1.1 Previous research on anaphora resolution

Prior work on AR has resulted in a great number of theoretical accounts that uniformly assume that, when the language under investigation allows for null subjects, the most prominent/salient entities are referenced with the most reduced anaphoric expression, i.e., the null form. Although the availability of null and overt pronouns is a grammatical property in null subject (pro-drop) languages, final interpretation of the subject pronoun (or pro) is regulated by various discourse factors. To this aim, researchers examined discourse notions such as Topicality (Givón 1983), Accessibility (Ariel 1990), Givenness (Gundel et al. 1993), or Informativeness (Allen 1997) to illustrate a hierarchy associating salient entities with the use of pronominal reference.

Accordingly, Arnold formulated the Expectancy Hypothesis through the investigation of the first-mention advantage. In two visual-world paradigm experiments Arnold, Eisenband, Brown-Schmidt & Trueswell (2000) investigated adults’ on-line pronoun resolution: first mention and recency were the factors examined as potential contributors to referent accessibility. In their first experiment, the two competing antecedents either matched or not in gender: In the same gender condition, the pronoun was temporarily ambiguous since it could refer to either entity, although visual cues provided evidence for only one possible antecedent. Results revealed that when the gender matched and the pronoun referred to the second-mentioned entity, gaze data differed from all the other conditions in that participants looked equally at both subject and object antecedents, although the target was the object one. No differences were found across conditions in the accuracy rates of the comprehension questions. In their second experiment, the authors increased the accessibility of the first-mentioned antecedent adding an interim sentence, where a pronoun repeated reference to this participant. They found that on-line processing of the target picture was enhanced across conditions in that effects appeared earlier than in the experiment with no interim sentence. In this experiment, however, accuracy decreased when the pronoun referred to the second-mentioned antecedent. Results suggested that participants expected the pronoun to refer back to the first-mentioned entity, unless grammatical cues (i.e., matching of the gender) biased interpretation to the object antecedent. Noticeably, it is not easy to conclude whether it is first-mention or subjecthood that guides interpretation in English, which was the language used in these experiments. Indeed, we may assume that a match between the grammatical role of the subject pronoun and the antecedent, i.e., a parallelism effect, is the one directing the processing and interpretation of the pronoun (Stevenson et al. 1995).

Much of previous research on AR on null subject languages such as Greek, Italian and Spanish focuses on the interpretation of null and overt subject pronouns in intra-sentential anaphora contexts (e.g., Tsimpli et al. 2004; Filiaci 2010). The findings typically converge on the conclusion that the most salient entities are referenced with the most reduced anaphoric expression, namely the null pronoun. Carminati (2002) suggested that null pronouns in Italian show a bias for subject antecedents, while overt pronouns prefer the less prominent discourse entities (i.e., usually the non-subject antecedent), as they are more informative. Carminati’s proposal has been supported by similar findings from Italian, Greek, Spanish and Catalan (e.g., Dimitriadis 1996; Alonso-Ovalle, Fernandez-Solera, Frazier & Clifton 2002; Miltsakaki 2002; 2007; Filiaci 2010; Jegerski et al. 2011; Keating, VanPatten & Jegerski 2011) although some studies have also found cross-linguistic differences among null subject languages (Tsimpli et al. 2004; Filiaci 2010), and even intra-language differences (Carminati 2002; but see also Tsimpli et al. 2004). Overall, these findings suggest that overt pronouns trigger a preference for non-salient antecedents in Italian, Catalan and Greek vs. Spanish, with Spanish not exhibiting a strong preference.

Previous evidence reporting clear differences between two overt pronouns comes from Turkish (e.g., Gürel 2003; 2004). Gürel tested native speakers of Turkish with sentences containing a null pronoun (∅ or pro), or the Turkish overt forms o (‘s/he’) and kendi- (‘self’). The author found that o only allowed disjoint reading, whereas the null pronoun and kendi- were unconstrained in binding, bringing up similarities to Greek. More specifically, Gürel suggested participants interpreted the overt o pronoun as referring to a non-subject entity, whereas null and kendi- pronouns were equally interpreted as the subject of the main clause or another entity (not) included in the previous context.

1.2 Age differences and cognitive decline in anaphora resolution

The vulnerability of the interfaces vs. that of pure grammatical aspects of language has been put forward by previous research in bilingualism and attrition (e.g., Tsimpli et al. 2004; Sorace & Filiaci 2006), since processing demands the coordination of various types of information, grammatical and discourse linked ones. Interpretation of reference presupposes cognitive control in the manipulation of linguistic (morpho-syntactic) and extra-linguistic (pragmatic) factors in real time. When competing antecedents are involved, it can be a process of disambiguation more difficult than structural ambiguity. Thus, when grammar does not disambiguate, as in the case of different gender cues, AR illustrates preferences and not categorical judgments, which however involve the integration of linguistic information (the form of the expression, the search for the optimal antecedent, the linguistic properties of the sentences in which the form and the antecedent appear), world knowledge and processing capacity. Under this rationale, AR should provide evidence of differences in the way people interpret specific anaphoric expressions and whether interpretations vary according to one’s cognitive abilities and whether they change over the years.

The idea stems from the fact that people with a high working memory (hence WM) capacity may be more sensitive to sentence ambiguity (e.g., Dwivedi & Golhawk 2009). Besides, WM capacity has been correlated with readers’/listeners’ ability to override a first, incorrect analysis of a sentence in view of new, conflicting information in sentences which include local ambiguity (as in garden-path structures) (Novick, Trueswell, & Thompson-Schill 2005). Crucially, separable syntactic, semantic and phonological components to WM, all of them being specialized WM capacities have been associated with the maintenance of different types of information (e.g., Waters & Caplan 2001). Adding to this, age related effects are attested in sentence comprehension of garden-path structures (e.g., Stine-Morrow, Ryan & Leonard 2000; DeDe, Caplan, Kemtes & Waters 2004) and are often correlated with measures of tasks tapping into WM abilities (e.g., Christianson, Williams, Zacks & Ferreira 2006), although no clear-cut relation between sentence comprehension and WM capacity is always reported.

Under this rationale, van Rij et al. (2013) implemented a computational model simulating the interpretation of subject pronouns. They suggested that WM load reduces individuals’ sensitivity to the various discourse related cues, resulting to a divergent interpretation of the pronouns examined. Their empirical testing of the model confirmed that WM constraints affect adults’ pronoun resolution. Actually, in a dual-task experiment, they manipulated WM load (with a digit task) and topic-shift (in a four sentence story-reading task, using a moving-window paradigm). They analysed online and off-line measurements and found that in the high WM load condition participants selected less often the subject of the first sentence only in stories with topic-shift, although story type yielded no differences in response times or during processing (contrary to their expectations). Moreover, WM load rendered the reading of stories slower in both story types.

Finally, Hendriks et al. (2014), examining referential choice across lifespan, suggested that WM and age-related effects condition AR. More specifically, using picture-based stories, the authors suggested that, in production, older adults were less strategic in encoding/retrieving the most important information from the discourse than younger ones and, therefore often overused ambiguous pronouns. They attributed their results to reduced cognitive capacities which help keep track of the prominence of discourse referents (see also Hendriks, Englert, Wubs & Hoeks 2008 for a theoretical explanation based on the framework of bidirectional Optimality Theory; bi-OT).⁵ In comprehension instead, they found that, as in other studies (e.g., Light & Capps 1986; Titone, Prentice & Wingfield 2000), elderly adults did not appear to have problems with disambiguating the pronoun on the basis of the information presented in the immediately preceding sentence. Interestingly, both production and comprehension results correlated positively with participants’ WM scores, in the case of children and elderly adults, while no such evidence was found for young adults (as in the case of van Rij et al. 2013). Similarly, age differences across children and adults were demonstrated in Papadopoulou, Peristeri, Plemenou, Marinis & Tsimpli (2015), who investigated the interpretation of null and overt anaphoric pronouns in children and adult Greek native speakers, using a self-paced listening and judgment task combined with pictures that matched in grammaticality. The authors found that the overt pronoun elicited a strong preference for the object antecedent in both children and adults, whereas children differed from adults in allowing null pronouns to refer to either subject or object antecedents; divergent anaphora resolution was associated with differences in language experience.⁶

Overall, these results shed some light to the idea that age-related individual differences may play a role in AR especially when additional linguistic factors increase complexity and the disambiguation of the pronoun demands more cognitive resources.

1.3 Word-order and anaphora resolution in Greek

Greek is a head-initial pro-drop language with a rich inflectional system that distinguishes subject and object positions with morphological case (nominative and accusative, respectively): Grammatical functions are marked with morphological case while subjects and objects show relative freedom in word-order patterns (Roussou & Tsimpli 2006). According to an early account (Philippaki-Warburton 1982), the basic word-order in Modern Greek is verb-subject-object (VSO; see 1a), implying that the subject occupies its canonical subject position only in this word-order. Tsimpli (1990; 1995) also suggested that the subject occupies a position lower than the verb. In the subject-verb-object order (SVO; see 1b) the subject occupies a Topic position which is adjoined to the Tense projection, and the same analysis is offered for the verb-object-subject (VOS; see 1c), the only difference being in the directionality of the adjunction which in VOS is to the right of the Tense projection. In both of these word-orders the grammatical subject is a pro, occupying the canonical subject position while the overt subject is a topic. In fact, a postverbal subject is possible in a wide-focus question of the ‘What happened?’ type. Nevertheless, the most frequent word-order in corpora is SVO (Lascaratou 1989), which has been shown to be easier to process (Kail & Diakogiorgi 1994).

One of the word-order patterns associated with changes in information structure (i.e., object topicalization or object focusing) is object-verb-subject (OVS; see e.g., Tsimpli 1995; 1998; Alexiadou 1999). In (2a), the OVS order includes a preposed object (e.g., ton Kosta = ‘the.ACC Kostas.ACC’) which may be interpreted as given information (i.e., already mentioned in the previous context) and topicalised while doubled by a pronominal clitic attached to the verb. Although the OVS order in Greek is also found in object focusing (2b), it is only in topicalization structures such as (2a) that a pronominal clitic is required to double (the structure known as clitic-left-dislocation; CLLD).⁷ The marked nature of the OVS order makes it more demanding in processing, albeit frequent in use.

Finally, notice that although Greek bears similarities to Italian, two pro-drop languages which allow postverbal subjects, they differ in that a vP peripheral New Information Focus position hosting postverbal subjects (Belletti 2009) is assumed only for Italian. In fact, a subject DP and an object DP cannot both occur in the same domain in Italian, but they can in Greek, where they spell out different features depending on their grammatical function (for a discussion see Roussou & Tsimpli 2006). Moreover, empirical data from research examining production and interpretation of pronominal pre- and postverbal subjects (Tsimpli et al. 2004) confirm differences between Greek and Italian.

When it comes to reference, previous psycholinguistic studies in Greek suggest that the null and overt options are ambiguous, although they exhibit specific preferences with respect to their choice of antecedent. Thus, there is a strong bias for the null pronoun in the Greek example in (3) to co-refer with the discourse salient subject antecedent (e.g., o astinomikós = ‘the.NOM policeman.NOM’) to maintain topichood, rather than the object (e.g., ton kléfti = ‘the.ACC thief.ACC’) (e.g., Tsimpli et al. 2004; Papadopoulou et al. 2015).

In contrast, the interpretation of the unstressed overt anaphoric expression aftós in (4) is associated with a preference for a less prominent referent, i.e., the object antecedent, marking topic-shift (e.g., Dimitriadis 1996; Miltsakaki 2002; 2007; Papadopoulou et al. 2015). This preference has been found in different types of behavioural tasks (e.g., self-paced reading or listening) for the pronouns aftós (‘he’, ‘this one’) and ekínos (‘that one’) but no experimental research has examined the anaphoric preferences of o ídhios (‘the same’/‘himself’) as yet. Noticeably, these overt subject expressions also have other functions: aftós is similar to the demonstrative ‘this (one)’ and ekínos, more marked, is used as the demonstrative ‘that (one)’ indicating a proximal/distal difference in reference.⁸ O ídhios has been suggested to be a pronominal anaphor (see Varlokosta & Hornstein 1993 and Everaert & Anagnostopoulou 2013 for relevant discussion), which behaves similarly to ‘himself’ with respect to emphatic uses.⁹

Dimitriadis (1996) used corpus-data to support that specific reference routines are inherent in the lexical entries of pronouns (and pro), according to which, overt pronouns (e.g. aftós/ekínos) cannot be construed with the most salient entities. Miltsakaki (2002; 2007), using corpus-data and sentence-completion tasks, suggested also that overt pronouns are preferentially anchored to the less salient antecedents. She claimed that this choice is led by the semantic relations between the predicates of the matrix and subordinate clauses, while order of mention of the referents in dependent clauses (intra- and inter-sententially) does not affect the choice.

Research investigating the processing load of backward anaphora and its resolution (Kaltsa et al. 2015; Papadopoulou et al. 2015), followed Tsimpli et al. (2004), creating a self-paced listening with sentence-picture matching grammaticality judgment task. Participants were presented with a picture and paced themselves through listening to a segmented (single) sentence, comprised of a main SVO clause and a subordinate one, introduced by an overt (aftós) or null pronoun. At the end of each sentence participants had to judge whether the picture-probe matched (or not) the sentence they had just heard. The pictures depicted the events mentioned in the subordinate clauses in three alternations: the action was performed either by the subject or the object of the matrix clause they had listened to, or by a third new entity (as in Tsimpli et al. 2004). Papadopoulou et al. (2015) tested monolingual children and adult speakers of Greek and found that the overt pronoun elicited strong preference for object antecedents across age groups, while only adults preferred to link the null pronoun to the most prominent/salient antecedent, i.e., the topic, contra Dimitriadis (1996) and Tsimpli et al. (2004). Offering an explanation to this finding, they suggested that the null option is unmarked in the grammar, allowing interpretation to either antecedent, while topic bias is actually the outcome of its contrast to the overt pronoun, also available. Moreover, as already mentioned, children differed from adults in allowing attachment to both subject and object antecedents in the null condition. Kaltsa et al. (2015) compared young and older adult monolingual speakers of Greek to heritage speakers and attriters in a Swedish-speaking environment. They found that, while monolinguals clearly preferred an object interpretation of the overt pronoun, the bilingual groups exhibited ambiguity, suggesting that the overt pronoun is vulnerable at the level of interface interpretation. They also spotted divergence in the preferences of younger compared to older participants in that the latter were more flexible in accepting either reference, while the former were shown to be more rigid in their choices.

The only eye-tracking research in Greek so far (Cunnings, Fotiadou & Tsimpli 2017) proposed that the topic-shift preference of overt pronouns in Greek may not be absolute. The study examined pronoun processing and resolution of L2 English among Greek speaking young adults: In their visual-world paradigm (following Arnold et al. 2000) they manipulated a purely grammatical cue, namely the gender of the subject and object competitors, creating ambiguous and unambiguous sentences (when different gender was used). Participants were simultaneously exposed to pictures depicting the two actors next to an object mentioned in the sentence (and a filler picture). Thus, the ambiguity of the aftós pronoun (when subject and object antecedents matched in gender) was resolved through visual cues. The authors investigated whether participants transfer L1 Greek to L2 English preferences. Relevant to the present research, they found that the L1 and L2 English speakers exhibited a subject antecedent preference during the pronoun time window. In other words, despite the fact that overt subject pronouns in Greek index a topic-shift, there was a general preference for the subject antecedent (i.e., discourse topic) to guide AR in English L2. On the other hand, in a second experiment investigating AR in Greek, the L1 Greek speakers’ eye-movements demonstrated no clear preference during processing of the overt pronoun, a finding that suggests that processing of overt pronouns in Greek does not necessarily trigger a topic-shift interpretation. In the comprehension questions, the participants exhibited a subject preference for the ambiguous pronoun in the subject bias ambiguous condition and an object preference in the object bias ambiguous condition, corroborating previous proposals. Interestingly, the attested object antecedent bias of the overt pronoun aftós found in previous judgement and self-paced reading or listening tasks (e.g., Tsimpli et al. 2004; Papadopoulou et al. 2015) may have been affected by the disrupted prosody imposed by the task methodology itself. Alternatively, the object antecedent bias may tap into processes that are not as early as the eye-movement data obtained during listening to continuous language stimuli. It is therefore likely that initial processing of the pronoun is not linked to a strong object or non-topic preference but instead it is genuinely ambiguous.

Finally, turning to properties of the subject expression o ídhios (‘the same’) which lacks any experimental research evidence, Varlokosta & Hornstein (1993) suggested that o ídhios qualifies as a regular pronoun, similar to aftós, when occurring in subject position. i.e., it does not require a sentence internal antecedent. The expressions differ in that o ídhios is necessarily focused, while aftós may but need not be focused. We assume that the lexical meaning of o ídhios (literally ‘the same’) carries an inherent focus feature which adds to its saliency and which in turn is responsible for imposing a preference for an antecedent in a parallel (grammatical subject) role. When in non-subject position, o ídhios has different properties and shows a mixed behavior with respect to an anaphor/bound variable status: It functions as an anaphor when non-emphatically used, or as an emphatic element modifying a discourse topic, in which case it does not need an antecedent in the sentence (cf. Iatridou 1986; Anagnostopoulou & Everaert 2013).¹⁰

Overall, when anaphora resolution does not involve categorical judgment but grammatical preference instead, different interpretive effects of null and overt subject pronouns are found in Greek (Table 1). As in a number of null-subject languages, null pronouns are the default option, co-referential with the previously mentioned subject, topic, or most prominent antecedent, although prior evidence from Greek is not very strong (cf. Dimitriadis 1996). On the other hand, overt subject pronouns (like aftós) constitute the ‘marked’ option indicating topic-shift and are thus often construed as co-referential with the syntactic object. However, previous studies (e.g., Tsimpli et al. 2004; Kaltsa et al. 2015; Papadopoulou et al. 2015; Cunnings et al. 2017) reveal vulnerability of the overt pronoun and suggest that topic-shift properties of overt pronouns may not be absolute in the language. In the absence of empirical evidence, the subject expression o ídhios is theoretically expected to behave similarly to aftós. Interestingly, behavioural and corpus studies so far (e.g., Miltsakaki 2007) have revealed no effects of word-order intra- or extra-sententially on speakers’ interpretive preferences of aftós.

2.1 Hypotheses

A syntactic implication arising from the default status of null subjects in NSLs is that overt pronominal subjects are associated with the interpretable features of topic-shift and focus (e.g. Tsimpli et al. 2004). Accordingly, we expected that participants’ gaze data would reflect information relevant to these interpretable features. In line with the properties of the two anaphoric expressions discussed before, we predicted that the subject picture would show longer fixation durations when adults listened to o ídhios compared to aftós (e.g., Varlokosta & Hornstein 1993). This subject preference should remain in the OVS structure as well, assuming that parallelism drives this expression’s processing and resolution. In essence both word-orders should reveal matching the subject grammatical role of o ídhios to the subject antecedent’s (Stevenson et al. 1995).

On the other hand, since the pronoun aftós has shown a preference for the less prominent, non-topic entity (e.g., Papadopoulou et al. 2015), we expected anaphoric type to interact with word-order in gaze data on the subject picture. The sensitivity of aftós to the discourse-related feature of topic-shift would trigger more ambiguity than o ídhios, particularly in the OVS structure where structural subject prominence competes with the topicalized object. Thus, we predicted that aftós should be more ambiguous following the OVS vs. the SVO structure.

Turning to predictions for the object picture, we expected that longer fixation durations should be manifested when adults listened to the pronoun aftós coming from the SVO compared to the OVS structure (e.g., Papadopoulou et al. 2015), showing a rapid sensitivity to topic-shift in Greek language (e.g., Tsimpli & Sorace 2006). The more frequent but ambiguous nature of the pronoun (cf. Cunnings et al. 2017) would interact with SVO and OVS word-orders used for the introduction of potential antecedents thus leading participants to invest more processing resources (longer fixation durations) to try to resolve the conflict between antecedents.

Age differences were also expected during online processing of these overt anaphoric expressions. Importantly, the ambiguous nature of the pronoun aftós and/or the more complex OVS structure were expected to be sensitive to age differences in both subject and object picture gaze duration, where young adults would show longer fixations compared to the elderly, indicating more processing when linguistic circumstances demand high cognitive resources. Age differences were not expected in the subject picture when adults listened to the pronoun o ídhios, especially when its potential referents were presented in the unmarked SVO order.

Overall, the online processing effects predicted during pronoun processing (fixation durations) were expected to pattern similarly with the participants’ response preference when answering the comprehension question. That is, a general preference for the subject antecedent was expected when adults listened to the pronoun o ídhios whereas the pronoun aftós should trigger a preference for the object antecedent. However, we also predicted the presentation of the OVS structure would reduce these differences especially for the pronoun aftós, suggesting the interpretation of the pronoun is challenging with the OVS word-order. Finally, because no time constraint was imposed on the comprehension question, age differences were not expected to play a role in response preference.

2.3 Procedure

Two individual sessions were held for each participant. Consent forms were signed at the beginning of these sessions. The first session started with the educational background questionnaire and the Mini Mental State Examination (Fountoulakis et al. 2000), which together took no longer than 15 minutes. After a five minutes’ break, we also assessed the backward digit span task, taking less than 10 minutes to complete. In total, the first session lasted about 30 minutes.

Two to three days later, the second session took place. Participants were presented with the online AR task which took approximately 40 minutes. Participants were seated in a comfortable chair at approximately 60 cm from the eye tracking monitor attached to a DELL laptop and advised to look on the computer screen throughout the experimental procedure to avoid any sudden movements. Light conditions in the environment were carefully controlled. Eye-movements were recorded with a Tobii X2-60¹² eye-tracker upon a 15.6″ screen (adapted to 4:3 format, 800 × 600 resolution). Auditory and visual stimuli were presented using the E-Prime software (Schneider, Eschman & Zuccolotto 2002), while fixation duration and response preference were recorded with the E-Prime extensions for Tobii software. A nine point calibration was followed by the instructions of the task. Three practice items ensured participants understood the whole procedure. They were instructed to listen to the short narratives while viewing the pictures, and subsequently answer the comprehension question by clicking with the mouse on the image illustrating their choice. A fixation cross surrounded by a red square during fixation (see Thaler, Schuütz, Goodale & Gegenfurtner 2013) was presented in the center of the screen for 750 ms before each trial. Then, the fixation cross disappeared and participants were simultaneously exposed to the auditory narrative and the pictures. Both experimental and filler items were randomized, to avoid any error variability due to familiarity and/or fatigue.

2.4 Data analysis

Different statistical analyses were conducted for the two measures of the online AR task. First, Linear Mixed Effects (LME) model analyses were carried out on the sum fixation duration (in milliseconds) for a time window of 1200 ms which began at the onset of the critical anaphoric expression, i.e., the beginning of sentence (3) and lasted until the onset of the object of the anaphoric clause, before the definite article introducing it, using the lmer function of the lme4 R package (Bates, Maechler, Bolker & Walker 2015). Overall fixation duration was preferred over division in bins to minimise the impact of the one-syllable difference between the duration of the two anaphoric expressions examined. Second, because response preference to the question for each picture was coded as a binomial measure (preferred vs. not preferred), we carried out a Mixed-Effect Logistic Regression (MELR) model analysis by using the glmer function of the lme4 R package. Both LME and MELR models are very powerful statistical analyses that account for both random and fixed effects. All models included Participants as the random factor, and Anaphoric type (aftós vs. o ídhios), Word-order (SVO vs. OVS), and Age (young vs. the elderly) as the fixed factors. Items were not included as a random factor because there were not enough data points (i.e., fixations) for each participant and item in the four types of pictures. The maximum fixed structure was always composed by a three-way interaction (i.e., anaphoric type × word-order × age). To control for working memory differences between young and elderly adults, we included the total score of WMC as a random slope of Participants factor. A minimum and maximum duration of 80–1000 ms. was established for fixation data (Manor & Gordon 2003). Outlier fixation duration data per anaphoric type, word-order and age group was detected by the Box-Whisker plot and replaced with the mean (3.46%).

The best fixed structure was found using stepwise model comparison from the most complex (the three-way interaction) to the simplest (a main effect) model, and selecting the one with lower AIC and BIC, and significant χ² test for the Log-likelihood, using the maximum likelihood. The p value of each significant fixed effect was calculated by the Anova function of the car package (Fox et al. 2018), and in case post-hoc comparisons were necessary we used the testInteractions function of the phia R package (De Rosario-Martínez 2013) with Bonferroni correction. Effect sizes for LME were informed by the explained deviance (dv) extracted by the pamer.fnc function of the LMERConvenienceFunctions R package (Tremblay & Ransijn 2015). This statistic serves as a generalization of R² because it measures the marginal improvement or reduction in unexplained variability in the fixed component after accounting for a given predictor. Effect size for MELR was reported using odds ratio (OR) and the 95% confidence interval (CI), by the lsmeans function of the lsmeans R package (Lenth 2017). OR indicates the constant effect of a specific predictor on the likelihood that one outcome will occur whereas CI estimates the precision of the OR (i.e., small CI signals high precision).

3.1 Fixation duration

To understand whether age differences in the processing of the ambiguous pronoun appeared only in the subject and/or object picture but not in the distractor picture, we performed a LME with age (young vs. elderly) and picture (subject, object and distractor) on fixation duration sum. This model showed a significant main effect of age, F(1) = 10.30, p < .01, dv = .88, with longer fixation durations in young (M = 603, SE = 15) than in elderly (M = 536, SE = 16) adults. In addition, the two-way interaction between age and picture was also significant F(2) = 9.28, p < .001, dv = .79 (see Figure 2), where post-hoc comparisons revealed longer fixation durations in young compared to elderly adults in the subject, χ²(1) = 8.65, p < .001, and the object picture, χ²(1) = 17.49, p < .001, but not in the distractor picture, χ²(1) = 2.25, p = 0.40.

Figure 2

Fixation duration sum (in milliseconds) in the subject and the object pictures of the online AR task, divided by age group and picture.

The fact that age differences were found just in the two (subject and object) antecedents discards the possibility that the distractor could be playing a role when explaining individual differences in AR processing. Thus, critical analyses were carried out only in the subject and object pictures (see Table 3 for means and standard errors).

Overall, descriptive measures show that the elderly preferred to link o ídhios subject expression to the subject antecedent when its possible referents were introduced in SVO utterances (no preference attested for the OVS word-orders). Young adults looked more to the object (than the subject) antecedent when o ídhios’ referents were introduced in SVO utterances and preferred the subject (over the object) antecedent when referents were introduced in OVS utterances. On the other hand, a subject preference is attested in all instances of aftós subject pronoun, irrespective of the word-order used for the introduction of its possible referents, in both age groups. Next, we present analyses performed for each antecedent picture.

3.1.1 Subject antecedent picture

To explore whether AR processing was explained by the type of overt anaphoric expression, word-order and/or age, we ran a LME with anaphoric type (aftós vs. o ídhios), word-order (SVO vs. OVS), and age group (young vs. elderly) on the fixation duration sum found in the subject picture. Similar to the previous result, the model manifested a significant main effect of age, F(1) = 7.43, p < .01, dv = 1.72, with longer fixation durations in young (M = 624, SE = 25) than in elderly (M = 516, SE = 27) adults. More importantly, the three-way interaction of anaphoric type, word-order and age was also significant, F(2) = 4.28, p < .05, dv = 1.05. No other effect reached significance (all ps > .05).

To follow-up on the three-way interaction, we divided the analysis by type of anaphoric expression (see Figure 3). The analysis performed for o ídhios trials demonstrated a significant interaction between word-order and age, F(1) = 7.88, p < .01, where the two age groups did not differ in SVO, χ²(1) = 0.70, p = .81, but they did in OVS, χ²(1) = 10.10, p < .01, where young adults showed longer fixation durations than the older ones. In contrast, the interaction of word-order and age was not significant in the pronoun aftós, F(1) = 0.03, p = .86, where compared to the elderly, young adults had longer fixation durations in both SVO and OVS structures. In other words, young adults spent significantly more time than elderly adults on the subject picture when o ídhios was preceded by a OVS word-order with a pre-posed object than an SVO word-order. Similarly, young adults looked at the subject picture more than elderly adults when they encountered the pronoun aftós in both word-order structures.

Figure 3

Fixation duration sum (in milliseconds) in the subject picture of the online AR task, divided by anaphoric type, word-order and age group.

3.1.2 Object antecedent picture

A third LME with anaphoric type (aftós vs. o ídhios), word-order (SVO vs. OVS), and age group (young vs. the elderly) was performed on the fixation duration sum of the object picture. The model showed a significant main effect of age, F(1) = 16.41, p < .001, dv = 1.72, where, once more, young adults had longer fixation durations than the elderly (M = 613, SE = 26 and M = 461, SE = 27, respectively; see Figure 4). No other effect was reached (all ps > .05). This effect suggests young adults looked at the object antecedent more than the elderly, regardless of the type of anaphoric expression or word-order used.

Figure 4

Fixation duration sum (in milliseconds) in the object picture of the online AR task, divided by anaphoric type, word-order and age group.

3.2 Response preference

To understand whether participants had a preference for the subject and/or object antecedents compared to the distractor character when resolving the ambiguity of the anaphoric expression, we examined the proportion of response preference (from 0 to 1) for each picture. Both age groups had a relatively high preference proportion for the subject (0.58 and 0.56 for young and the elderly, respectively) and the object antecedent (0.40 and 0.41), whereas proportion was very small for the distractor picture (0.02 and 0.03). Accordingly, we carried out a MELR analysis with word-order (SVO vs. OVS), anaphoric type (aftós vs. o ídhios) and age (young vs. elderly) including only preference proportion for the subject or object antecedents (see Table 4 for means and standard errors).

This model showed a significant main effect of anaphoric type, χ²(1) = 23.54, p < .001, with more preference for the subject in o ídhios (M = 0.66, SE = 0.10) than in aftós (M = 0.52, SE = 0.10; OR = 0.62, CI –0.86, –0.37) and a significant two-way interaction of anaphoric type and word-order, χ²(1) = 9.89, p < .01 (see Figure 5). No other effects reached significance (all ps > .05). Post-hoc comparisons in the two-way interaction showed all participants preferred the subject antecedent with SVO compared to OVS when they listened to o ídhios, χ²(1) = 12.25, p < .001 (OR = 0.62, CI –1.08, –0.17); however, no differences between word-order structures were found with aftós, χ²(1) = 0.81, p = .74 (OR = 0.15, CI –0.28, –0.59). Thus, when the SVO word-order was presented, adults of younger or older age had a preference for the subject antecedent when listening to the subject expression o ídhios. Fronting the object in the preceding clause boosted the expression’s less preferred object readings. However, no preferences were found with the more ambiguous pronoun aftós.

Figure 5

Response preference (from 0 to 1) for the subject antecedent (reverse pattern for the object) in the online AR task, divided by anaphoric type and word-order.

4.1 Online anaphora processing: Fixation durations

We first discuss our findings for the online processing of the subject picture and then proceed with a discussion of the fixations on the object picture.

According to our hypotheses, we predicted the online processing of the subject picture to show longer fixation durations when adults listened to o ídhios (than aftós) irrespective of the word-order used in the clause introducing the possible referents, due to this expression’s lexical meaning (‘the same’) (e.g., Varlokosta & Hornstein 1993) and the associated parallelism preference. In contrast, given that the pronoun aftós is suggested to be sensitive to discourse manipulation of topichood (i.e., topic-shift), we expected it would be preferentially linked to the less prominent discourse entity (in line with Papadopoulou et al. 2015 a.o.). However, we expected the online anaphora processing to show an interaction between anaphoric type and word-order. Thus, aftós was expected to be more prone to ambiguity in cases where a pre-verbal object was used, since in OVS structures the structural subject and the topicalized object compete for prominence. Age was hypothesized to play an important role during processing of the subject expressions, specifically for the pronoun aftós. In this case, we predicted age effects due to (non-structural) ambiguity requiring more cognitive resources for reference interpretation, i.e., the manipulation of linguistic (morpho-syntactic) and extra-linguistic (pragmatic) factors in real time.

Analyses of our gaze data revealed no significant effect of the anaphoric type or the word-order variables, nor of their interaction, suggesting that these linguistic factors alone could not explain online anaphora processing. Nonetheless, anaphoric type and word-order interacted with age. Although we suggested that age differences would be significant for the pronoun aftós, gaze data revealed differences in the subject looks for the expression o ídhios as well. These differences came from OVS structures: Young adults showed longer fixations than the elderly in the subject picture; no such differences were evinced when the expression’s competing referents were presented in an SVO structure (see Figure 3). In other words, when competing referents were presented in a structure marking topic-shift, elderly adults’ looks to subject antecedent pictures were significantly less than the younger ones’; young adults did not show this word-order effect. This finding suggests that whereas both groups have preserved ability to process o ídhios (‘the same’/‘himself’) when preceded by SVO sentences, young adults outperform older ones when the anaphoric expression is preceded by an OVS sentence, in the sense that their looks to the subject picture do not drop as is the case with the elderly participants.

Turning to data from aftós, we found that, on hearing the pronoun, young adults spent more time than the elderly looking at the subject picture, irrespective of the word-order used to introduce the referents. Although we did not have clear predictions for the pronoun aftós, these results can be explained by the fact that the more frequent but ambiguous nature of aftós caused a conflict between the two types of antecedents (subject and object) requiring more processing time, thus more time spent looking on both potential referents. Young adults were better detecting the conflict than elderly adults by increasing the processing time spent on the subject picture.¹³

Overall, fixation durations in the subject picture suggest that, compared to aftós, the online processing of the pronoun o ídhios is preserved in elderly adults when structural complexity is lower (i.e., in SVO). This is because o ídhios is a less ambiguous pronoun that looks for an antecedent in a parallel (subject) position, due to the focus feature it triggers, as assumed by its lexical meaning.

Results found in the subject picture are also consistent with those studies interpreting lower performance of older adults in AR as a result of cognitive decline in WMC (e.g., Fraundorf et al. 2012; Hendriks et al. 2014), which has been also associated with higher sensitivity to ambiguity (e.g., Dwivedi & Golhawk 2009), and syntactic complexity (e.g., Stine-Morrow et al. 2000). Indeed, Stine-Morrow et al. (2000) found that both younger and older adults performed similarly in subject relative clauses, whereas older adults had problems comprehending object relative clauses. However, in our study, WMC was assessed by an independent task (i.e., the backward digit span task) and the total score of each participant was included as a random slope in all statistical models to control for individual variability possibly due to WMC. Although it is possible that we cannot capture all aspects of WMC with one task, our results suggest that other aspects of cognitive decline (e.g., speed of lexical access) could explain why the elderly experience difficulties during the online processing of o ídhios after OVS (but not SVO) sentences, and aftós after both SVO and OVS sentences. Overall, however, it seems that elderly adults have fewer processing resources compared to young adults as evinced by the lower overall fixation results in all conditions of the study.

In the object picture, we expected longer fixation durations when participants listened to the pronoun aftós coming from the SVO compared to the OVS structure. However, due to the more ambiguous nature of aftós, we expected differences between OVS and SVO structures to be attenuated, showing the need to invest more processing in the latter (than the former) condition, due to the conflict between the pronoun’s topic-shift feature (favouring object antecedents) and the SVO word-order (pointing to a subject reference). The only significant effect here was age group, where young adults manifested longer fixation durations than the elderly in all conditions, demonstrating a better processing ability.

Although our data cannot clarify whether the lack of a difference between the OVS and SVO structures with the pronoun aftós was only due to ambiguity (as the same pattern also appeared in the pronoun o ídhios), it seems our online AR task imposed a high cognitive demand. Thus, discrepancies in the preferred reading of aftós (‘he’) found in our study, when compared to previous studies may be explained in terms of methodology. Differently from previous studies investigating AR by using self-paced comprehension (e.g., Papadopoulou et al. 2015), our task was presented through continuous speech decreasing the time available for participants to process story content. In addition, instead of assessing AR by providing only one sentence, as in the previous eye-movements study using the visual-world paradigm (Cunnings et al. 2017), we presented stories which were three-sentences-long. Finally, our task involved inter-sentential reference, while prior research focused on intra-sentential reference. Importantly, visual stimuli in previous research depicted events, while we only included pictures of referents. The lack of depicted events meant that moderating processing load or disambiguating the anaphoric expression was not possible using visual cues. The presence vs. absence of contextual information prior to the introduction of the ambiguous pronominal and the differences in methodology (self-paced reading or listening vs. the visual world paradigm) require further research. Still, these differences could have contributed to the lower fixation durations during online anaphora processing in the elderly group of adults.

Interestingly, the pattern of results observed in the object picture contrast with the pattern found in the subject picture. In the subject picture, elderly adults manifested sensitivity to variation in relation to anaphoric type and word-order: they clearly looked at the subject in the o ídhios-SVO while their looks in the o ídhios-OVS was less than in all other conditions (i.e., aftós-SVO and aftós-OVS). This differential pattern is easier to understand when we think of it in terms of subject prominence. The subject is by default the most salient entity of the matrix clause, frequently in the first position of a sentence and, therefore, it keeps a grammatical parallel position with that of the pronoun (see Stevenson et al. 1995). In addition, the subject is usually the topic of the sentence and, in Greek it is clearly morphologically marked by nominative, offering many cues to its prominence and interpretation. Finally, our online AR task included only verbs that were thematically neutral between the two antecedents (e.g., ‘to meet’), which means that there were no semantic cues that could facilitate either subject or object prominence.

Altogether, our findings demonstrate that variation in the type of anaphoric expression and syntactic complexity affects elderly adults’ online anaphora processing when related to the subject (but not the object) antecedent.

4.2 Offline anaphora interpretation: Response preference

Our hypotheses for anaphora interpretation in response to the comprehension question were expected to confirm the effects observed in online anaphora processing. More concretely, we predicted a preference for the subject picture when adults listened to the pronoun o ídhios (similar to the antecedent reference suggested by the theoretical analysis of Varlokosta & Hornstein 1993), but a preference for the object when listening to the pronoun aftós (in line with previous research results for Greek, as in Papadopoulou et al. 2015). In addition, we expected an interaction between anaphor type and word-order, where the presentation of the OVS structure would affect especially the pattern suggested for aftós, a pronoun previously suggested being ambiguous in nature (e.g., Cunnings et al. 2017). Age differences were not expected to play a role in response preference, as both young and elderly adults were allowed to take as much time as they needed to answer the comprehension question and in line with previous literature which suggests that comprehension is not affected by age (e.g., Hendriks et al. 2015).

Indeed, according to our results and in line with Varlokosta and Hornstein’s (1993) theoretical proposal, the main effect of anaphor type confirmed adults had a preference for the subject picture in the pronoun o ídhios, which had not been empirically demonstrated as yet. However, in contrast with previous research demonstrating the pronoun aftós showing a preference for the object antecedent (e.g., Papadopoulou et al. 2015), preference rates for aftós were at chance level (about 0.5; see Figure 5), which illustrates the ambiguous nature of this pronoun. This result corroborates the idea that the pronoun aftós in Greek is open to either antecedent (subject and object) and does not necessarily mark topic-shift inherently (Cunnings et al. 2017).

Previous research on the relation between reference and word-order of dependent clauses is scarce. However, first mention and recency that have been examined as factors enhancing referent accessibility (e.g., Arnold et al. 2000) are relevant to our discussion. Our results revealed an interaction between anaphor and word-order in response preference. For the subject expression o ídhios, our (offline) response results suggest that although there is a clear preference for the subject antecedent referent, OVS structures boosted the less preferred object reading. This finding indicates that the interpretation of the pronoun o ídhios was challenging when a syntactically more complex word-order was presented. The finding possibly relates to the idea that the prominence of the element in clause-initial position and the advantage of the first-mentioned referent (the subject in SVO and the object in OVS) reduce preference for the subject antecedent. Interestingly, the preference for the subject picture after listening to the pronoun o ídhios coming from SVO sentences is consistent with our fixation duration results on the subject antecedent pictures, where longer fixations were registered for both young and elderly adults (compared to fixations coming from OVS sentences).

In order to investigate whether anaphora processing may in fact predict the subsequent offline interpretation across adulthood, we ran a correlation analysis between fixation duration and offline responses (as suggested by one of the reviewers). The analysis gave no significant results for either group (all ps > .01) probably because both the online and offline measurements were distributed close to their means (fixation durations for the two potential antecedents and final interpretations of the subject expressions; see Tables 3 & 4 respectively).¹⁴

On the other hand, no differences due to word-order used were observed in aftós. The pronoun was equally linked to both subject and object antecedent pictures, irrespective of word-order. Thus, locality/recency of specific antecedents did not provide direct resolution of the ambiguous pronoun, traditionally suggested to mark topic-shift. Instead, no change in the preferences of the participants was noted with SVO vs. OVS manipulation. Noticeably, responses in our task illustrated preferences (not categorical judgments) as the ambiguity was never resolved, but interpretation was based in individuals’ sensitivity to discourse related cues, world knowledge and processing capacity. Recall that the complexity of the task may partly explain this discrepancy between our data and data from previous research. Alternatively, our results may further support the claim that aftós is open to either interpretation and does not necessarily mark topic-shift (in line with Cunnings et al. 2017).

However, the results of the two subject expressions suggest that online anaphora processing may predict the subsequent offline anaphora interpretation in adults, at least in the easy to process and most frequent SVO word-order structures.

Finally, as expected, no age effect was found in response preferences a finding in line with previous literature arguing that language comprehension does not decline due to aging (e.g., Light & Capps 1986; Titone et al. 2000; Hendriks et al. 2014).