 90%
of the time while the above-mentioned theories predicted the referent just above
the level of chance (Kehler et al., 2008). Coherence relations accomplish this
by appealing to the relation between ideas, and not just to syntax. One major
advantage of this approach is that it can account for the ”pragmatic factors” that
have been so carefully avoided by previous approaches. This is done by describing
the relation between two propositions in a discourse (see Kehler, 2002).
For the sake of space and clarity, we discuss three Coherence Relations in detail,
parallel, result and occasion. Parallel relations are used when a speaker or writer
wishes to express the similarity between participants, actions, or participants and
actions in two propositions. Sentence (1) above is a parallel relation: the similarity
between Stephen’s participation in the action and Stephen’s action (seeing) with
John’s action (calling) is highlighted for the person who hears/reads the sentence.
Sentences related by parallel can be paraphrased as X, and similarly Y.
Result relations are used to express a causal relationship between two events,
where the first event causes the second event. In the sentence Jessica helped Angela,
and she was thankful, the event expressed in the first clause results in the event in
the second clause. The relation between the two sentences can be paraphrased by
X, and as a result Y.
Occasion are related by temporality, such that X represents the state of events
before Y, although unlike result relations there is no causal relation between the
two. An example would be Steve went to the store and bought a book. Unlike in
a result relation, going to the store does not necessarily imply that someone will
buy a book.
In addition to predicting the referent of pronouns, Kehler et al. (2008) claims
that coherence relations are predictive of accent distribution. Phonetically, a word
is prosodically accented if it is made acoustically prominent. Prosody has a different motivation in different coherence relations: accent has different motivation in
parallel sentences than in result sentences, but its effect on pronouns is predicted
to be the same. Accent is said to occur in parallel sentences to indicate which
constituents (i.e., NP, VP) are being related. In John emailed Jeff, Stephen
called Lawrence, the parallel between Stephen and John is being asserted. In result relations, however, accenting occurs on constituents that are improbable given
the situation being described – the speaker draws the listeners attention to it. In
spite of the different functions of accent, Kehler et al. (2008) and Kehler (2005)
predicts that accenting a pronoun in parallel or result sentences have the same
effect, irrespective of Coherence Relation (3):
(3)
John emailed Jeff, and...
a. Stephen called him.
No-accent
CHAPTER 2. EXPERIMENT 1
b.
Stephen called him.
24
Object–accent
For both parallel and result, when a pronoun is stressed, its referent is switched.
In (3b) this switch is from Jeff to John Importantly, they provide a different
motivation for the switch in each Relation. In parallel, nouns are stressed when
their relation to something in the previous clause is emphasized. Since it does
not make sense to stress something’s relation to itself (as would be the case with
pronouns), it instead switches the referent. In result relations, words that are
improbable given world knowledge are stressed.
Venditti et al. (2002) tested the interaction of coherence relations and accent
using the Visual World Paradigm. Prosody was used to change the referent of a
pronoun. More particularly, eye-tracked participants looked at a picture as they
heard a short narrative of like (4):
(4)
a.
b.
c.
d.
The zebra and the pig wanted to wash the car together.
The zebra put a bucket of soapy water next to the pig near the front
of the car.
(i) Then he got out some sponges.
(ii) Then he got out some sponges.
And together they started washing the hood and the fenders.
p.3
Two characters were introduced in the first sentence (4a). In each of the critical
conditions was a sentence starting with then (4c). Their hypothesis is that the
coherence relation between (4b) and (4c-i) is occasion, and between (4b) and (4c-ii)
is parallel.
This auditory, on-line experiment provides evidence that accenting switches
the referent of a pronoun under some conditions. They found that when asked
which character was represented by he in (4c) participants looked at a different
referent in the accented condition than in the neutral condition.
There is, however, a theoretical problem with Venditti et al. (2002). They
state that their stimuli are ambiguous between the relations parallel and occasion.
But then is far from being a neutral conjunction. It implicitly encodes temporal
adjacency: X then Y implies that X occurred first and Y followed. In comparison with then, and is Coherence Relation-neutral; the Coherence Relation is thus
biased towards occasion. and is therefore a better choice of conjunction for an
experiment testing parallel relations. Hence, in the experiment we present below,
the stimuli are either unambiguously result or unambiguously parallel.
Our experiment constitutes a necessary clarification of the results in Venditti
et al. (2002). Unlike this previous study, our stimuli were unambiguously of one
coherence relation or another, and two coherence relations were tested. And, as
mentioned, while subject pronouns have most often been experimented on, we
CHAPTER 2. EXPERIMENT 1
25
experimented on object pronouns. For these pronouns, gapping is not an option
and thus we could be more certain that pronominalisation was the most natural
choice.
In the experiment below, we tested whether the existence or absence of accent
on an object pronoun changes the referent. Kehler et al. (2008) suggested that
accent will change the referent of an ambiguous antecedent in parallel and result
sentences. This experiment constitutes an empirical test of intuitions presented
by Kehler et al. (2008) regarding the interaction of Coherence Relations, Accent
and Pronoun Reference.
2.1
2.1.1
Method
Participants
Twenty-seven native speakers of English (nine women, one left-handed) residing
in Edmonton (Western Canada) participated in the experiment.
2.1.2
Materials and design
The experiment used a 2 × 3 design, taking Coherence relation (result, parallel )
as one factor, and Accent (No–accent, Subject–accent, Object–accent) as another.
The Participants heard only one version of a given stimulus; there were 60 sentences
in total, presented in a pseudo-randomized manner 1 . To assure that parallel
-related sentences could not also be result sentences, the second verb in parallel related sentences was a near synonym (e.g. phone–email), following Kehler et al.
(2008).
We thought it possible that participants would adopt the strategy of picking
the second NP heard as the referent in No–accent conditions and the first NP
heard as the referent in Object–accent condition. To prevent this strategy, there
was a nested factor in the result condition. This factor, NP, had two levels, NP1
and NP2 . The stimuli and referents are illustrated in Figure 2.1.
There were 18 items in NP1 and 12 in NP2 . In NP1 sentences, reference bias
is towards the first NP, and for NP2 towards the second NP.
Additionally, following Kehler et al. (2008) we tested Smyth’s EFMH hypothesis as a nested factor within the parallel condition. This factor had the levels
Partial and Complete. As per the the discussion above, the two clauses either had
exactly the same grammatical roles, or there was an additional grammatical role in
one of the clauses. In table (5) we see that the grammatical roles are not parallel,
1
(Stimuli available from http://ryanchristophertaylor.googlepages.com/scripts%2Cusefulsoftware%26links)
CHAPTER 2. EXPERIMENT 1
26
Nicole1 yelled at Caroline2 , and Isaiah pleaded with her1 to relax.
Sarah1 invited Vivian2
over for dinner, and Evan gave her2 the address.
Figure 2.1: The referent of pronouns in NP1 and NP2 sentences without accent.
because in the Complete condition the syntactic structure of both clauses is the
same. In the Partial condition the pronoun is expressed with a preposition.
(5)
a.
b.
Sally called Monica, and Robert summoned her.
Amanda wrote Alisha, and Stevie made amends with her.
Complete
Partial
Some sentences were modified versions of stimuli from Smyth (1994), and Kehler
et al. (2008). The proper names in the first sentence had the same gender so that
there was a potential ambiguity in the referent of the pronoun. The proper names
used were chosen to be unambiguously male or unambiguously female. All items
were recorded by a native speaker of English for presentation over speakers or
headphones.
2.1.3
Procedure
Participants rated which person named in the previous sentence they felt to be
the antecedent of the pronoun using a seven-point Likert-like scale.
To lessen the memory load on participants, the names of the characters of the
first two sentences were included on the response sheet. This allowed participants
to deal with the large number of names. Prosodic naturalness ratings from 1–5
were also collected. Participants were asked to go with their first intuition and not
to over-think their answers.
The instructions were read by participants and explained orally by the experimenter. Particular efforts were taken to assure participants that there was no
prescriptive rule and that their intuitions would not be judged for correctness, but
simply analysed.
CHAPTER 2. EXPERIMENT 1
2.2
27
Results
We present below a series of ANOVA analyses. To compensate for violations of
the sphericity assumption, we applied the Greenhouse Geisser correction where
appropriate. The original degrees of freedom are reported.
2.2.1
Prosodic naturalness
In order to ascertain that any results found in further analyses were not attributable to unnatural prosody, an ANOVA analysis was performed on the naturalness ratings of the items (Figure 2.2).
Figure 2.2: Analysis of prosodic naturalness. 1 = highly unnatural, 5 = highly
natural.
There was no interaction of coherence and accent (F1 (2, 5) = 1.7, p = .186;
F2 (2, 114) = 42, p = .657)) there were main effects of Coherence Relation (F1 (2, 48) =
37.0, p < .001; F2 (1, 57) = 9.5, p < .003) and Accent (F1 (1, 24) = 8.8, p < .01;
F2 (2, 114) = 34.1, p < .001). For Coherence the level result (3.9, SE = .1) was
higher ranked than parallel (3.7, SE = .1). For Accent, the No–accent condition (4.2, SE = .1) was ranked highest, followed by the Subject–accent condition
(4.0, SE = .2), with the Object–accent condition (3.2, SE = .2) ranked least natural.
CHAPTER 2. EXPERIMENT 1
2.2.2
28
Referent preference
We took the ratings performed by the participants, and averaged them by condition. This average was the preference for a particular NP as referent. The scores
were then analysed by items and by participants using a 2 × 3 ANOVA designs,
taking Coherence Relation (parallel, result) as one factor, and Accent (No–accent,
Subject–accent, Object–accent) as the other, shown in Figure 2.3.
Figure 2.3: Likelihood of choosing the preferred referent by Coherence Relation.
There was a main effect of Coherence Relation (F1 (1, 24) = 20.1, p < .001; F2 (1, 57) =
27.3, p < .001). The participants were more certain of the antecedent of pronouns
in result (1.8, SE = .1) than in parallel (0.9, SE = .1). There was also a main
effect of Accent (F1 (2, 48) = 43.0, p < .001; F2 (2, 114) = 104.6, p < .001).
Participants preferred the other NP as antecedent in Object–accent condition
(0.2, SE = 0.2) relative to the Subject–accent (2.0, SE = 0.1) and the No–accent
conditions (all ps < .001). The No–accent condition and the Subject accent condition, however, were not significantly different from each other (F s < 1). These
main effects were, however, qualified by the existence of an interaction between
Coherence Relation and Accent. This interaction was highly significant in both by
participant and by items analyses (F1 (2, 4) = 11.1, p < .001; F2 (1, 57) = 16.0, p <
.001).
Here too the relationship between the Object–accent condition and the other
two conditions was the locus of this effect. So, in both conditions we find a similar
pattern, which is more pronounced in the parallel than in the result condition.
CHAPTER 2. EXPERIMENT 1
29
To find out whether this is true, we compared parallel and result conditions at
each of the levels of Prosody. We found that there was a significant difference
between the parallel and result Relations in all Accent conditions: No–accent
(F1 (1, 26) = 14.4, p < .001; F2 (1, 57) = 6.9, p < .05), Object–accent (F1 (1, 26) =
56.1, p < .001; F2 (1, 58) = 41.7, p < .001) and Subject–accent (in the F1 but not
F2 analysis). (F1 (1, 26) = 6.0, p < .05; F2 (1, 58) = 2.3, p = .131).
Visual inspection suggested that although the difference between the Accent
conditions was significant in all three conditions, the difference was most pronounced in the Object–accent condition; This was confirmed by post-hoc pairwise
comparisons of the effect of Coherence Relation between the different levels of
Accent. All analyses taking Subject–accent and Object–accent as levels were statistically significant (F > 12, p < .001). When No–accent and Subject–accent
were taken as a factor alongside Coherence, the interaction of Accent and Coherence is not significant, showing instead a trend of interaction (F1 (1, 57) = 2.8, p =
.10; F2 (1, 26) = 3.0, p = .12).
In order to determine whether a referent switch had in fact occurred, we tested
the means of the different groups against the value zero, the central value of the
scale. Values significantly different from zero have therefore changed referent as
a result of Accent. All of the result conditions and two of the parallel conditions
were significantly greater than zero (all ts> 5.750, p < .001). Object–accent had,
however, reduced the mean of participant responses below zero in parallel (t1 (26) =
−3.2, p < .01; t2 (29) = −3.4, p < .01).
2.2.3
NP1 and NP2 verbs
We included the variable NP to make sure that participants were not blindly choosing the second NP in the no-accent condition and subject accent conditions. An
exploratory analysis indicated that the Coherence Relation parallel leaned strongly
toward NP2 reference, while result was split between NP1 and NP2 preference. Result sentences were predicted to pattern with the parallel verbs. We tested this
prediction with the nested factor in result, NP, which was analysed to see if differences between result and parallel could be ascribed to the NP that was referred
to by the second verb in the No-accent condition. To that end, we took NP1 vs.
NP2 (henceforth NP) as a factor in an ANOVA. Thus, NP was a factor on the one
hand, and Accent (No–accent, Object–accent, and Subject–accent) on the other,
yielding a 2 × 3 design (Figure 2.4).
ANOVA analysis showed that there was no main effect of NP for subjects or
Items (F < .15). There was a trend toward an interaction (F1 (2, 52) = 2.3, p =
.12; F2 (2, 56) = 2.7, p = .09). Bonferroni post-hoc tests found that in the Subject–
accent condition there was a trend toward the ambiguous pronoun being judged as
more clearly associated with the preferred NP in the NP2 condition (2.3, SE = .2)
CHAPTER 2. EXPERIMENT 1
30
Figure 2.4: Likelihood of choosing the preferred anaphor with different intonations
in the Result condition by preferred NP.
than in the NP1 condition (1.9, SE = .2)(p1 = .09; p2 = .10). No other post-hoc
results were significant (p > .26).
2.2.4
Grammatical role parallelism
Further assessing the correctness of the EFMH (Smyth, 1994), we included the
nested variable Grammatical Role Parallelism with the factors Complete and Partial within the Parallel condition. Combined with Accent, this gave a 2 x 3 design
(Figure 2.5. An ANOVA analysis comparing Complete (.9, SE = .1) and Partial
(1.0, SE = .1) found no main effect (F < .1). All values were tested against the
central value zero; this determined that the values were not significantly different
from each other, and Partial sentences had the same referent as their Complete
analogues (all ps less than .05).
There was a trend toward interaction of Grammatical Role Parallelism with Accent in the by subject but not by items analysis (F1 (2, 52) = 2.0, p = .143; F 2(2, 54) =
.9, p = .4). Post-hoc analyses indicate that this trend of interaction in the by subject condition was due to significant differences between The Object-accent Complete condition (−.9, SD = .2) and No-accent Partial condition (−.6, SE = .3;
F 1(26, 1) = 2.3, p < .141). All other F s < 1.6.
CHAPTER 2. EXPERIMENT 1
31
Figure 2.5: Likelihood of choosing the preferred anaphor with different intonations
in the Result condition by preferred NP.
2.3
Discussion
Our findings were not totally in line with the expectations of Kehler et al. (2008).
Whereas Kehler et al. (2008) predicted that both result and parallel relations
would respond the same way to pronouns, we found that it was in fact only in
parallel-related sentences that accented pronouns selected a different referent than
their unaccented analogues.
The prosodic naturalness rating showed that the stimuli used were possible and
natural sentences of English. We ensured that Grammatical Role Parallelism was
not responsible by including a nested factor within the parallel condition. This
demonstrated that with respect to Grammatical Role Parallelism, partially parallel
structures were interpreted no differently than fully parallel structures. We tested
the effect of accenting pronouns in the result condition that referred to the the
first or the second NP: Again we found no difference.
Furthermore, this experiment clarified the results of Venditti et al. (2002).
Because our experiment separated Coherence Relation from Accent, and examined
object pronouns, our results gave a more effective account of the interaction of
prosody and coherence relation. Accented pronouns in a parallel-related sentence
change referent, those in result-related sentences do not.
At first blush, the results suggest that coherence relations determine how pro-
CHAPTER 2. EXPERIMENT 1
32
nouns are interpreted, and how accented pronouns are interpreted. Accenting
object pronouns switched the referent in the parallel condition, but not in the
result condition. Again, this interaction was modified by neither grammatical role
parallelism, nor NP. These results demonstrate that accenting plays a different
role in parallel -related sentences than it does in result sentences.
While the coherence relations account proved to be the best of the previous
models, it failed to account for the accented result condition. The model introduced here is able to account for this difference, and offers greater parsimony.
2.3.1
An account that incorporates the conceptual level,
coherence and information structure
We advocate an approach based on information structure and plausibility. In
this account parallelism functions as described by Kehler et al. (2008), but the
distinction between cause–effect and occasion is not relevant. We suggest instead
that it is not causality or contiguity that is important, but the plausibility of
a candidate referent in the situation described. If one candidate is plausible in
the situation, and the other implausible, then the plausible candidate becomes
the referent of the pronoun. If both candidates are plausible, then information
structure determines the pronoun.
It is difficult to say whether the form of a parallel sentence is mitigated by
Coherence relations or information structure. Parallelism, as Kehler et al. (2008)
describes it, makes strong requirements on information structure. When something is Paralleled to something else, it appears in an analogous position; thus, the
information is structured to make the coherence relation clear. Separating Coherence from information structure is thus tricky and perhaps undesirable altogether.
Expanding the discussion, the relationship between the cognitive representation of
a situation (reflected in part by situational plausibility, coherence structure and
information structure) is very close. Attempting to separate coherence structure
from either of these is rather like trying to peel the latex from a balloon and have
it remain intact; the situation is described by the coherence structure, which in
turn constraints the information structure. What can be said, is that information
structure is important to pronoun interpretation.
2.3.2
Information structure and intonation
In line with incorporating information structure, we suggest that accenting is the
result of kontrast, as proposed by Vallduvı́ and Vilkuna (1998). Kontrast is a
feature of information structure which exists in every language, although it can be
expressed differently (see Vallduvı́ and Vilkuna (1998) for a review). According
CHAPTER 2. EXPERIMENT 1
33
to them, kontrast in English is indicated by a pitch accent (L+H*). Kontrasting
an expression causes a set of things like the expression to be generated, and the
accented word is contrasted with that set. So in the sentence Steve wanted to
go to the restaurant we imagine that the other individuals involved in making
the decision to go to the restaurant (the rest of the set) were not in agreement.
How then does this affect something with variable reference like a pronoun? For
accented pronouns, the set generated comprises the entities that the pronoun could
plausibly refer to.
A theory of information-structural intonation that accounts for the results has
been elaborated by Steedman (2007), whose theory in turn builds in part on ideas
the idea of kontrast put forward by Vallduvı́ and Vilkuna (1998). Steedman proposes that English intonation indicates whether information is thematic or rhematic, and whether the speaker and hearer are agreed. In Table 2.1 Steedman
predicts that accent is different if it is thematic than if it is rhematic. Accent is
different if the information is linked to previous information or not2 .
Table 2.1:
Accents classified by thematicity and speaker/hearer agreement (after Steedman, 2007)
Agreement
Thematic
Rhematic
+
L+H*
H*,(H*+L)
L*+H
L*,(H+L*)
Furthermore, the boundary tones (the tune of the sentence at a place where
a speaker pauses) indicates the speaker’s beliefs and what she believes about the
hearer’s opinion (Table 2.2) Importantly, this means that when the speaker says
Table 2.2:
Boundary tones classified by Speaker/Hearer Commitment(after Steedman, 2007)
Commitment
Speaker
Hearer
Boundary Tone
L, LL%, HL%
H, HH%, LH%
something she believes, but that is not agreed upon by both parties, this is indicated in the intonation.
2
The distinction between thematic information and old information is discrete but important.
Essentially the theme links an idea with a previous idea and the other information is the rheme.
Given information is something that has already been mentioned, whereas new information has
not. Thus something may both new and thematic or old and rhematic or any other combination
for that matter. Vallduvı́ and Vilkuna (1998), Mel’čuk (2001) and Bott (2007) give rationale
and discussion of these distinctions.
CHAPTER 2. EXPERIMENT 1
34
How then does this effect pronoun interpretation? On encountering a pronoun,
hearers know the gender of the referent and that it can refer to things that are
active. Entities that are arranged in a parallel manner are usually paralleled to
one another, but the intonation states that the referent is not agreed upon. Thus
the referent shifts to the next most suitable canditate3 .
Our account like that of Hendriks (2004), sees the effect of coherence relations
on pronoun interpretation as epiphenomena. It does, however, diverge from her
account in that we take plausibility–not causality–as primary to pronoun interpretation. Causality, when one event causes another, is closely related to plausibility,
the perceived likelihood of an event. For example, chopping a tree trunk with an
axe will cause it to fall down, merely going into the woods with an axe will not.
However going into the woods with an axe makes it very plausible that a tree
will be felled. The inferences that one makes to connect two ideas, going into the
woods and cutting down a tree. Provide some evidence for this. These inferences
are based on frequent scenarios, and likely outcomes.
Importantly, while cause–effect sentences are generally biased towards one referent in their plausibility, this is also intuitively at play in contiguity relations (6)
(Kehler et al., 2008, p.23). In his description of the contiguity relations occasion
Kehler et al. (2008) suggests that the first sentence is used as an initial state for a
system and the second as the end state. What then links the two states together?
Our argument is that it is general and situational knowledge.
A critique that could be levelled against our proposal is that the topic is not
constant in our sentences. However, the sentences are correct in the context of
a question about both characters in subject position. For example, What did the
kids do today? (Kehler et al., 2008, p.10). This is compatible both with both
Kehler et al.’s approach and with information structure (Rooth, 1992).
Our hypothesis is consistent with the work of Gordon in stating that sentences
in sequence with a constant topic will be processed more quickly, even if a constant
topic is not completely necessary for comprehension in a situation with strong
plausibility to one referent only (Gordon & Hendrick, 1999, 1998). In particular,
if there are two sentences with the same meaning, one with a constant topic and
the other without, the sentence that maintains the topic structure will most likely
be preferred.
Another obvious step in this line of research is to investigate whether there is
any interaction of plausibility and topic consistency with other factors—particularly
in on-line experimentation. Future studies should sort out differences between output and the on-line processes that achieve this output.
3
Accented pronouns are thus different from other accented words only in that they are so
semantically empty, thus their referent is defined only by information packaging and context.
Other words differ in that lexical information plays a role in determining plausible referents.
Chapter 3
Experiment 2:
Discourse structure and models of
syntactic parsing
We investigated pronoun and discourse structure ambiguity processing. Ambiguity has played a crucial role in determining which models of sentence processing
best describe human language (Clifton Jr. & Staub, 2008; Traxler, Pickering, &
Clifton Jr., 1998). In the processing of adjunct attachment ambiguities, perhaps
surprisingly, ambiguous expressions are read faster than unambiguous expressions
in eye-tracking and self-paced reading experiments (Clifton Jr. & Staub, 2008).
These results are quite robust and have been repeated in different sentence types:
if two interpretations are equiplausible, then the ambiguous sentence is read faster
(Clifton Jr. & Staub, 2008). Previous findings indicate that semantic or morphological information helps with processing; the effect of pronoun reference ambiguity
has not yet been tested.
Pronouns thus provide an excellent possibility for researching ambiguities. The
model of pronoun comprehension we describe in Chapter 1 suggests that if the
pronoun is not clearly topic, and the situation does not make one referent more
plausible, then ambiguity will occur.
To say that ambiguity effects are robust is not to say that they have always
been a matter of consensus. Rather, they have been a major point of contention
between different models, especially because competition models are said to make
different predictions than other models.
3.0.3
Sentence processing models
Competition models predict that at each word there is a competition between
possible parses Traxler et al., 1998. After a threshold has been reached, the parser
35
CHAPTER 3. EXPERIMENT 2
36
continues. This causes the parser to stall at places where there is ambiguity as
the information for the options is computed and a threshold is reached for one
option or another (Traxler et al., 1998; but see Green & Mitchell, 2006). These
predictions, however, go against experimental results; Traxler et al. (1998) found
that globally ambiguous sentences (1a) were processed faster than temporarily
ambiguous sentences (1b or c), shown in:
(1)
a.
b.
c.
The son of the driver that had the moustache was pretty cool.
(globally ambiguous)
The driver of the car that had the moustache was pretty cool.
(high attachment)
The car of the driver that had the moustache was pretty cool.
(low attachment)
It makes sense for either the son or the driver to have a moustache, as in (1a).
It does not make sense though for a car to have a moustache, as is a possible
interpretation in (1b) and (1c). In these sentences, Traxler et al. (1998) found
that the first of these sentences was processed faster than the other two. In order
to account for these results, they proposed the race model.
The race model is not the only model that makes the correct predictions for the
parsing of syntactic ambiguities. The good-enough parser offers another solution:
the parser does not bother to create the structure. The structure will only be
fully constructed if the sentence stops making sense or if there is some reason to
pay particular attention to the parse. Readers will not pay attention to who is in
possession of a moustache it is not contradictory or useful information (answering a
comprehension question in an experiment or disambiguating characters in a novel).
The unambiguous sentences take more time because dependencies between phrases
are fully established.
It is entirely possible that that the good-enough parser and the race model
are different strategies. In cases where participants have specific comprehension
goals, they will attempt to parse the sentence as quickly as possible. Sentence
comprehension must be quick because conversations proceed quickly. Although
conversations proceed quickly, there is often extraneous information, and hearers can usually ask for clarification or re-read a passage if the parser fails. An
obviously implausible sentence will be found, but otherwise the sentence will be
under-represented.
The good-enough parser and the race model can, in theory, be distinguished
in a self-paced reading study by looking at the distribution of RTs at the place
an ambiguity is resolved. The good-enough parser will show faster processing
at ambiguities because it does not construct the parse but rather holds onto the
information and either discards it or uses it to construct the proper meaning. The
CHAPTER 3. EXPERIMENT 2
37
good-enough parser predicts that RTs for ambiguous sentences will cluster around
a single mean, as will unambiguous sentences.
Not so for unambiguous sentences in the race model. Some RTs reflect correct
first parses and others reflect incorrect first parses — the race model predicts
a bimodal distribution (although this may not be observed because of variance
between participants; van Gompel, p.c.). First parse RTs that are correct will
cluster around one mean, and incorrect first parses will cluster around another.
3.0.4
Discourse structure ambiguities
The separateness of the language module has been a long-standing debate in linguistics. The models that best account for the data are those which assume a
neurologically dispersed, unencapsulated parser. Both of the parsers discussed
above assume a greedy parser, that is one that will draw its information from any
source. However, so far as we are aware, no one has investigated whether discourse
structures are parsed in the same way as syntactic structures.
Discourse structure comprises the coherence and information structure of a
text. One part of the information structure is the topic structure. Grosz and Sidner
(1986) suggests that there is a hierarchic relation between topics and subtopics in
a discourse. Subtopics describe the topic that they are dominated by. These
structures can be ambiguous. In (2a) presents a topic, (2b) a subtopic, and it is
not obvious whether (2c) is a subtopic of (2a) or (2b); it is ambiguous.
(2)
a.
b.
c.
Frank showed Steve the new time machine in the laboratory.
Steve sent Frank back in time.
He wrote from within the laboratory about what happened.
Explaining the phenomenon in a more abstract way, when two sentences are arranged such that A precedes B, the topic of a sentence A is likely to be the topic
of sentence B (Van Vliet, 2008). However, when a sentence could refer to the topic
or subtopic – a temporary ambiguity is created. If the same system that processes
syntax also resolves pronouns, then pronouns in the ambiguous condition will be
read faster than in the unambiguous condition.
If the ambiguous condition is faster than the unambiguous sentences, this provides support for both the good-enough model and the race model. Furthermore,
showing that pronoun ambiguities speed processing suggests that the same cognitive system that parses syntactic structure also parses discourse structure.
CHAPTER 3. EXPERIMENT 2
3.1
3.1.1
38
Method
Participants
We tested 19 native speakers of Dutch. Participants were linguistics students at
the University of Groningen who received course credit for participation.
3.1.2
Materials and design
The sentences that we presented took the same form as (2). First one proper
name was mentioned Frank, then a verb, then proper name two, Steve. In the
next sentence, the ordering of the names was changed. Last, there was a sentence
starting with a personal pronoun, verb, disambiguating phrase, and a final phrase.
(3)
Frank1 showed Steve2 the-new-time-machine-in-the-laboratory.
Name1 verb
Name2 final-phrase
(4)
Steve2 sent Frank1 back-in-time.
Name2 verb Name1 final-phrase
(5)
He? wrote from-within-the-laboratory about-what-happened.
Pro verb disambiguating-phrase
final-phrase
There were two manipulations in this experiment. We created the distinction
between ambiguous and unambiguous pronouns by manipulating the gender of
the proper names. Sentences either had two male proper names, or one male
and one female proper name (all names were unambiguously male or female). In
unambiguous gender sentences, there was one male and one female referent, thus
participants could infer the referent of the pronoun without the sentence context.
In the ambiguous gender sentences, both referents were male, so the referent could
only be figured out with the help of the disambiguating phrase.
We also deemed it important to manipulate the referent that the disambiguating phrase referred to; in this way, any advantage could not be the result of topic
continuation. In one manipulation, the referent was the topic of the first sentence,
and in the other it was the topic of the second sentence. As shown in (6), this
resulted in disambiguating phrases that could refer either to the Topic (Name1 ) or
Subtopic (Name2 ). We verified the referent was correct with an off-line plausibility
judgment task.
(6)
He? wrote from-within-the-{time-machine1 ,laboratory2 }
Pro verb disambiguating-phrase-{Name1 ,Name2 }
about-what-happened.
final-phrase
CHAPTER 3. EXPERIMENT 2
(7)
(8)
(9)
39
a.
Frank1 showed Steve2 the new time machine in the laboratory.
b.
c.
Steve2 sent Frank1 back in time.
(i) He? wrote from within the laboratory about what happened.
(Ambiguous, Name1 )
(ii) He? wrote from within the time machine about what happened.
(Ambiguous, Name2 )
a.
Sarah1 showed Steve2 the new time machine in the laboratory.
b.
c.
Steve2 sent Sarah1 back in time.
(i) He2 wrote from within the laboratory about what happened.
(Ambiguous, Name1 )
a.
Frank1 showed Sarah2 the new time machine in the laboratory.
b.
c.
Sarah2 sent Frank1 back in time.
(i) He1 wrote from within the time machine about what happened.
(Ambiguous, Name2 )
This produced a total of 4 conditions, ambiguous–topic, ambiguous–subtopic, unambiguous–
topic, unambiguous–subtopic. Note that the referent always had the correct gender.
3.1.3
Procedure
Participants pressed a key to move from word to word in a self-paced reading task.
A computer displayed the words centred in the screen one at a time. Following
each three-sentence set, participants had to say whether they thought the series
of events was realistic.
3.2
Interim Results
We analysed the RTs in the last sentences of the stimuli by breaking them into
regions. The first two words, the verb and the pronoun were analysed together,
the disambiguating phrase was analysed by looking at the first word that clearly
disambiguated the sentence, analysing the three words before it and the three
words after it, shown in Figure 3.1.
Ambiguous pronouns were read numerically faster than unambiguous pronouns.
In fact, sentences with ambiguity were read faster overall than unambiguous sentences. This evidence counters the predictions of the competition model and provides evidence in favour of either the good-enough account or the race model. It
also suggests that the syntactic parser is not so syntactic after all.
CHAPTER 3. EXPERIMENT 2
40
Figure 3.1: RTs and standard errors from the last sentence of a stimuli Experiment
2. On the left, word one was always he, and two was always a verb. 0 is the
disambiguating word, with the three words before it and the three words after it.
The Unambiguous words were slower than Ambiguous words in all regions except
the disambiguating word
At the point of disambiguation responses were slower than elsewhere. This was
likely due in part to the fact that the disambiguating word was usually a noun,
and therefore was longer than the words around it.
3.3
Discussion
Our results show that ambiguous pronouns are read faster than unambiguous
pronouns. This is further evidence that either the good-enough parser or the race
model are correct. This also provides strong evidence that discourse structural
ambiguities are processed in the same way as syntactic ambiguities. The syntactic
parser is likely one and the same as the discourse parser.
Chapter 4
Future directions
Proposed experiments and
implementation plan for year two
Our goal is to understand the role of prosody in pronoun resolution. In line with
this, we have tried to understand pronouns in environments both contextually enriched and impoverished. In Chapter 2, we examined the off-line resolution of the
pronouns, manipulating accenting and coherence relations. This created a context–
rich environment. In Chapter 3 we examined the on-line resolution of pronouns;
how resolution occurs in a context-impoverished environment. These two directions partition the analysis and will reconverge on a more exact comprehension of
pronoun resolution.
Chapter 3 seemed to provide support for the good-enough model of processing or the race model over other models of processing, but this must be further
investigated. To that end we propose Experiments 2b and 3.
4.1
Context-poor environments
In Experiment 2 we manipulate the referent of the pronoun with a disambiguating
phrase. This experiment appeared to show that resolution happens later rather
than sooner. This lends support to the good-enough model, because it predicts
that the syntax can remain underspecified. The race model, in contrast, predicts
that sentence structure is always specified: if it cannot be inferred, it will be
guessed. We intend to test this hypothesis by partially replicating Swets, Desmet,
Clifton, and Ferreira (2008). Swets et al. (2008) showed sentences with adjunct
attachment ambiguities and manipulated the question following the sentence. Either the question asked very specific information, which required the sentence to be
41
CHAPTER 4. FUTURE DIRECTIONS
42
fully syntactically parsed, or it asked more general information. The good-enough
model predicts that when the questions are more specific, then the ambiguity advantage should disappear, whereas the race model predicts that there should be
no difference between closely scrutinized sentences and those that are just glossed
over.
Experiment 2b
Our proposal is to manipulate the question following the stimuli used in Experiment 2 (1):
(1)
a.
b.
c.
d.
Frank showed Steve the new time machine in the laboratory.
Steve sent Frank back in time.
He wrote from within the laboratory about what happened.
Who wrote about what happened? (Frank/Steve)
As it stands, Experiment 2 does not allow us to know whether the pronoun is ever
resolved. Adding an additional condition solves this problem. If the pronoun is
unresolved, then a neutral disambiguation phrase (2) will be no faster than either
of the other two conditions.
(2)
a.
Frank1 showed Steve2 the new time machine in the laboratory.
b.
c.
Steve2 sent Frank1 back in time.
(i) He? wrote in his journal about what happened.
(Ambiguous,
Ambiguous)
(ii) He? wrote from within the laboratory about what happened.
(Ambiguous, Name1 )
(iii) He? wrote from within the time machine about what happened.
(Ambiguous, Name2 )
Experiment 3
It may be the case that faster response times are not indicative of easier processing.
The ambiguity advantage might result from participants trying to race ahead to
get more information. Using ERP allows us to “look under the hood” to see
whether the faster reaction times are in fact indicative of greater processing effort.
In ERP, greater processing difficulty results in a greater deflection of the signal
Van Berkum (in press). The ERP signal that Van Berkum (in press) describe,
the N-ref, is conventionally assumed to be a negativity (hence the “N”). However,
CHAPTER 4. FUTURE DIRECTIONS
43
it may be the case that it is actually a positivity, a P-ref 1 . However, it seems
relatively clear that the negative deflection that the Nref describes is an indication
of greater difficulty: Van Berkum et al. (2007) found that establishing one referent
caused a greater signal deflection than none, and two a greater deflection than one.
4.2
Context-poor environments
Experiment 4
We looked at the effects of discourse context on pronoun resolution in the form
of coherence relations and plausibility in Chapter 2. We intend to follow this up
at the end of the second year and beginning of the third year. We will do this in
Experiment 4 with a series of 2 sub-experiments. The first of these will manipulate coherence while keeping plausibility constant, and the second will manipulate
plausibility while keeping the coherence relation constant. For example, in (3),
we manipulate a disambiguating phrase so that it refers either to the first name
mentioned or the second (that is to say, either the first name is the most plausible,
the is the most plausible second, or either is plausible):
(3)
a.
Frank1 showed Steve2 the new time machine in the laboratory.
b.
c.
Steve2 sent Frank1 back in time.
(i) From within the laboratory wrote he? about what happened .
(Ambiguous, Name1 )
(ii) From within the time machine wrote he? about what happened.
(Ambiguous, Name2 )
(iii) In his journal wrote he? about what happened .
(Ambiguous,
Ambiguous)
Note that the prepositional phrase is now first, so that when the listener/reader
arrives at he the referent is already resolvable.
1
N-ref conventionally has no dash, but Pref seems like an abbreviation of preference, an
association we wish to avoid.
Appendix A
Schedule
A.1
Year 2
Table A.1: Research activities for year two
Experiment 2b
Testing
September ’09
Analysis
September ’09
Write-up (CUNY)
Deadline: December ’09
Write-up for AMLAP/CogSci
Deadline: April ’10
Write-up w/ exp. 2 for submission February’10
Experiment 3
Stimulus design
Until May ’10
Testing
June ’10
Analysis
August ’10
Experiment 4
Stimulus design
August ’10
Testing
September ’10
Write-up
October ’10
Attend/present at CUNY, AMLAP, CogSci, Tindag,
Tabudag, BCN posterday, BCN symposium
44
APPENDIX A. SCHEDULE
45
Table A.2: Training activities for year two
Masterclass (Gussenhoven)
Dutch Niveau III
September–November ’09
Statistics proseminar
September–November ’09
BCN Posterday
January ’10
LOT Winter School
January ’10
LOT Summer School
June ’10
Project Management
June ’10
ELG group
ongoing
EEG group
ongoing
Neurolinguistics
ongoing
A.1.1
Year 3
In year 3 we will continue to research the place of context, in the form of plausibility
and coherence relations. One option is to expand the research by using naturalistic
examples found in corpus research, as opposed to the shorter, more controlled
stimuli we are now using. We will also write up the results.
Experiments 5 & 6
Experiment 5 will take whichever factor contributes the most to pronoun resolution
(or both as the case may be) and manipulate accent in an antecedent preference
task as discussed in Chapter 2. Experiment 6 will take the same materials and
put them on-line in an ERP study.
A.1.2
Year 4
Year 4 will be consumed with the writing of the dissertation and the writing of
manuscripts for submission to journals.
Appendix B
Quick and Dirty Guide to
Optimality theory
Optimality theory1 (OT) was initially developed to deal with phonology, although
it has since been expanded to deal with syntax, semantics, and language processing
(Hoeks & Hendriks, 2005). OT consists of two modules, gen and eval. Gen
functions by producing every possible output form, a trait referred to as freedom
of analysis. It is so-named because by producing every possible output form, any
linguistic output is possible. However, most of these forms are never actually
uttered, because they are evaluated by eval to see how well they conform to a set
of constraints.
Simply put, constraints are defined and arranged hierarchically in graphical
form constraints further to the left are more important than constraints to the
right. The winning output is the one for which the first violation occurs closest
to the right. If there are several outputs are tied, then these outputs are tested
again ignoring the first violation until there is one or several winning candidates.
To illustrate the usage of these graphical representations of hierarchies or tableaux,
we discuss the trivial although informative example of what to eat for dinner.
Imagine preparing a dinner for several guests, one of whom is lactose intolerant,
another is a health food nut, and the third sort of feels like something italian, but
does not mind that much.
*Dairy The food must not contain milk products.
*Fatty The food should be low in fat
Italian The food should be Italian
1
For a more complete introduction to OT, see Kager, 1999.
46
APPENDIX B. QUICK AND DIRTY GUIDE TO OPTIMALITY THEORY 47
In Figure B.1 we can see how this hierarchy effects the choice of food. The tableau
is evaluated from left to right. Gelatto loses immediately because of the constraint
*Dairy. Notice that it does not make any difference that it is Italian, because this
is a low ranked constraint. Sushi would win, except that it is not Italian, leaving
Rotini as the winner (hold the parmesan).
Input
+
Gelatto
Donair
Sushi
Rotini in tomato sauce
*Dairy
*
*Fatty
*
Italian
*
*
Figure B.1: OT Applied to Choice of Dinner
Things are made slightly more complex by two phenomena: Equally ranked
constraints, and gradable constraints. In the example above, *Dairy and *Fatty
are in fact equally important. This is indicated by eliminating the line between
the two categories. Additionally, some things are fattier than others, thus there
can be more than one violation (asterisk) in the *Fatty category.
Input
Gelatto
Falafel
Donair
Sushi
Rotini in tomato sauce
*Dairy
*
*
*Fatty
*
*
**
Italian
*
*
*
Figure B.2: OT Applied to Choice of Dinner 2
Again, Rotini would win. But if Rotini were not available, then Sushi would
win. Even though Sushi violates some constraints, it still violates the fewest. If
the only options are Gelatto and Falafel, then Gelatto wins because Falafel is not
Italian (for a more complete introduction to OT, see Kager, 1999).
For a less concrete, but more linguistic example, we present an OT tableau of
the number, gender and case constraints on pronouns.
AgreeNumber A pronoun must agree in number with its referent
AgreeGender A pronoun must agree in Gender with its referent
AgreeCase A pronoun in subject position takes Nominative case, a pronoun in
object position takes Accusative case2
2
For the sake of clarity we will pretend that examples like Jason asked Frank and I about
APPENDIX B. QUICK AND DIRTY GUIDE TO OPTIMALITY THEORY 48
In Figure B.3 we use a tableau to illustrate number and gender agreement in
pronouns, following the constraints described above. In the first example sentence,
Input Pronoun = Steve1
Dan talked to her1
Dan talked to them1
Dan talked to he1
+
Dan talked to him1
AgreeNumber
*
AgreeGender
*
∼
AgreeCase
*
Figure B.3: OT Applied to Number, Gender, and Case agreement in pronouns
her does not have the correct gender, in the second, the pronoun, them has the
wrong number, and in the third, he, the case is incorrect. There are no violations
in the last sentence, so it is the winning candidate.
going to see a movie tonight do not exist since their existence is cursory to the discussion at
hand.
References
Aikhenvald, A., & Dixon, R. (1998). Dependencies between grammatical systems.
Language, 56–80.
Arregui, A., Clifton Jr., C., Frazier, L., & Moulton, K. (2006). Processing elided
verb phrases with flawed antecedents: The recycling hypothesis. Journal of
Memory and Language, 55 (2), 232–246.
Beaver, D. (2004). The optimization of discourse anaphora. Linguistics and
Philosophy, 27 (1), 3–56.
Bott, S. (2007). Information Structure and Discourse Modelling. Unpublished
doctoral dissertation, Universitat Pompeu Fabra, Barcelona, Spain.
Caramazza, A., Grober, E., Garvey, C., & Yates, J. (1977). Comprehension of
anaphoric pronouns. Journal of Verbal Learning and Verbal Behavior , 16 (5),
601–9.
Chafe, W. L. (1973). Language and memory. Language, 50 (1), 261–281.
Clifton Jr., C., & Staub, A. (2008). Parallelism and competition in syntactic
ambiguity resolution. Language and Linguistics Compass, 2/2 , 234–250.
Comrie, B. (1999). Grammatical gender systems: a linguist’s assessment. Journal
of Psycholinguistic research, 28 (5), 457–466.
Connell, L., & Keane, M. (2006). A model of plausibility. Cognitive Science: A
Multidisciplinary Journal , 30 , 95–120.
Connell, L., & Keane, M. T. (2004). What plausibly affects plausibility? concept
coherence and distributional word coherence as factors influencing plausibility
judgments. Memory and Cognition, 32 (2), 185–197.
Crawley, R., Stevenson, R., & Kleinman, D. (1990). The use of heuristic strategies
in the interpretation of pronouns. Journal of Psycholinguistic Research, 19 (4),
245–264.
Crinean, M., & Garnham, A. (2006). Implicit causality, implicit consequentiality
and semantic roles. Language and Cognitive Processes, 21 (5), 636–648.
De Hoop, H. (2004). On the interpretation of stressed pronouns. In R. Blutner
& H. Zeevat (Eds.), Optimality theory and pragmatics. New York, NY: Palgrave
Macmillan.
Garvey, C., Caramazza, A., & Yates, J. (1975). Factors influencing assignment of
49
References
50
pronoun antecedents. Cognition, 3 (3), 227—243.
Gordon, P. C., & Hendrick, R. (1998). The representation and processing of
coreference in discourse. Cognitive Science, 22 , 389–424.
Gordon, P. C., & Hendrick, R. (1999). Comprehension of coreferential expressions.
New Brunswick, NJ, United States: Association for Computational Linguistics.
Green, M. J., & Mitchell, D. C. (2006). Absence of real evidence against competition during syntactic ambiguity resolution. Journal of Memory and Language,
55 , 1–17.
Grober, E., Beardsley, W., & Caramazza, A. (1978). Parallel function strategy in
pronoun assignment. Cognition, 6 , 177–133.
Grosz, B., & Sidner, C. (1986). Attention, intentions, and the structure of discourse. Computational Linguistics, 12 (3), 175–204.
Hagoort, P., & Van Berkum, J. J. A. (2007). Beyond the sentence given. Philosophical transactions of the Royal Society, 362 , 801–811.
Hendriks, P. (2004). Coherence relations, ellipsis and contrastive topics. Journal
of Semantics, 21 , 122-153.
Hendriks, P., Englert, C., Wubs, E., & Hoeks, J. (2008). Age differences in adults’
use of referring expressions. Journal of Logic, Language and Information, 17 (4),
443–466.
Hoeks, J. C. J., & Hendriks, P. (2005). Optimality theory and human sentence
processing: The case of coordination. Proceedings of the 27th Annual Meeting of
the Cognitive Science Society, Erlbaum, Mahwah, NJ , 959–964.
Hoeks, J. C. J., Prinsen, I., & Stowe, L. A. (2003). Semantic illusion in sentence
processing: a right-hemisphere mechanism? In F. Schmalhofer, R. M. Young, &
G. Katz (Eds.), (pp. 175–180). Hillsdale, NJ: Erlbaum.
Hoeks, J. C. J., Stowe, L. A., & Doedens, G. (2004). Seeing words in context: the
interaction of lexical and sentence level information during reading. Cognitive
Brain Research, 19 (1), 59–73.
Hoeks, J. C. J., Vonk, W., & Schriefers, H. (2002). Processing coordinated structures in context: the effect of topic-structure on ambiguity resolution. Journal
of Memory and Language, 46 (1), 99–119.
Kaan, E., Wijnen, F., & Swaab, T. (2004). Gapping: Electrophysiological evidence
for immediate processing of “missing” verbs in sentence comprehension. Brain
and Language, 89 (3), 584–592.
Kager, R. (1999). Optimality theory. Cambridge: Cambridge University Press.
Kehler, A. (2002). Coherence, Reference, and the Theory of Grammar. Stanford,
United States: Center for the Study of Language and Information Publications.
Kehler, A. (2005). Coherence-driven constraints on the placement of accent. In
Proceedings of the 15th conference on semantics and linguistic theory (salt-15)
(pp. 98–115). Fort Washington, United States: CLC Publications.
References
51
Kehler, A., Kertz, L., Rohde, H., & Elman, J. (2008). Coherence and coreference
revisited. Journal of Semantics(25), 1–44.
Kertz, L., Kehler, A., & Elman, J. (2006a). Evaluating a coherence-based model
of pronoun interpretation. In R. Artstein & M. Poesio (Eds.), Ambiguity in
anaphora workshop proceedings (pp. 49–56).
Kertz, L., Kehler, A., & Elman, J. (2006b). Grammatical and coherence-based
factors in pronoun interpretation. Proceedings of the 28th Annual Conference of
the Cognitive . . . .
Koornneef, A. W., & Van Berkum, J. J. A. (2006). On the use of verb-based
implicit causality in sentence comprehension: Evidence from self-paced reading
and eye tracking. Journal of Memory and Language, 54 (4), 445-465.
Kuno, S. (1974). The position of relative clauses and conjunctions. Linguistic
Inquiry, 5 (1), 117–136.
Levinson, S. (1983). Pragmatics. Cambridge: Cambridge University Press.
Lewis, R. L., & Vasishth, S. (2005). An activation-based model of sentence processing as skilled memory retrieval. Cognitive Science, 29 (3), 375–419.
Lewis, R. L., Vasishth, S., & Van Dyke, J. A. (2006). Computational principles
of working memory in sentence comprehension. TRENDS in Cognitive Sciences,
10 (10), 447–454.
Mel’čuk, I. (1988). Dependency syntax: Theory and practice (M. Aronoff, Ed.).
Albany, NY, United States: State University of New York Press.
Mel’čuk, I. (2001). Communicative organization in natural language: The
semantic-communicative structure of sentences (Vol. 57). John Benjamins Publishing Company, Amsterdam, The Netherlands.
Miller, G. A. (1956). The magical number seven plus or minus two: some limits on
our capacity for processing information. The Psychological Review , 63 , 81–97.
Nieuwland, M. S., & Van Berkum, J. J. A. (2008). The neurocognition of referential
ambiguity in language comprehension. Language and Linguistics Compass, 2 ,
603-630.
Pickering, M. J., & Majid, A. (2007). What are implicit causality and consequentiality. Language and Cognitive Processes, 22 (5), 780–788.
Pickering, M. J., & Van Gompel, R. P. G. (in press).
In M. J. Traxler & M. A. Gernsbacher (Eds.), (pp. 455–503). San Diego CA:
Elsevier.
Prince, E. F. (1981). Radical pragmatics. In P. Cole (Ed.), (pp. 223–255). London:
Academic Press.
Rohde, H., Kehler, A., & Elman, J. (2006). Event structure and discourse coherence biases in pronoun interpretation. In Proceedings of the 28th annual conference of the cognitive science society (pp. 697–702). Vancouver, BC, Canada.
Rooth, M. (1992). A theory of focus interpretation. Natural language semantics,
References
52
1 (1), 75–116.
Schwarzschild, R. (1999). Givenness, avoidf and other constraints on the placement
of accent. Natural Language Semantics, 7 , 141–77.
Sitnikova, T., Kuperberg, G., & Holcomb, P. (2003). Semantic integration in videos
of real-world events: An electrophysiological investigation. Psychophysiology, 40 ,
160–164.
Smyth, R. (1994). Grammatical determinants of ambiguous pronoun resolution.
Journal of Psycholinguistic Research, 23 (3), 197–229.
Steedman, M. (2007). Information-structural semantics for English intonation. In
D. B. Chungmin Lee M Gordon (Ed.), Topic and focus: Cross-linguistic perspectives on meaning and intonation (pp. 245–264). Dordrecht: Springer.
Swets, B., Desmet, T., Clifton, C., & Ferreira, F. (2008). Underspecification of
syntactic ambiguities: Evidence from self-paced reading. Memory and Cognition,
36 (1), 201.
Taylor, R. C., Stowe, L. A., Redeker, G., & Hoeks, J. C. J. (2009). The referent
of accented pronouns is determined by coherence relations. In Proceedings of the
31st annual conference of the cognitive science society.
Traxler, M., Pickering, M., & Clifton Jr., C. (1998). Adjunct attachment is not
a form of lexical ambiguity resolution. Journal of Memory and Language, 39 ,
558–592.
Vallduvı́, E., & Engdahl, E. (1996). The linguistic realisation of information
packaging. Linguistics, 34 (3), 459–519.
Vallduvı́, E., & Vilkuna, M. (1998). On rheme and kontrast. In P. Culicover &
L. McNally (Eds.), The limits of syntax and semantics (pp. 79–108). San Diego
CA: Academic Press.
Van Berkum, J. J. A. (1999). When does gender constrain parsing? Journal of
Psycholinguistic Research, 28 (5), 555–571.
Van Berkum, J. J. A. (in press). The electrophysiology of discourse and conversation. In M. Spivey, M. Joanisse, & K. McRae (Eds.), The cambridge handbook
of psycholinguistics. Cambridge, UK: Cambridge University Press.
Van Berkum, J. J. A., Koornneef, A., & Otten, M. (2007). Establishing reference in
language comprehension: An electrophysiological perspective. Brain Research,
1146 (158–171).
Van Gompel, R. P. G., Pickering, M. J., Pearson, J., & Jacob, G. (2006). The
activation of inappropriate analyses in garden-path sentences: Evidence from
structural priming. Journal of Memory and Language, 55 , 335-362.
Van Vliet, S. (2008). Proper nouns and pronouns: The production of referential
expressions in narrative discourse (Vol. 175). Utrecht, The Netherlands: LOT.
Venditti, J., Stone, M., Nanda, P., & Tepper, P. (2002). Discourse constraints on
the interpretation of nuclear-accented pronouns. Speech Prosody 2002 .
References
53
Wolf, F., Gibson, E., & Desmet, T. (2004). Discourse coherence and pronoun
resolution. Language and Cognitive Processes, 19 (6), 665–675.