University of Edinburgh The University of Edinburgh School of

					        The University of Edinburgh
           School of Philosophy,
      Psychology & Language Sciences


    Undergraduate Psychology Honours
         Final Year Dissertation


Processing pronouns: Effects of content-based
   properties of potential antecedents and
        similarity-based interference.


           SARAH RUSBRIDGE

          Supervisor: Patrick Sturt
CONTENTS



ABSTRACT .............................................................................................................. 1

INTRODUCTION.................................................................................................... 3

          Constraints on antecedent resolution ............................................................. 4

          The initial-filter model ................................................................................... 7

          The interactive-parallel constraint model ...................................................... 9

          The memory load hypothesis ......................................................................... 12

          The present study ........................................................................................... 14

METHODS ............................................................................................................... 17

          Participants ..................................................................................................... 17

          Design ............................................................................................................ 17

          Materials ........................................................................................................ 20

          Procedure ....................................................................................................... 22

RESULTS ................................................................................................................. 24

          Analysis.......................................................................................................... 24

          Pre-critical region........................................................................................... 26

          Pronoun region ............................................................................................... 27

          Critical region ................................................................................................ 27

          Post-critical region ......................................................................................... 29

          Probe recognition ........................................................................................... 30

DISCUSSION ........................................................................................................... 32

REFERENCES ......................................................................................................... 44
APPENDICES .......................................................................................................... 51

          Appendix A: Experimental materials............................................................. 51

          Appendix B: Statistics for word reading times and probe responses ............. 55
ABSTRACT



This report examines the process of antecedent identification for pronouns, which is an

essential part of language comprehension in general. The interactive-parallel constraint

model proposes that antecedent resolution is the result of many structural and non-

structural constraints competing simultaneously, suppressing and enhancing the

activation levels of potential antecedents within the discourse representation

(Gernsbacher, 1989; MacDonald & MacWhinney, 1990). This model is supported by

previous evidence that the gender of grammatically-inaccessible antecedents, according

to Principle B of binding theory (Chomsky, 1981), affects antecedent resolution

(Badecker & Straub, 2002). Using a word-by-word self-paced reading methodology, this

experiment investigated whether this „multiple-candidate effect‟ can instead be attributed

to the basic difficulty encoding a sentence containing two less distinguishable referents of

the same gender as compared to different genders, termed the „memory load hypothesis‟.

The results did not replicate the multiple-candidate effect, and found no support for the

memory load hypothesis in early processing. Results in later processing revealed a

conflicting effect of slower reading times for words located towards the end of a sentence

that contained two referents of different genders as opposed to being gender congruent.

This effect could be attributable to a process in which readers attempt to bind the pronoun

to the most-recently processed grammatically-inaccessible antecedent within the

sentence, or it might be evidence of gender-based priming in referent encoding. Finally,

evidence for the memory load hypothesis was found on response times in the probe

recognition task; responses were longer if the probe followed a sentence containing two




                                                                                            1
proper names of congruent gender, irrespective of whether the sentence required

antecedent resolution. A type of similarity-based interference was apparently affecting

probe responses; it seems the same processes that are accessed in initial sentence

processing are re-accessed in later probe sentence-checking.




                                                                                          2
INTRODUCTION



       Understanding how anaphoric dependencies are processed is an essential part of

comprehending how language is processed in general. Anaphoric dependencies constitute

the relationship between a referentially dependent expression, such as a pronoun, and its

antecedent noun phrase. Identifying an antecedent for a pronoun is necessary so that a

reader can understand what a pronoun refers to and thereby build a coherent

representation of the discourse being processed. Research has shown considerable interest

in investigating the process by which such dependencies are resolved. A number of

constraints have been identified as contributing towards antecedent resolution, primarily

grammatical constraints such as binding theory (e.g. Chomsky, 1981; Reinhart, 1976);

morphological constraints of the pronoun, such as person, gender, and number (e.g.

Arnold, Eisnband, Brown-Shmidt & Trueswell, 2000; Badecker & Straub, 1994;

Badecker & Straub, 2002; Cacciari et al, 1997; Ehrlich, 1980; Sturt, 2003; Van Gompel

& Liversedge, 2003); and the accessibility of referents within the discourse representation

(e.g. McKoon, Ward, Ratcliff, & Sproat, 1993; Van Gompel & Majid, 2004). This

experiment investigates the role of both grammatical and morphological gender

constraints on antecedent resolution imposed by the pronoun by manipulating the gender

of referents within the discourse representation built during comprehension, examining

whether gender information facilitates or hinders the process of antecedent resolution of

pronouns, or whether the manipulation of gender information is simply related to general

effects of similarity based-interference in language comprehension.




                                                                                            3
Constraints on antecedent resolution

       In sentence comprehension, it is generally assumed that a discourse model is

formed in temporary working memory (Baddeley, 2007) and incrementally updated as

the sentence is read (Carreiras, Garnham, Oakhill & Cain, 1996). The discourse model is

a coherent representation of the set of entities evoked by the discourse and the

relationships that exist between those entities, commonly referred to as „discourse

referents‟. Language comprehension necessitates such a representation so as to allow for

the retrieval of earlier words in a sentence in order to relate them to later words in a

sentence (Just & Carpenter, 1992). Interpreting referentially dependent expressions

requires accessing and searching the discourse representation for referents already

encoded in the representation of the text so far to which they can refer back to (Lucas,

Tanenhaus & Carlson, 1990). A number of factors have been identified which constrain

the resolution of such referential dependencies, some of which are now discussed.

       Grammatical constraints on the allowability of coreferential relationships are

imposed by binding theory (Chomsky, 1981). Binding theory proposes that structural

relations within a sentence limit the possible set of antecedents for a particular

referentially dependent expression, which are classified into three categories: anaphors

(reflexives and reciprocals, e.g. himself, each other), pronominals (non-reflexive

pronouns, e.g. he, her), and referential-expressions (henceforth R-expressions, defined as

uniquely identifiable entities such as full NPs, e.g. John, the duchess). The use of the

term „anaphor‟ here is distinct from the more general term used earlier to describe a

referentially-dependent expression. The concept of „binding‟ refers to coreference within

a discourse, and is based on the principle of c-command (Reinhart, 1976) and governing




                                                                                           4
categories. The use of the term „governing category‟ here is used to refer to the minimal

clause containing both the referentially-dependent expression and its grammatical

subject.

       Chomsky (1981) proposed three principles constraining the grammaticality of

coreferential relationships within a sentence based on structural and locality restrictions

imposed on the antecedent, and there evidence suggests that native English speakers‟

judgement of the acceptability of different kinds of coreferential relationships are in

agreement with these principles (Gordon & Hendrick, 1977). Principle A of binding

theory states that an anaphor must be bound locally within its governing category, thus

explaining sentence (1a). Principle B states that a pronominal must be bound, but must be

free within its governing category, which explains sentence (1b). Principle C states that

an R-expression must be free everywhere, which explains the ungrammaticality of

sentence (1c), except in the case that the two R-expressions refer to separate referents.



       (1a) Billi thought that Johnj saw himself*i/j.

       (1b) Billi thought that Johnj saw himi/*j.

       (1c) Johni saw John*i.



       Principle A specifies the appropriate antecedent of an anaphor, whereas Principles

B and C only identify those antecedents that a pronominal or R-expression cannot refer

to, thus other discourse processing procedures may thereby be required to select a

suitable antecedent from the grammatically-appropriate set. This report is mainly

concerned with the processing of pronouns constrained by Principle B, although some




                                                                                              5
comparisons will be made with reflexives governed by Principle A. Throughout this

report, those referents located in binding theory compatible positions will be referred to

as „grammatically-accessible‟, and those referents not in binding theory positions will be

referred to as „grammatically-inaccessible‟.

       Antecedent resolution has also been shown to be affected by the morphological

content-based constraints specified by a particular referentially-dependent expression.

Agreements of gender, number, and person between a pronoun and its potential

antecedents have been shown to facilitate antecedent selection (e.g. Arnold et al, 2000;

Badecker and Straub, 1994; Badecker and Straub, 2002; Cacciari et al, 1997; Sturt, 2003;

Van Gompel & Liversedge, 2003). These morphological features identify those entities

within a discourse that could be potential antecedents for pronouns, thereby assisting

resolution (Rigalleau & Caplan, 2000). Arnold et al (2000) has shown that gender

information about the pronoun is rapidly accessed in resolution, finding an influence of

gender within 200ms following the offset of the pronoun in cross-modal eye-tracking

experiments.

       Salience within the discourse representation is similarly thought to constrain

selection for the initial antecedent set. Pronouns are resolved faster when the antecedent

is presented in a focused position within the discourse, suggesting that ease of

accessibility within the discourse representation affects the difficulty of antecedent

resolution (McKoon et al, 1993). Badecker and Straub (2002) argue that only highly

focused referents are initially considered as antecedents of referentially-dependent

pronouns. Van Gompel and Majid (2004) found longer reading times for pronouns with




                                                                                             6
infrequent antecedents as opposed to frequent antecedents, a finding they propose to be

due to the infrequent antecedents being more salient within the discourse representation.

       Research has recently focused on examining the time-course of antecedent

resolution, investigating at what point in processing binding theory constrains

coreference, and moreover the impact of other non-structural constraints. A number of

different models have been proposed to attempt to explain the interplay between these

constraints in the process of antecedent resolution, some of which are discussed below.



The initial-filter model

       The „initial-filter model‟ proposes that grammatical constraints immediately

restrict the potential antecedent set to include only grammatically-accessible referents,

and only if this be insufficient for resolution are other discourse processing procedures

used later in processing (Nicol & Swinney, 1989). Cross-modal priming studies using

sentences such as those in (2a) and (2b) support this proposal (Nicol, 1988), finding that

only the grammatically-accessible referent the doctor is primed immediately after the

reflexive in (2a), whereas in (2b) the doctor is grammatically-inaccessible and

accordingly not primed following the pronoun. The two equally possible grammatically-

accessible referents the boxer and the skier are both primed after the pronoun in (2b),

consistent with the idea that Principle B functions to indicate only those referents

inaccessible to the pronoun.



       (2a)    The boxer told the skier that the doctori for the team would blame himselfi

               # for the recent injury.




                                                                                             7
       (2b)    The boxer(i) told the skier(i) that the doctor for the team would blame him(i)

               # for the recent injury.



       Participants‟ responses were faster only for probe words corresponding to

grammatically-accessible referents of the different referentially dependent expressions,

which seems indicative of a process in which there is immediate triggering of the

potential set of antecedents for referentially dependent expressions, composed only of

those that are grammatically-accessible according to the structural constraints imposed by

binding theory (Nicol & Swinney, 1989). Indeed, grammatically-inaccessible referents

appear not to have been considered as potential antecedents at any point in processing.

       Further support for the „initial-filter model‟ comes from a study by Clifton,

Kennison and Albrecht (1997), who manipulated grammatical number to investigate how

potential antecedents for pronouns are selected, using self-paced reading of sentences

such as (3a) to (3d).



       (3a) The supervisors paid him yesterday to finish typing the manuscript.

       (3b) The supervisor paid him yesterday to finish typing the manuscript.

       (3c) The supervisors paid his assistant yesterday to finish typing the manuscript.

       (3d) The supervisor paid his assistant yesterday to finish typing the manuscript.



       In (3a) and (3b), the grammatical subject the supervisor(s) is not in a

grammatically-accessible antecedent location for the accusative pronoun him according to

Principle B, whereas in (3c) and (3d) the subject is in a grammatically-accessible position




                                                                                            8
for the possessive pronoun his, although only in (3d) does the subject match the pronoun

in number. Reading times were faster in the number-congruent sentence (3d), and longer

in the number-incongruent sentence (3c), in which there was a mismatch and further

processing was required. Number congruency had no effect on reading times in either

(3a) or (3b), which was taken as evidence that binding theory is imposed as an initial-

filter on the potential antecedent set, with number information only being utilised later in

processing if binding theory was unsuccessful in antecedent resolution.



The interactive-parallel constraint model

       An alternative model of antecedent resolution is offered by what Badecker and

Straub (2002) refer to as the „interactive-parallel constraint model‟. This model proposes

that antecedent resolution is based on multiple constraints competing in parallel,

including binding theory, morphological agreement (i.e. gender, person and number),

discourse focus, and order of mention, all of which result in the suppression and

enhancement of activation levels of referents within the discourse representation

(Gernsbacher, 1989). The final activation level of a referent is regarded to be the sum of

these competing facilitatory and inhibitory activation levels. Accordingly, probe response

time studies reveal evidence of both facilitation effects for antecedents and inhibition

effects for non-antecedents (Gernsbacher, 1989). MacDonald and MacWhinney (1990)

report similar findings and conclude that as a discourse is processed, the level of

activation of particular elements within the discourse representation constantly changes

depending on the contributions of such parallel-acting constraints. The fundamental idea

of this model is that antecedent resolution is affected by many different kinds of lexical




                                                                                             9
information and grammatical constraints, in a similar manner to how lexical and syntactic

ambiguities are processed (e.g. MacDonald, Pearlmutter & Seidenberg, 1994; Trueswell,

1996).

         Previous research has reported evidence of different constraints acting on

antecedent resolution, whereby content-based properties of grammatically-inaccessible

referents have affected the processing of pronouns (e.g. Badecker & Straub, 1994;

Badecker & Straub, 2002; Sturt, 2003). Badecker and Straub (2002) designed an

experiment to test the predictions of both the initial-filter model and the interactive-

parallel constraint model, using sentences such as (4a) to (4d) (in their Experiment 1),

which all include more than one highly salient referent.



         (4a) John thought that Bill owed him another chance to solve the problem.

         (4b) John thought that Beth owed him another chance to solve the problem.

         (4c) Jane thought that Bill owed him another chance to solve the problem.

         (4d) Jane thought that Beth owed him another chance to solve the problem.



         Using self-paced word-by-word reading tasks, sentences in which grammatically-

accessible referents mismatched the morphological gender of the pronoun, as in the „no-

antecedent‟ conditions in (4c) and (4d), displayed longer reading times in the post-

pronoun region (corresponding to the two words immediately following the pronoun)

than in sentences in which there was a corresponding gender match, as in (4a) and (4b),

which was regarded as evidence that gender information is automatically used in

antecedent resolution. The proper names used in this study were not morphologically




                                                                                           10
marked for gender but were assumed to have conventional gender (Carreiras et al, 1996).

Other studies have revealed that definitional gender (i.e. king) and even stereotypical

gender (i.e. minister) influence antecedent resolution of reflexives (Kreiner, Sturt &

Garrod, 2008). Likewise, Cacciari et al (1997) highlight the importance of morphological

gender information in anaphor resolution in Italian, in which the gender of the antecedent

is explicitly marked and thereby aids referential resolution.

       The key finding in Badecker and Straub‟s (2002) Experiment 1 was that of longer

reading times in the post-pronoun region in multiple-match conditions in which both the

grammatically-accessible referent and the grammatically-inaccessible referent matched

the morphological gender of the pronoun, termed the „multiple-candidate effect‟, as in

(4a), than in single-match conditions in which only the grammatically-accessible referent

matched the pronoun‟s gender. Similar reading time differences were found for reflexives

governed by Principle A, and also for manipulations of morphological number. The

initial-filter model proposes that grammatically-inaccessible referents be immediately

discarded from the candidate antecedent set; however, Badecker and Straub‟s (2002)

results indicate that content-based properties of grammatically-inaccessible antecedents

affect antecedent resolution at least initially. The constraints of binding theory were

clearly not being imposed as an initial-filter on the candidate antecedent set. Instead, all

referents were receiving activation according to the congruence of their content-based

properties, in accordance with the predictions of the interactive-parallel constraint model.

       Sturt (2003) examined reflexives governed by Principle A using eye-tracking

methodology and similarly reported evidence of content-based properties affecting

antecedent resolution. Participants experienced processing difficulty when




                                                                                           11
grammatically-accessible antecedents mismatched the morphological gender of the

reflexive. Crucially, the gender of the grammatically-inaccessible discourse focused

antecedent also affected antecedent resolution, although this effect appeared substantially

later in processing than the effects of the grammatically-accessible antecedent, arising

only in second-pass and regression-path times, which was taken as evidence that this

effect was not part of the initial process of binding. Sturt (2003) proposed that binding

theory acts as an initial-filter on antecedent resolution, but may be overridden at a later

stage in favour of other non-structural constraints, thereby acting as a „defeasible filter‟.



The memory load hypothesis

       Badecker and Straub (2002) posited that the multiple-candidate effect might not

be due to content-based properties of the grammatically-inaccessible referent affecting

antecedent resolution, but instead may be due to an earlier effect imposed by

comprehending material prior to the pronoun. The effect might have arisen from the basic

difficulty of encoding two referents of the same gender as compared to encoding two

more easily distinguishable referents of different genders. Within the discourse

representation, different gender referents are supposedly more distinctly represented

compared to referents of the same gender (Garnham & Oakhill, 1985), a finding based on

the idea that it is easier to build and maintain a discourse representation in working

memory when the discourse referents within it are more distinguishable (Just &

Carpenter, 1992). Memory demands are indeed directly related to sentence

comprehension difficulty (Van Dyke & McElree, 2006). Encoding referents of the same




                                                                                              12
gender might result in a type of similarity-based interference within the discourse

representation.

       Danks (1986) and Magliano, Graesser, Eymard, Haberlandt and Gholson (1993)

claim that self-paced word-by-word reading tasks induce an artificial buffering strategy in

which participants demonstrate a tendency to traverse text at a rate that outpaces their

comprehension of complex material, which is instead subsequently processed towards the

end of the clause. In self-paced reading tasks, reading is constrained by the requirement

of response key pressing; the response keys cannot be pressed as fast as the eyes can

move and comprehend material, consequently readers invoke an artificial strategy of

simply rapidly pressing the response key and delay processing the material until an

appropriate moment arises (Danks, 1986). Badecker and Straub (2002) proposed that the

basic memory problems arising from the added complexity of encoding two referents of

the same gender might be delayed by the constraints of the self-paced reading task,

thereby resulting in the reading time differences observed in their original experiment.

Accordingly, similar differences ought to occur if the pronoun is replaced by another

proper noun and the number of gender congruent proper nouns within each sentence is

manipulated, as in (6a) and (6b). For ease of mention throughout this report this

prediction will henceforth be referred to as the „memory load hypothesis‟.



       (6a) John thought that Beth owed Jim another opportunity to solve the problem.

       (6b) John thought that Bill owed Jim another opportunity to solve the problem.




                                                                                           13
       If the multiple-candidate effect can instead be explained by the memory load

hypothesis, then this finding will no longer offer support for the parallel-interactive

constraint model. In fact, the findings in Badecker and Straub‟s (2002) Experiment 2

revealed a reading time difference only in the sentences containing the pronoun, as in (4a)

and (4b), thereby replicating the multiple-candidate effect. Manipulating the gender, and

thereby the distinguishability, of the proper nouns in (6a) and (6b) failed to invoke the

processing-load effects predicted by the memory load hypothesis, which thereby led to

the conclusion that the reading time differences found do reflect antecedent resolution

processes. Content-based properties of both grammatically-accessible and grammatically-

inaccessible referents can affect antecedent resolution, thus supporting the interactive-

parallel constraint model. Binding theory is apparently not acting as an initial-filter;

congruent properties of grammatically-inaccessible referents are allowing for such

referents to be included in the initial candidate antecedent set.



The present study

       There is a methodological concern with Badecker and Straub‟s (2002) Experiment

2 that must be considered before disregarding the memory load hypothesis entirely. If

participants are particularly sensitive to sentence complexity, it is feasible that the

materials used in their experiment were simply unsuitable for revealing differences in

encoding difficulty. If adding referents to the discourse representation increases memory

load, then adding a third referent in place of the pronoun can only add to that memory

load and increase any effects of similarity-based interference. The sentences exemplified

in (6a) and (6b) distinguish between either three referents of congruent gender, or of




                                                                                            14
different genders. However, since gender is dichotomous, the sentences containing

referents of different genders are biased towards one gender (i.e. two males and one

female). There may be an inherent difference between these conditions, and the pronoun

conditions which only contain two referents, and thereby the gender of the referents can

be equally distributed. It is possible that this added referent has breached some memory

load threshold in the materials in (6a) and (6b). Indeed, an examination of the data in

Badecker and Straub‟s (2002) Experiment 2 reveals that reading times in the critical

region (the two words following the third proper name) in the proper name sentences in

(6a) and (6b) are numerically comparatively similar to reading times in the corresponding

region in the multiple match pronoun sentences in (4a), all of which are higher than the

single match sentences in the pronoun condition in (4b).

       An alternative idea that might not breach the suggested threshold would be to

replace the pronoun with a non-referential pronoun that does not introduce a new referent

into the discourse, such as the indefinite pronoun everyone, which does not require an

antecedent and furthermore is gender-neutral. This would then allow a gender-

manipulation of the prior two proper names in the sentence comparable to that in the

referential pronoun conditions in (4a) and (4b); the indefinite pronoun cannot contribute

to the distinguishability of the referents encoded within the discourse representation since

it bears no gender. If the interactive-parallel constraint model is accurate, and the

difference in reading times found in Badecker and Straub‟s (2002) experiments are

attributable to the multiple-candidate effect arising from antecedent resolution processes,

then reading time differences should only occur between the definite pronoun sentences.

However, if the memory load hypothesis is correct, and if the additional complexity of




                                                                                           15
encoding two referents of the same gender in the discourse representation is contributing

to the difference in reading-times found in Badecker and Straub‟s (2002) experiments,

then reading times should be longer in sentences containing two referents of the same

gender in both the definite pronoun conditions and the indefinite pronoun conditions.




                                                                                        16
METHODS



Participants

         40 undergraduate students at the University of Edinburgh participated in the

experiment, with no reported history of language or reading disorders. All were native

speakers of English and had normal or corrected-to-normal vision.



Design

         The experiment consisted of a word-by-word self-paced reading task, followed by

a post-sentential probe-recognition task and intermittent yes-no comprehension questions.

These secondary tasks were included in order to test reading comprehension, and to

encourage participants‟ full attention to the content of the sentences. The experiment

follows the same structure as Badecker and Straub‟s (2002) Experiment 2.

         The main experimental task had four conditions, organised into a 2 x 2 design

with the within-subjects factors pronoun type (definite vs. indefinite pronoun) and match

type (single vs. multiple match). Example sentences from one complete set are shown in

(7a) to (7d). The full set of experimental items adheres to this basic structure, which is

based on the structure used in Badecker and Straub‟s (2002) Experiment 2.




                                                                                             17
(7a)   definite pronoun condition, single match

       Barbarai assumed that Michael saw heri disappear quietly from the party.

(7b)   definite pronoun condition, multiple match

       Barbarai assumed that Melissa saw heri disappear quietly from the party.

(7c)   indefinite pronoun condition, single match

       Barbara assumed that Michael saw everyone disappear quietly from the party.

(7d)   indefinite pronoun condition, multiple match

       Barbara assumed that Melissa saw everyone disappear quietly from the party.



       Sentences (7a) and (7b) constitute the definite pronoun conditions, containing the

referentially-dependent definite pronoun her/him bound to an antecedent located within

the sentence. Principle B of binding theory states that a pronoun must not be bound

locally within its governing category (the constraint corresponding approximately here to

the embedded clause it is located within). Accordingly, the antecedent of the definite

pronoun her is the grammatically-accessible (non-local) matrix subject Barbara, as

opposed to the grammatically-inaccessible (local) embedded subject Michael or Melissa.

Match conditions were varied by manipulating the gender congruency of the

grammatically-inaccessible proper name with the morphological gender of the definite

pronoun. In the single match condition, only the gender of grammatically-accessible

proper name matched the morphological gender of the definite pronoun. In the multiple

match condition, both the gender of the grammatically-accessible proper name and the

grammatically-inaccessible proper name matched the morphological gender of the

definite pronoun. Both the interactive-parallel constraint model and the memory load




                                                                                         18
hypothesis predicted slower reading times in the multiple match condition, thus the

multiple-candidate effect found by Badecker and Straub (2002) was expected to arise.

The interactive-parallel constraint model assumes that both binding theory and gender

congruency affect antecedent resolution in parallel, and thereby the gender of the

grammatically-inaccessible proper name was predicted to affect antecedent resolution,

resulting in slower reading times. The memory-load hypothesis assumes that two

discourse referents of the same gender are less distinguishable than referents of different

gender, and thereby predicted difficulty in encoding and maintaining the discourse

representation, which would be demonstrated by slower reading times.

       Sentences (7c) and (7d) comprised the indefinite pronoun conditions, containing

the non-referential indefinite pronoun everyone/everybody. Indefinite pronouns do not

refer to any specific entity and thereby require no antecedent to be bound to;

consequently, antecedent resolution does not occur in the indefinite pronoun conditions.

Match conditions were varied by manipulating the gender of the two proper names in the

sentence, thereby varying the distinguishability of the referents evoked by the discourse.

The indefinite pronouns selected were gender-unspecified, and thereby indefinite

pronoun match conditions were equivalent to the definite pronoun match conditions in

terms of the number of gender congruent referents. The multiple match condition

contained two proper names of the same gender, and the single match condition

contained two proper names of different genders. The interactive-parallel constraint

model predicted no difference in reading time between the single and multiple match

conditions, since no antecedent resolution is required. The memory load hypothesis

predicted that the multiple match condition would induce processing difficulty, resulting




                                                                                          19
in slower reading times, since the discourse referents within the discourse representation

are less distinguishable when they match in gender, and thus encoding the referents and

maintaining the discourse representation ought to be more difficult.



Materials

       The main experiment consisted of 24 experimental sentences (see appendix A),

four variants of each corresponding to the four conditions described in (7a) to (7d),

combined with 77 filler sentences of varying syntactic structures and lengths. Within the

experimental sentences, the morphological gender of the pronoun was balanced across

the definite pronoun condition items; half of the sets contained the feminine version of

the third person singular objective pronoun her, and half contained the masculine version

him. The sentences were constructed so that the definite pronoun her could not be

interpreted as a possessive pronoun referring to the embedded subject proper name. The

realisation of the indefinite pronoun was balanced across the indefinite pronoun condition

experimental items; half of the sets featured everyone, and half featured everybody. The

morphological gender of the definite pronoun controlled the order of the gender of the

preceding proper names within the sentence in single match conditions; grammatically-

accessible proper names were gender congruent and grammatically-inaccessible proper

names were gender-incongruent. In multiple match conditions both the proper names

matched the definite pronoun‟s morphological gender. The order and selection of the

proper names in the corresponding indefinite pronoun conditions was identical.

       Similar to Badecker and Straub‟s (2002) materials, proper names were selected as

opposed to full noun phrases, since proper names denote referents that are more




                                                                                           20
prominently represented within the discourse representation (Sanford, Moar, & Garrod,

1988). Selection of male and female gendered proper names was largely based on

intuition, excluding uncommon or gender-ambiguous names. Proper names in English are

not morphologically marked for gender; however, whether they are male or female is

generally considered to be an inherent property of the proper name itself (Carreiras et al,

1996). English proper names are thereby assumed to be conventionally gendered. For

each experimental set, all proper names were matched by letter length, and the gender-

manipulated male and female proper names in the embedded subject position were

matched by letter onset in the single and multiple match conditions.

       The secondary tasks consisted of a post-sentential probe recognition task and

intermittent comprehension questions, both of which required a “yes” or “no” response

and correct responses were balanced equally between the two. A single probe word was

created for each experimental set (see appendix A) and, where appropriate, a single

comprehension question was created for each of the experimental sets. Pronouns and

proper names were never selected as probe words or as the focus of the comprehension

questions. “Yes” probes were selected equally from content words located in sentence-

initial, sentence-medial, and sentence-final positions to avoid any potential cuing. “No”

probes were either semantic associates of content words in the sentence, morphological

neighbours of content words in the sentence, or were both semantically and

morphologically unrelated to words in the sentence. Comprehension questions were

developed for one quarter of the experimental and filler sentences, for which correct

responses never required antecedent resolution to occur within the sentence, so the proper

names or pronouns were never highlighted as of interest in the experiment.




                                                                                         21
Procedure

       The 24 experimental sentences were divided into four lists using a Latin-square

counterbalanced design; each list contained only one condition of each item, and all

conditions were equally represented. Each list was combined with the 77-filler item set

and pseudo-randomized so that no two experimental sentences occurred consecutively.

Participants were divided equally between the four lists.

       Participants were tested individually in a small and quiet room, each testing

session lasting approximately 20 minutes. The experiment was run on an Ergo Preceptor

3 laptop model N-30N3 using DMDX software (Forster & Forster, 2003). Participants

were presented with a set of written instructions informing them that they would be

partaking in a study investigating how people read different types of sentences.

Participants read each sentence one word at a time in a self-paced manner and responded

to probe words and comprehension questions. Each sentence was initially presented as a

series of white-lines on a black screen. Clicking the mouse-button initiated each trial, and

participants‟ repeated clicks of the mouse button resulted in successive word-by-word

presentation of the sentence. Words appeared in a white size 12 courier font on a black

background. Participants used their dominant hand for mouse clicks, of which response

times were recorded as corresponding to reading times of individual words.

       Immediately following the presentation of each sentence probe words appeared in

isolation in the centre of the screen, and participants pressed the left- or right-hand mouse

button for “yes” or “no” responses respectively (corresponding to the locations of a Y and

N on screen), to indicate whether they thought the probe was present in the just-read

sentence. On one quarter of the trials comprehension questions followed the probe words,




                                                                                          22
and a similar “yes” or “no” response was required. On trials without a comprehension

question, the experiment immediately continued to the initial presentation screen of the

following sentence. Feedback was not given in response to any of the tasks. Response

times in the probe recognition task and comprehension task were recorded.

        Two breaks were included in the experiment which participants could utilise if

required. Importantly, participants were instructed to read the sentences at a quick,

comfortable pace, but with care so as to be able to respond accurately on the question

tasks following each sentence.




                                                                                           23
RESULTS



Analysis

        Several different types of data were collected in this experiment. Reading time

data and probe recognition and comprehension question data were compiled. Probe words

and comprehension questions were included in the experiment to encourage participants‟

understanding and attention to the content of the experimental materials. Using Badecker

and Straub‟s (2002) criteria, only data from participants scoring above 80% accuracy on

comprehension and probe questions were included in the analyses. In this case, all

participants attained greater than 80% accuracy, and consequently all collected data was

included. Comprehension questions were only present in one quarter of the experimental

trials, and thus were not included in further analyses. Probe word accuracy and response

times were included in the main analyses.

        Condition means and standard errors of reading times for all words are reported.

Repeated measures analyses of variance with the factors pronoun type and match type

were computed for single words and for planned-text groupings, specifically the two

words following the pronoun on which past research has typically shown a spill-over

effect of antecedent resolution (e.g. Badecker & Straub, 1994; Badecker & Straub, 2002).

The ANOVAs were based on means computed for each participant (F1) and each item

(F2) in each condition. All statistical tests are reported at the p < 0.05 level, unless

otherwise indicated. Planned comparisons were carried out on any significant

interactions, comparing single and multiple match conditions within the definite pronoun

conditions and the indefinite pronoun conditions.




                                                                                           24
       For the purpose of discussion and clarification only, the sentence was divided into

the following regions:



Word 1.        Barbara (initial region)

Word 2.        assumed (initial region)

Word 3.        that (pre-critical region)

Word 4.        Melissa/Michael (pre-critical region)

Word 5.        saw (pre-critical region)

Word 6.        her/everyone (pronoun region)

Word 7.        disappear (critical region)

Word 8.        quietly (critical region)

Word 9.        from (post-critical region)

Word 10.       the (post-critical region)

Word 11.       party (post-critical region)



       The initial region contained the matrix subject (the grammatically-accessible

antecedent in the definite pronoun conditions) and the matrix verb, and was not included

in the analyses. It was assumed that reading times would not differ in this region since the

lexical material was identical in all conditions. Analyses were carried out on all regions

from the pre-critical region onwards, which comprised the words from the onset of the

complement clause up to the word preceding the pronoun. The lexical material did not

differ in this region, with the exception of the gender manipulation of the embedded

subject proper name. This region was thus included in the analyses in case any effects of




                                                                                             25
encoding multiple referents predicted by the memory load hypothesis arose immediately.

The pronoun region consisted of either the definite pronoun her/him or the indefinite

pronoun everyone/everybody. It was predicted that reading time differences would not

occur in this region, unless the processing difficulty of the multiple-candidate effect

appeared immediately, although previous findings suggest the effects of processing

referentially-dependent pronouns be somewhat delayed (e.g. Badecker & Straub, 2002;

Ehrlich & Rayner, 1983). The critical region contained the two words immediately

following the pronoun. It was expected that reading time differences would arise in this

region, indicative of either the multiple-candidate effect or the memory load hypothesis.

The post-critical region consisted of all remaining words in the sentence. This region was

included in case there was any effect spill-over from the prior critical region. Results

from within these established regions will now be discussed.



Pre-critical Region

       As expected there were no reading-time differences between the sentences in this

region, since this region did not differ between any conditions. For word 3 there was no

effect of pronoun type or match type (all F‟s < 1), and no interaction between the two

(F1(1,39) = 1.027, p > 0.05; F2(1,23) = 1.200, p > 0.05). For word 4 (the embedded

subject) there was no effect of pronoun type (F1(1,39) = 3.992, p > 0.05; F2(1,23) =

2.105, p > 0.05), no effect of match type (both F‟s < 1), and no interaction between the

two (both F‟s < 1). For word 5 there was no effect of pronoun type or match type, and no

interaction between the two (all F‟s < 1).




                                                                                           26
Pronoun Region

       From this region onwards one might expect some effects to arise since the

sentences differ in their content depending on condition. Word 6 was either the definite

pronoun her/him or the indefinite pronoun everyone/everybody. For word 6 there was a

significant main effect of pronoun type both by-participants (F1(1,39) = 6.806, p < 0.05)

and by-items (F2(1,23) = 7.562, p < 0.05). Response times were faster in the definite

pronoun conditions (single match = 417.50 ms; multiple match = 420.63 ms) than in the

indefinite pronoun conditions (single match = 460.63 ms; multiple match = 463.67 ms).

Match type was not significant, and there was no interaction between pronoun type and

match type (all F‟s < 1).



Critical Region

       Reading-time differences were expected to arise in this region. Planned-text

groupings of the critical region, combining word 7 and word 8, revealed no effect of

pronoun type or match type, and no interaction between the two (all F‟s < 1). Analyses

were also carried out on the two words separately. For word 7 there was no effect of

pronoun type (F1(1,39) = 1.050, p > 0.05; F2 < 1). The effect of match type was not

significant (both F‟s < 1), and there was no interaction between pronoun type and match

type (F1(1,39) = 1.233, p > 0.05; F2 < 1). For word 8 the effects of pronoun type and

match type were not significant, and there was no interaction between the two (all F‟s <

1). Standard errors in this region were very large, indicative of much variance in the

reading times (see Figures 1 and 2, and appendix B).




                                                                                           27
Figure 1. Word-by-word mean reading times in milliseconds for the single match and
multiple match sentences within the definite pronoun condition. Error bars indicate
standard errors of participant means.

                                         Definite Pronoun Conditions

                                       Single Match Condition           Multiple Match Condition
   Reading Time (ms)




                       600
                       550
                       500
                       450
                       400
                       350




                                                                                  m
                               a




                                                                       ly
                                                                         r
                                                        w




                                                                                        e
                                     at




                                                                                              r ty
                              ed




                                                 a




                                                                        r
                                                                     he
                             ar




                                                                                      th
                                                                     ea
                                               ss




                                                                              fr o
                                                     sa
                                   th




                                                                   iet
                           m
                           rb




                                                                                            pa
                                           eli




                                                                  pp

                                                                qu
                         su
                        Ba




                                                               sa
                                         l/M
                       as




                                                             di
                                      ae
                                   ich
                                   M




                                                            Word Position




Figure 2. Word-by-word mean reading times in milliseconds for the single match and
multiple match sentences within the indefinite pronoun condition. Error bars indicate
standard errors of participant means.

                                        Indefinite Pronoun Conditions

                                   Single Match Conditions              Multiple Match Conditions
   Reading Time (ms)




                       600
                       550
                       500
                       450
                       400
                       350
                                                                                  m
                               a




                                                                   ly
                                                        w




                                                                                        e
                                     at




                                                                                              r ty
                                                                 ne
                              ed




                                                 a




                                                                    r
                             ar




                                                                                      th
                                                                ea
                                               ss




                                                                              fr o
                                                     sa
                                   th




                                                               iet
                                                              yo
                           m
                           rb




                                                                                            pa
                                           eli




                                                             pp

                                                           qu
                         su
                        Ba




                                                           er

                                                          sa
                                         l/M
                       as




                                                        ev

                                                        di
                                      ae
                                   ich
                                   M




                                                            Word Position




                                                                                                     28
Post-Critical Region

       For word 9 there was no effect of pronoun type or match type (all F‟s < 1), and no

interaction between the two (F1 < 1; F2(1,23) = 1.222, p > 0.05). For word 10 there was

no effect of pronoun type (both F‟s < 1), and the effect of match type was not significant

(F1(1,39) = 1.538, p > 0.05; F2(1,23) = 1.073, p > 0.05). The interaction between the two

was not significant by-participants (F1(1,39) = 1.945, p > 0.05), but was marginally

significant in the by-items analysis (F2(1,23) = 3.562, p = 0.072).

       Two planned comparisons were carried out on the reading times for word 10 to

examine the predictions of the multiple-candidate effect and the memory load hypothesis

in the two pronoun conditions. The first comparison compared the reading times for the

single and multiple match conditions within the definite pronoun condition, exemplified

in (7a) and (7b). Reading times in the definite pronoun conditions were 43.17 ms longer

in the single match conditions than in the multiple match conditions, as shown in figure 1,

a difference that reached marginal significance both by-items and by-participants (single

match = 424.88 ms; multiple match = 381.71 ms; t1 (39) = 2.157, p < 0.05; t2(23) = 2.035,

p = 0.054). This is the opposite to the predictions of both the multiple-candidate effect

and the memory load hypothesis. The second comparison compared the reading times for

the single and multiple match conditions within the indefinite pronoun sentences,

exemplified in (7c) and (7d). There was no difference between the reading times in this

condition (single match = 387.29 ms; multiple match = 396.75 ms; t1 (39) = -0.365, p >

0.05; t2(23) = -0.437, p > 0.05).

       Only 23 items were included in the analysis for word 11, since one of the

sentences contained only 10 words. For word 11 there was no effect of pronoun type




                                                                                            29
(both F‟s < 1). The effect of match type was significant by-participants (F1(1,39) = 7.383,

p < 0.01), and was marginally significant by-items (F2(1,22) = 2.955, p = 1.00). Reading

times were consistently faster in the multiple match conditions (definite pronoun =

451.30 ms; indefinite pronoun = 454.04 ms) than in the single match conditions (definite

pronoun = 507.09 ms; indefinite pronoun = 506.22 ms). This result is contrary to the

predictions of the memory load hypothesis. There was no interaction between pronoun

type and match type (both F‟s < 1).



Figure 3. Mean response times in milliseconds for the probe recognition task for each
condition. Error bars indicate standard errors of participant means.

                                       Probe Word Response Times

                                                Single Match     Multiple Match

                        1550
   Response Time (ms)




                        1500
                        1450
                        1400
                        1350
                        1300
                        1250
                                     Definite Pronoun                   Indefinite Pronoun
                                                        Pronoun Condition




Probe Recognition

                        All participants scored greater than 80% accuracy on the probe words. For probe

word accuracy there was no effect of pronoun type or match type, and no interaction

between the two (all F‟s < 1). For the probe word response times there was no effect of

pronoun type (both F‟s < 1). The effect of match type was significant both by-participants


                                                                                                      30
(F1(1,39) = 4.382, p < 0.05) and by-items (F2(1,23) = 5.518, p < 0.05). Response times

were consistently faster in the single match conditions (definite pronoun =1306.63 ms;

indefinite pronoun = 1359.42 ms) than in the multiple match conditions (definite pronoun

= 1443.75 ms; indefinite pronoun = 1417.71 ms), as shown in Figure 3, which is the

direction predicted by the memory load hypothesis. There was no interaction between

pronoun type and match type (both F‟s < 1).




                                                                                         31
DISCUSSION



       The results of this experiment are somewhat complex, and therefore each finding

will be examined in turn. Results are discussed in terms of effects found in early and late

processing of the sentence (in terms of reading times on early and late words in the

sentence). Effects found late within the sentence are assumed to be related to processing

earlier material, given that the content following the pronoun did not differ between

conditions. Since a word-by-word analysis was conducted, the results of particular words

will be examined according to their sentence-location. Examining effects in early

processing first, this experiment did not explicitly replicate the multiple-candidate effect

found by Badecker and Straub (1994, 2002). There were no immediate differences

between the processing of the definite pronoun in sentences in which both the

grammatically-accessible and grammatically-inaccessible referents match the

morphological gender of the definite pronoun (multiple match), and sentences in which

only the grammatically-accessible antecedent matches the morphological gender of the

definite pronoun (single match), as demonstrated by the lack of effects found either in the

pronoun region or the critical region. Likewise, no effects were found on these regions

within the indefinite pronoun conditions, thereby offering no support for the memory load

hypothesis. There were no differences in processing the gender congruent or gender

incongruent second proper name in the pre-critical region either, indicating that effects of

the memory load hypothesis did not arise at least in early processing. Reading times on

the pronoun region within the definite pronoun conditions were faster than those within

the indefinite pronoun conditions, an effect attributable to the fundamental difference in




                                                                                           32
word length between the definite pronouns her/him and the indefinite pronouns

everyone/everybody (Just, Carpenter, & Wooley, 1982).

       Two somewhat conflicting effects in processing arose late in the sentence. Firstly,

reading times of words located within the post-critical region in a sentence containing

two proper names of different gender (single match) were longer than in a sentence

containing two proper names of the same gender (multiple match). This effect was firstly

restricted to the definite pronoun condition on word 10, but then extended to include the

indefinite pronoun condition also on word 11, which indicates that the effect is not

necessarily related to antecedent resolution. Secondly, the opposite effect was found on

response times in the probe recognition task, whereby probe responses were longer when

the sentence to be checked for correspondence contained two proper names of the same

gender (multiple match) than two proper names of different gender (single match). This

effect was not related to response accuracy, and thereby seems indicative of a delay in

processing. The first effect seems to oppose the memory load hypothesis, whereas the

second effect is in accordance with its predictions.

       The results of this experiment appear to suggest that early processing of the

referentially-dependent definite pronoun is not affected by content-based properties of the

grammatically-inaccessible antecedent. In either the pronoun region or the subsequent

critical region there was no difference in reading times between sentences in which the

grammatically-inaccessible referent matched or mismatched the morphological gender of

the definite pronoun. Participants did not seem to have more difficulty processing a

sentence in which the two proper names matched the gender of the definite pronoun;

hence the multiple-candidate effect was not replicated in this experiment. However, it




                                                                                           33
does not follow that content-based properties of the grammatically-inaccessible referent

did not affect processing at all, since there are later effects indicating that the gender of

the grammatically-inaccessible referent was in fact affecting resolution, hence it would be

inappropriate to reject the interactive-parallel constraint theory based on this finding

alone. Nevertheless, the grammatically-inaccessible referent did not appear to influence

antecedent resolution in early processing at least. Furthermore, the lack of effects found

in early processing in both the definite and indefinite pronoun conditions provides little

support for the memory-load hypothesis.

        Before forming any assumptions regarding whether binding theory is constraining

initial antecedent resolution, it is worth highlighting that, although not significant, there is

a numerical difference in the reading times from the definite pronoun condition in the

critical region; reading times were faster in single match than multiple match conditions.

This difference is in the direction that would be expected by the multiple-candidate

effect; however, there is a large amount of variability in the means which may account

for the lack of significant effects. Larger participant sample sizes might reveal the effect

more strongly. Comparing the data from this experiment with Badecker and Straub‟s

(2002) Experiment 2 data reveals that word-by-word reading times in this experiment

were consistently higher than those in Badecker and Straub‟s (2002). Higher reading

times are generally indicative of greater variability; indeed numerically, the standard

errors are much larger in this experiment than Badecker and Straub‟s (2002), which

might explain the failure to replicate the multiple-candidate effect.

        On the other hand, it has been proposed that readers do not resolve anaphoric

dependencies on-line unless task demands require such a resolution (Green, McKoon &




                                                                                                34
Ratcliff, 1992). It is possible that readers were simply not resolving the definite pronoun

since this was a more complex process than was required for the immediate task. The

probe words and comprehension questions were constructed so as to never require

antecedent resolution. However, later effects found within the sentence that can

potentially be explained in terms of antecedent resolution processes seem to cast doubt

over this explanation.

        Turning to effects in late processing, there were significant effects on several

words within the post-critical region, which consisted of all the words from the end of the

critical region to the end of the sentence. It must be pointed out that the length of this

region was not controlled for since no effects were expected to arise within it other than

potential spill-over effects from the critical region preceding it, thus it is difficult to make

any generalisations regarding the effects found on words within this region. There are

two potential explanations offered for the effects found within this region.

        The initial, albeit rather weak, effect within this region was restricted to the

definite pronoun conditions. Longer reading times were found on word 10 for sentences

in which only the grammatically-accessible referent matched the morphological gender of

the definite pronoun (single match) than for sentences in which both the grammatically-

accessible and grammatically-inaccessible referents matched the definite pronoun‟s

morphological gender (multiple match), an effect opposite to the predictions of the

memory load hypothesis. Similar effects have been found in other research. Using a

naming latency methodology, Rigalleau and Caplan (2000) found longer naming

latencies for pronouns in which the prior clause contained two proper names of different

genders, only one of which matched the morphological gender of the pronoun (as in the




                                                                                             35
single match condition), than when the clause contained two proper names of the same

gender, both of which matched the pronoun‟s morphological gender (as in the multiple

match condition). Sturt (2003), examining reflexives using eye-tracking methodology,

found a comparable difference in reading times on the pre-final region in the sentence,

which corresponds approximately to the post-critical in this experiment. This was

assumed to be related to sentence wrap-up effects, given that the effect occurred only in

regressions as opposed to in first-pass reading. It was proposed that discourse preferences

only affected antecedent resolution in later processing, whereby participants attempted to

bind the reflexive to the discourse-focused but grammatically-inaccessible referent, and

processing difficulty entailed only in the case of a gender-mismatch. On the contrary, in

the case of a gender-match, readers may have actually successfully bound the reflexive to

the grammatically-inaccessible referent, a notion supported by the greater number of

ungrammatical interpretations of such gender-match sentences in a follow-up experiment

(Sturt, 2003).

       The first explanation thus offered for the results found in the post-critical region is

that participants may have bound the definite pronoun to a grammatically-inaccessible

antecedent in sentences in which the grammatically-inaccessible referent matched the

morphological gender of the definite pronoun (multiple match). However, in the case of a

gender mismatch in the single match condition, in which the gender of the

grammatically-inaccessible referent did not match the morphological gender of the

pronoun, further processing was required in order to bind the pronoun to the

grammatically-accessible gender-congruent referent. The grammatically-inaccessible

referent may have been more focused in the discourse representation, thereby the




                                                                                           36
preferential antecedent, although the sentences were constructed with the aim of giving

both referents within the sentence equal salience, since both were grammatical subjects,

which are generally regarded as more prominent (Järvikivi, Van Gompel, Hyönä, &

Bertram, 2005). There is a common assumption that there is an advantage of first-

mention, whereby the first mentioned referent is easier to access from the discourse

representation than later-mentioned referents (e.g. Gernsbacher & Hargreaves, 1988;

Gernsbacher, Hargreaves, & Beeman, 1989), which seems to contradict this explanation

of salience. On the other hand, it is equally possible that the embedded subject referent

was being regarded as more salient due to the simple advantage of recency (Gernsbacher

et al, 1989). Research has found that responses to a probe word located in the second of

two clauses are faster than if the probe word is located in the first clause (Chang, 1980;

Von Eckardt & Potter, 1985). The grammatically-inaccessible referent, located in the

embedded clause, was most recently processed and may thereby be receiving the most

activation resulting in higher salience and easier access within the discourse

representation, therefore readers may initially attempt to bind the definite pronoun to it.

       If this explanation is appropriate, antecedent resolution appears to have been

delayed somewhat towards the end of the sentence, arising later in processing than was

found by Badecker and Straub (2002). The delay in antecedent resolution may of course

be due to the artificial buffering strategy, as suggested earlier (Danks, 1986; Magliano et

al, 1993), whereby readers traverse text in a self-paced reading experiment at a rate faster

than they comprehend, and so delay antecedent resolution until later in the sentence. It

must be noted that the explanation just offered is particularly tentative, since the effect

arising on this word was only marginally significant. Further research, perhaps




                                                                                              37
investigating which of the referents is receiving most activation, would be required in

order to form any further conclusions. Conversely, it might be that antecedent resolution

itself was not delayed, but that it occurred in two stages. It is possible that binding theory

was acting immediately in antecedent resolution, but that the constraint was violated later

in processing, as suggested by the defeasible-filter view (Sturt, 2003). The self-paced

methodology used does not sufficiently separate effects of early and late measures, thus

eye-tracking methodology might be more suitable for investigating this suggestion

further.

           On the following word in the sentences (word 11), the effect just explained

expanded to include the indefinite pronoun conditions. Sentences including two proper

names of the same gender had faster reading times for word 11 than sentences including

two proper names of different genders, irrespective of whether the sentence contained a

referentially dependent pronoun. As pointed out earlier, sentence length in this region

was not controlled for, since no effects were expected other than possible spill-over

effects from the critical region. However, with the exception of a small number of

sentences, this was either the final word in the sentence or the penultimate word. Longer

reading times are generally found on end-of-sentence words, indicative of integrative

processes (Just et al, 1982). Participants appear to have experienced more difficulty

integrating material from sentences containing referents of different gender than of the

same gender, an effect contrary to the predictions of the memory load hypothesis.

However, there is an alternative, equally speculative, explanation for the effects found

within the post-critical region, that of gender-based priming. This explanation is based on

findings that the response latency to a target word can be facilitated following a gender-




                                                                                            38
congruent word as compared to a gender-incongruent word (Banaji & Hardin, 1996). It

seems reasonable to suggest that when comprehending sentences containing two proper

names of the same gender, participants‟ encoding of the second gender-congruent proper

name was subject to priming, since its content-based properties matched those of the first

proper name, thereby potentially aiding encoding. Given that no baseline condition was

present in this experiment, no conclusions can be inferred as to whether the priming was

facilitative or inhibitory. Nevertheless, building and maintaining a discourse

representation containing referents that are all the same gender may well be comparably

easier than if the discourse contains referents that are both male and female, since the

gender feature primes further encoding of gender-congruent referents.

       This effect of gender-based priming may either be separate to the effect found on

the previous word, or it may be that the same effect is occurring on both words, but that it

simply arose earlier within the definite pronoun conditions. Indeed, the pronoun is also

gendered and thereby matches the gender of at least one of the proper names, conceivably

contributing to the priming effect and potentially facilitating encoding. It is slightly

unclear, of course, as to why this potential priming effect arose only in later processing

rather than exhibiting priming on the actual proper names or on the gender-marked

definite pronoun. However, if an artificial buffering strategy is being used by participants

as suggested earlier, then it is plausible that processing might not take place until towards

the end of the sentence (Danks, 1986; Magliano et al, 1993).

       The final effect found in this experiment occurred in the probe recognition task

that immediately following the presentation of each sentence. Probe recognition

responses were slower when participants had to decide whether the probe was located in




                                                                                             39
a sentence containing two referents of the same gender (multiple match) than a sentence

containing two referents of different genders (single match). This effect occurred in both

the definite and indefinite pronoun conditions, so it cannot be related to antecedent

resolution processes. A type of similarity-based interference seems to have taken place.

Probe recognition accuracy was not affected; participants performed equally well in all

conditions, however the decision of whether the probe was present in the just-read

sentence appears to have been delayed in the multiple match conditions. This finding

supports the memory-load hypothesis, although the effect arose much later than expected,

and indeed not even within the actual sentence being processed. Nevertheless,

participants did appear to be experiencing processing difficulty when the referents within

the sentence were of the same gender, and thereby less distinguishable.

       One explanation is that processing whether the probe word was present in the just-

read sentence involves the same factors as initially processing the discourse and building

and maintaining the corresponding discourse representation (Garnham & Oakhill, 1985).

The self-paced reading methodology may not have been sensitive enough to reveal such

effects when participants were initially reading the sentence, potentially due to the use of

an artificial buffering strategy (Danks, 1986; Magliano et al, 1993); however when

participants had to make a firm decision in order to accurately respond to the probe

recognition task, and they had time to do so, the effect arose. The probe words never

corresponded to either the pronoun or the proper names; however, “yes” probes were

selected equally from initial, medial, and final locations within the sentence in order to

avoid cuing, and so it is plausible that participants may have been processing the entire

sentence to establish whether the probe word was present. Half of the probes were “no”




                                                                                             40
probes and thus not located within the sentence, which further supports the idea that

sentence checking must have occurred since the effect clearly arose on these probes also.

The difficulty of such sentence checking may thus depend on “the ease with which

information can be read out of a (referentially determinate) representation that has been

constructed as the sentence was read” (Garnham & Oakhill, 1985, p. 395). It is easier to

maintain such a discourse representation when the events within it are easier to

distinguish (Just & Carpenter, 1992). Two referents of the same gender are less

distinguishable than two referents of different gender, since they cannot be identified in

terms of gender information alone, and thus re-accessing the discourse representation

may be harder and cause more processing difficulty, resulting in the longer response

times.

         That this effect arose on the probe recognition task even though the task was not

directly examining the discourse referents suggests that it is perhaps a particularly

dominant effect. Warren and Gibson (2005) report a related effect examining different

types of noun phrases, finding that comprehension accuracy decreased when post-

sentential comprehension questions were about sentences containing two noun phrases

that matched in type (e.g. name and name) as opposed to differing in type (e.g.

description and name). The similarity of the noun phrases appears to have heightened

interference in processing, in a manner comparable to the way in which the similar

content-based properties of proper names in this experiment decreased the

distinguishability of the corresponding referents within the discourse representation,

thereby increasing interference, resulting in processing difficulty.




                                                                                            41
       Although mere conjecture, if the explanation of gender-based priming is

appropriate to explain the effects found within the post-critical region (as opposed to

antecedent resolution processes), and since support was found for the memory-load

hypothesis on probe response times, one might potentially conclude that the multiple-

candidate effect found in Badecker and Straub‟s (2002) Experiment 2 was perhaps in fact

an effect of the memory-load hypothesis, as Badecker and Straub (2002) initially

proposed. Probe response times were not given by Badecker and Straub (2002), so there

is no way of knowing whether a similar effect arose on the probe words. The change in

materials might have been sufficient to bring about the effect, supporting the suggestion

of the existence of some distinguishability threshold breached in Badecker and Straub‟s

(2002) materials. Stipulating that the multiple-candidate effect was in fact an effect of the

memory load hypothesis eradicates the support it was thought to offer for the interactive-

parallel constraint model. If the word-by-word self-paced reading methodology is indeed

inducing an artificial buffering effect that is delaying sentence processing, perhaps the

effects found within this study would arise earlier if a more sensitive methodology which

does not induce the strategy, such as eye-tracking, was used to investigate the same

effects. As suggested earlier, this technique would also allow much greater accuracy in

determining which effects arise in early or late processing, as could be determined by

differences in first- or second-pass reading. Alternatively, if the effects found within the

post-critical region were in fact related to antecedent resolution processes, then this offers

support for the proposal that content-based properties of grammatically-inaccessible

antecedents affect resolution, as suggested by the interactive-parallel constraint model.

Furthermore, the slower reading times in this experiment resulted in much greater




                                                                                            42
variability, which might have concealed any credible effects that could have offered

support for the multiple-candidate effect and thereby the interactive-parallel constraint

model.

         In summary, this experiment failed to replicate the multiple-candidate effect. In

early processing, the content-based properties of grammatically-inaccessible antecedents

did not affect antecedent resolution. Similarly, there was no supportive evidence of the

memory-load hypothesis in early processing either. Effects arose in later processing

towards the end of the sentence which potentially indicate that readers were attempting to

bind the referentially-dependent pronoun to the most recently processed but

grammatically-inaccessible referent, unless in the case of a gender-mismatch, at which

point further processing was required. An alternative explanation offered was that

gender-based priming was occurring, whereby sentences containing referents of

congruent gender were primed and thereby easier to encode than those with referents of

incongruent gender. It is beyond the scope of this report to establish which of these

explanations is appropriate. Finally, support for the memory-load hypothesis was found

on the probe recognition task, in which responses were longer if the probe followed a

sentence containing two referents of the same gender. This is explicable if one assumes

that the same processes that occur when initially processing the sentence similarly occur

in later sentence-checking.




                                                                                             43
REFERENCES



Arnold, J.E., Eisenband, J.G., Brown-Shmidt, S., & Trueswell, J.C. (2000). The rapid use

       of gender information: evidence of the time course of pronoun resolution from

       eyetracking. Cognition, 76, B13-B26.



Baddeley, A. (2007). Working Memory, Thought, and Action. Oxford: Oxford University

       Press.



Badecker, W., & Straub, K. (1994). Evidence that binding principles participate in a

       constraint satisfaction process. Poster session presented to the Seventh Annual

       CUNY Sentence Processing Conference, New York.



Badecker, W., & Straub, K. (2002). The processing role of structural constraints on the

       interpretation of pronouns and anaphors. Journal of Experimental Psychology, 28,

       748-769.



Banaji, M.R., & Hardin, C.D. (1996). Automatic stereotyping. Psychological Science, 7,

       136-141.



Cacciari, C., Carreiras, M., & Cionini, C.B. (1997). When words have two genders:

       Anaphor resolution for Italian functionally ambiguous words. Journal of Memory

       and Language, 37, 517-532.




                                                                                          44
Carreiras, M., Garnham, A., Oakhill, J., & Cain, K. (1996). The use of stereotypical

       gender information in constructing a mental model: Evidence from English and

       Spanish. The Quarterly Journal of Experimental Psychology, 49A, 639-663.



Chang, F.R. (1980). Active memory processes in visual sentence comprehension: Clause

       effects and pronominal reference. Memory and Cognition, 8, 58-64.



Chomsky, N. (1981). Lectures on Government and Binding. Dordrecht, Netherlands:

       Foris.



Clifton, C., Kennison, S.M., & Albrecht, J.E. (1997). Reading the words her, his, him:

       Implications for parsing principles based on frequency and on structure. Journal

       of Memory and Language, 36, 276-292.



Danks, J.H. (1986). Identifying component processes in test comprehension: Comment

       on Haberlandt and Graesser. Journal of Experimental Psychology: General, 115,

       193-197.



Ehrlich, K. (1980). Comprehension of pronouns. Quarterly Journal of Experimental

       Psychology, 32, 247-255.




                                                                                         45
Ehrlich, K., & Rayner, K. (1983). Pronoun assignment and semantic integration during

       reading: Eye movements and immediacy of processing. Journal of Verbal

       Learning and Verbal Behaviour, 22, 75-87.



Forster, K.I., & Forster, J.C. (2003). DMDX: A Windows display program with

       millisecond accuracy. Behaviour Research Methods, Instruments & Computers,

       35, 116-124.



Garnham, A., & Oakhill, J. (1985). On-line resolution of anaphoric pronouns: Effects of

       inference making and verb semantics. British Journal of Psychology, 76, 385-393.



Gernsbacher, M. (1989). Mechanisms that improve referential access. Cognition, 32, 99-

       156.



Gernsbacher, M., & Hargreaves, D.J. (1988). Accessing sentence participants: The

       advantage of first mention. Journal of Memory and Language, 27, 699-717.



Gernsbacher, M., Hargreaves, D.J., & Beeman, M. (1989). Building and accessing

       clausal representations: the advantage of first mention versus the advantage of

       clause recency. Journal of Memory and Language, 28, 735-755.



Gordon, P.C., & Hendrick, R. (1997). Intuitive knowledge of linguistic co-reference.

       Cognition. 62, 325-370.




                                                                                         46
Greene, S., McKoon, G., & Ratcliff, R. (1992). Pronoun resolution and discourse models.

       Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 266-

       283.



Järvikivi, J., Van Gompel, R.P.J., Hyönä, J., & Bertram, R. 2005. Ambiguous pronoun

       resolution: contrasting the first-mention and subject-preference accounts.

       Psychological Science, 16, 260-264.



Just, M.A., & Carpenter, P.A. (1992). A capacity theory of comprehension: Individual

       differences in working memory. Psychological Review, 99, 122-149.



Just, M.A., Carpenter, P.A., & Wooley, J.D. (1982). Paradigms and processes in reading

       comprehension. Journal of Experimental Psychology: General, I11, 228-238.



Kreiner, H., Sturt, P., & Garrod, P. (2008). Processing definitional and stereotypical

       gender in reference resolution: Evidence from eye-movements. Journal of

       Memory and Language, 58, 239-261.



Lucas, M.M., Tanenhaus, M.K., & Carlson, G.N. (1990). Levels of representation in the

       interpretation of anaphoric reference and instrument inference. Memory and

       Cognition, 18, 611-631.




                                                                                         47
MacDonald, M.C., & MacWhinney, B. (1990). Measuring inhibition and facilitation from

       pronouns. Journal of Memory and Language, 29, 469-492.



MacDonald, M.C., Pearlmutter, N.J., & Seidenberg, M.S. (1994). Lexical nature of

       syntactic ambiguity resolution. Psychological Review, 101, 676-703.



Magliano, J.P., Graesser, A.C., Eymard, L.A., Haberlandt, K, & Gholson, B. (1993).

       Locus of interpretive and inference processes during text comprehension: A

       comparison of gaze durations and word reading times. Journal of Experimental

       Psychology: Learning, Memory, and Cognition, 19, 704-709.



McKoon, G., Ward, G., Ratcliff, R., & Sproat, R. (1993). Morphosyntactic and pragmatic

       factors affecting the accessibility of discourse entities. Journal of Memory and

       Language, 32, 56-75.



Nicol, J. (1988). Coreference processing during sentence comprehension. Unpublished

       doctoral dissertation, Massachusetts Institute of Technology.



Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment during

       sentence comprehension. Journal of Psycholinguistic Research, 18, 5-19.



Reinhart, T. (1976). The Syntactic Domain of Anaphora. Ph.D. thesis, MIT.




                                                                                          48
Rigalleau, F., & Caplan, D. (2000). Effects of gender marking in pronominal

       coindexation. The quarterly journal of experimental psychology, 53A, 23-52.



Sanford, A., Moar, K., & Garrod, S. (1988). Proper names as controllers of discourse

       focus. Language and Speech, 31, 43-56.



Sturt, P. (2003). The time-course of the application of binding constraints in reference

       resolution. Journal of Memory and Language, 48, 542-562.



Trueswell, J.C. (1996). The role of lexical frequency in syntactic ambiguity resolution.

       Journal of Memory and Language, 35, 566-585.



Van Dyke, J.A., & McElree, B. (2006). Retrieval interference in sentence

       comprehension. Journal of Memory and Language, 55, 157-166.



Van Gompel, R.P.G., & Liversedge, S.P. (2003). The influence of morphological

       information on cataphoric pronoun assignment. Journal of Experimental

       Psychology: Learning, Memory, and Cognition, 29, 128-139.



Van Gompel, R.P.G., & Majid, A. (2004). Antecedent frequency effects during the

       processing of pronouns. Cognition, 90, 255-264.




                                                                                           49
Von Eckardt, B., & Potter, M.C. (1985). Clauses and the semantic representation of

       words. Memory and Cognition, 13, 371-354.




                                                                                     50
APPENDICES



APPENDIX A: Experimental materials



The 24 experimental sentences used in the experiment are listed below. Sentences listed

are from the single match/definite pronoun condition. Alternative proper names for the

multiple match variants and the alternative pronoun for the indefinite pronoun variants

are given in parentheses. The probe words from the probe recognition task that were

presented following each sentence are given in parentheses at the end of each complete

sentence.



1. John mentioned that Beth (Bill) cooked him (everyone) spaghetti bolognaise for dinner

that night. (Mentioned)



2. Henry explained that Jenny (Jacob) brought him (everyone) beautiful daffodils last

week. (Flowers)



3. Brian complained that Helen (Harry) packed him (everyone) coleslaw sandwiches for

lunch that day. (Coleslaw)



4. Bruce insisted that Julia (Jason) sold him (everybody) inexpensive furnishings for the

house. (Insist)




                                                                                          51
5. Gordon demanded that Rachel (Robert) pay him (everyone) substantial compensation

for the accident. (Accident)



6. Timothy proposed that Miranda (Malcolm) lend him (everyone) champagne glasses for

the party. (Umbrella)



7. Donald remembered that Louise (Lionel) challenged him (everyone) somewhat

stupidly to a fight. (Remembered)



8. Roger said that Alice (Aaron) gave him (everybody) wonderful presents for Christmas

this year. (Gifts)



9. James revealed that Laura (Lewis) passed him (everybody) answers discretely in the

exam hall. (Passed)



10. Greg believed that Jane (Jack) asked him (everybody) particularly difficult questions

in the test. (Difficulty)



11. Richard explained that Jessica (Jeffrey) offered him (everyone) substantial amounts

of money as a bribe. (Money)




                                                                                          52
12. Frank stated that Sarah (Simon) sent him (everybody) important documents in the

post that day. (Cheese)



13. Samantha announced that Jonathon (Jennifer) invited her particularly enthusiastically

to the party tonight. (Announced)



14. Nancy thought that Peter (Penny) owed her (everybody) another opportunity to solve

the problem. (Trouble)



15. Anne suggested that Tony (Tina) get her (everyone) another quotation for the

building works. (Quotation)



16. Sarah admitted that Frank (Fiona) offered her (everybody) additional lessons before

the exam. (Offer)



17. Polly appreciated that Jason (Janet) called her (everyone) extremely remorsefully last

night to apologise. (Night)



18. Alison explained that Trevor (Trisha) appreciated her (everyone) asking questions

about the incident. (Prepare)




                                                                                        53
19. Angela imagined that George (Gloria) made her (everyone) delicious chocolate tarts

for dessert. (Imagined)



20. Barbara assumed that Michael (Melissa) saw her (everyone) disappear quietly from

the party. (Vanish)



21. Maggie divulged that Thomas (Teresa) loaned her thirteen thousand pounds at the

races. (Thirteen)



22. Sheila implied that Andrew (Amanda) told her (everyone) confidential information

about the case. (Cases)



23. Sally guessed that Tommy (Tanya) overheard her (everybody) talking loudly about

the surprise party. (Surprise)



24. Lauren requested that Steven (Sophie) advance her (everybody) several hundred

pounds of this month‟s pay. (Concert)




                                                                                       54
APPENDIX B: Statistics for word reading times and probe responses



The participant means of word-by-word reading times and probe word response times in

milliseconds are given below for each condition. Standard errors of participant means are

given in parentheses.



                         Definite Pronoun Conditions      Indefinite Pronoun Conditions
                                  (Figure 1)                         (Figure 2)
                        Single Match Multiple Match      Single Match Multiple Match
                         Condition        Condition        Condition         Condition
Barbara assumed
That                      416 (19)        420 (22)          423 (19)         404 (20)
Michael/Melissa           456 (35)        473 (39)          440 (25)         436 (40)
Saw                       448 (24)        432 (24)          447 (28)         454 (25)
Her/Everyone              417 (20)        420 (21)          460 (25)         463 (34)
Disappear                 527 (54)        484 (36)          523 (34)         527 36)
Quietly                   530 (38)        536 (55)          524 (24)         518 (28)
From                      437 (22)        452 (22)          441 (17)         441 (19)
The                       424 (27)        381 (15)          387 (17)         396 (28)
Party                     507 (35)        451 (23)          506 (39)         454 (22)
PROBE                    1307 (45)       1444 (75)         1359 (47)        1418 (64)




                                                                                        55