Attrition as bias strengthening: revisiting previous findings from interface phenomena

2.1 A working definition of attrition

Within the present paper, we take the term attrition to refer to any changes within an individual speaker’s L1 due to the acquisition and use of an L2 after the L1 had been acquired.^[1] Therefore, attrition is distinct from language change due to pathology or language loss/change at the community level. Additionally, we restrict ourselves to adult attrition. In this sense, attrition is typically studied in migration contexts in which the potential attriter (i.e., the person undergoing attrition) has moved from their L1 community to their L2 community where they are likely to receive substantially less L1 input as well as substantially more L2 input.

An aspect of this definition that is worth highlighting is that we do not presuppose that attrition is the result of transfer from the L2. That is to say, we will not exclude the possibility that attrition may be the result of cognitive effects related to becoming bilingual. Although we take this assumption to be shared by other researchers (e.g., Chamorro et al. 2016a; Martín-Villena 2023; Sorace 2011), we believe that explicitly stating this assumption allows us to identify novel and potentially interesting patterns, as will become relevant below.

2.2 Pronominal biases in Italian

A fruitful domain for attrition research has been the syntactic optionality offered by Italian-style null-subject languages. In these languages, null subjects are licensed by agreement (Rizzi 1982, 1986) meaning that the subject of a well-formed finite clause may be phonologically overt or null (1a). This contrasts with non-null-subject languages like English (1b).

Despite this syntactic optionality, null and overt pronouns in Italian are constrained by discourse factors such as topicality and focus, with overt pronouns representing the ‘marked’ option (Cardinaletti and Starke 1999). When the referent of the subject represents a topic continuity, native Italian speakers prefer the production of null subjects. Conversely when the subject represents a shift in topic, native speakers disprefer null pronouns in favour of overt pronouns or non-pronominal DPs (e.g., Di Domenico et al. 2020; Sorace et al. 2009).

This asymmetry extends to comprehension, under the assumption that matrix subjects are generally treated as default sentential topics (Reinhart 1981). For example, Carminati (2002) found that native Italian speakers overwhelmingly interpreted null pronouns in sentences like (2) as referring to the matrix subject (80 %) and overt pronouns as referring to a non-subject antecedent (83 %). Subsequent studies with globally ambiguous items have supported this finding (e.g., Vogelzang et al. 2020: null 86 % subject, overt 61 % non-subject; Contemori and Di Domenico 2021: null 75 % subject, overt 80 % non-subject).

Convergent results have also been reported by work on the online processing of pronouns in Italian. Using pragmatically disambiguated items like (3), Carminati (2002) found that native Italian speakers read the second clause slower when null pronouns were biased to corefer with the object of the preceding clause than when they were biased to corefer with its subject. The opposite pattern was observed with overt pronouns. These findings have since been replicated (Filiaci et al. 2014).

2.3 Attrition in null-subject languages

To investigate how the interpretive biases for (intrasentential) pronominal resolution are affected by attrition, Tsimpli et al. (2004) conducted a picture selection task. For this, the authors recruited L1-Greek and L1-Italian speakers who had achieved ‘near-native’ proficiency in their L2-English and had been immersed in their L2 for a minimum of six years. They additionally recruited a control group for each language. Items consisted of biclausal sentences in which the ambiguous pronominal form always appeared in the embedded clause. In half of their items, the matrix clause preceded the embedded clause (4a). In the other half, the order was reversed (4b).

Sentences were presented along with three pictures corresponding to the logically possible interpretations of the pronoun: coreferential with the matrix subject, the matrix object, or some unspecified new referent. Participants were invited to select any and all images that corresponded to the given item.

When presented with overt pronouns, the control participants were expected to prefer matrix-object or new-referent interpretations (i.e., topic-shift, [TS], readings). When presented with a null pronoun, they were expected to prefer matrix-subject interpretations (i.e., non-TS readings). For the experimental groups, the authors predicted that English would influence the interpretable discourse features on overt pronouns in Italian (i.e., representational changes to the L1), but not the availability of null subjects. Specifically, they predicted a bleaching of the [+TS] featural specification for those pronouns, resulting in indeterminacy in their interpretation. For this task, the authors only presented the Italian data.^[2]

For embedded-first items (4b) with null pronouns, the authors reported a significant preference for matrix-subject readings in both L1-Italian groups, with no difference between the groups. For embedded-first items with overt pronouns, the authors reported an interaction driven by a significantly stronger preference for new-referent readings in the control group than in the experimental group. No significant differences were reported for the rates of matrix-subject or matrix-object readings.

For matrix-first items (4a), significant interactions were observed for items with both null and overt pronouns. Within null items, this was driven by the fact that the experimental group opted for significantly more matrix-subject readings and significantly fewer matrix-object readings than the control group. For matrix-first items with overt pronouns, the interaction was driven by the experimental group selecting significantly more matrix-subject readings than the control group. No differences were reported between matrix-object or new-referent readings.

Tsimpli et al. (2004) interpret these results to support their hypothesis that attrition only affects interpretable features for two reasons. First, they observed an attrition effect in the interpretation of overt pronouns. Second, no effect of attrition was observed in the embedded-first items with null pronouns. However, there are some inconsistencies with their account. First, the attrition effect for overt pronouns in the embedded-first items (4b) is not capturable in terms of a change in the [+TS] specification of overt pronouns. That trend was driven by fewer new-referent interpretations by the experimental group who instead selected numerically – but not significantly – more matrix-object and matrix-subject interpretations. Second, the increase in matrix-subject interpretations for overt pronouns by the experimental group in the matrix-first items (4a) was modest (control: 7 %; experimental 21 %) and did not result in a loss of their TS bias. This is not what would be expected if the feature had become underspecified. Third, the increased non-TS bias observed in the matrix-first items (4a) with null pronouns cannot be captured under their account. Instead, the authors speculate that the experimental group may have misanalysed the embedded clause as being non-finite like (5), resulting in obligatory control by the matrix subject, a point to which we will return after considering data from other studies.

Tsimpli et al. (2004) also report the results from their Greek participants from a picture selection task investigating the interpretation of a related property of null-subject languages, the syntactic optionality of pre- and postverbal subjects. In that task, an initial sentence always introduced a set of possible referents (6). A second sentence always contained an indefinite subject DP that could be interpreted as old or new information. Half the items contained a preverbal subject (6a) and half a postverbal subject (6b). After each item, participants were presented images corresponding to the two logically possible readings (old information, new information) as well as an invalid option. Participants were invited to select any and all images that matched the sentence. When participants selected both valid options, this was not recorded as a new and an old response, but rather as a separate category ‘both’.

The control group was expected to prefer old readings for preverbal subjects and exhibit no clear bias for postverbal subjects. Under attrition, the experimental group was predicted to differ significantly in their interpretation of preverbal subjects only. In that case, they were expected to admit more new readings under the influence of English which requires a preverbal subject regardless of information status.

In the preverbal items, the experimental group selected significantly^[3] fewer new readings and significantly more ‘both’ readings relative to the control group. No difference was reported for old readings. For the postverbal items, both groups exhibited indeterminacy, but in different ways. For the control group new, old, and ‘both’ were selected at similar rates. Conversely, the experimental group exhibited a clear bias for ‘both’ interpretations and opted for new and old readings around a fifth of the time. In this condition, the two groups differed significantly in their rates of old and ‘both’ readings.

Again, Tsimpli et al. (2004) interpret their results to support their hypothesis that attrition affects interpretable features only. They base this conclusion on the increased rate of ‘both’ readings for preverbal subjects which they take to indicate increased indeterminacy in the interpretation of preverbal subjects with regard to the feature [±Topic]. However, we should also note that their post hocs indicated no difference in old readings while also indicating that the experimental group selected significantly fewer – not more – new (i.e., non-Topic) readings. Thus, it seems more reasonable to attribute the increase in ‘both’ responses to the decrease in pure non-Topic readings. To see this more clearly, consider what would have happened if the authors had only considered the overall response rates of new and old, rather than treating ‘both’ as a separate category. The control group would have shown a modest bias in line with the expected preference for old readings (new: 46 % of items; old: 66 %, approximately^[4]) whereas the experimental group would have shown a much more noticeable bias (new: 52 %; old: 86 %, approximately).

Applying the same logic to the results from postverbal items also affects our interpretation, albeit to a lesser extent. On the one hand, Tsimpli et al. (2004) interpret the experimental group’s (i) significantly higher rates of ‘both’ responses and (ii) lower rates of old responses to indicate increased indeterminacy. On the other, if we were to only consider the overall rates of new and old readings collapsing both, we see that the two groups pattern together and appear to show no bias (control: new: 57 %; old: 57 %; experimental: new: 72 %; old 70 % approximately^[5]). Thus, it seems that the two groups do not actually differ in their interpretive biases (both showing indeterminacy) in line with Tsimpli et al.’s (2004) predictions, although the experimental group is overall more permissive.

Integrating the Greek and Italian results, it seems that only one of Tsimpli et al.’s (2004) attrition effects (matrix-first items with overt pronouns in Italian) provides clear evidence that attrition surfaces as the relaxation of restrictions relating to discourse features. Conversely, if the reader is willing to accept our re-interpretations presented above, two of their attrition effects (matrix-first items with null pronouns in Italian and preverbal subjects in Greek) appear to indicate that attriters exhibited a more pronounced version of an interpretive bias already present in the L1. Finally, the remaining two effects do not appear to relate to (interpretable) discourse features.^[6] First, we should not attribute the decrease in other responses for embedded-first items with overt pronouns in Italian to a change in the [+TS] feature due to the lack of a significant difference in the rate of matrix-subject readings. Second, the increased acceptance rate for any interpretation of postverbal subjects in Greek cannot be attributed to a featural change.

A similarly mixed pattern can be observed in Gürel (2004). In that study, they were interested in how prolonged exposure to English might affect pronominal resolution in Turkish, another null-subject language (Şener and Takahashi 2010). In addition to null pronouns (ø), Turkish has two overt pronominal forms that may occupy the subject position: o (‘s/he’) and kendisi (‘self’). When these forms appear in contexts like in (7), o cannot be coreferential with the matrix subject. This contrasts with pronominal subjects in English (8) which admit either interpretation. The same asymmetry also does not hold for kendisi or ø.

For Turkish, Gürel (2004) predicted attrition to be selective. This is couched in terms of the Activation Threshold Hypothesis (ATH, M. Paradis 1993, 2007) which proposes that the less a given form is used, the higher its threshold for activation (i.e., it is less available) and vice versa. Under this proposal, attrition is predicted when a form in the L1 has a ‘competing’ form in the L2 with a lower activation threshold. Relevant to the present discussion, Gürel (2004) assumed that the overt pronoun o in Turkish competes with overt English pronouns such as he. Therefore, they predicted o would begin to admit matrix-subject readings in sentences like (7) under pressure from the L2. However, English does not admit null subjects, nor does it have a pronoun analogous to kendisi in subject position. Consequently, those forms were predicted to remain stable.

To test this idea, Gürel (2004) recruited a group of native Turkish speakers who had emigrated after the age of sixteen and had lived in English-speaking countries for a minimum of ten years. L2 proficiency was not assessed, but participants were assumed to be highly proficient in their L2-English given their professional and educational backgrounds. A control group still residing in Turkey was also recruited. This group did not consist of idealized monolinguals and were deemed competent enough in English to take part in the story task.^[7]

Participants were presented three tasks. In the written interpretation task, participants saw a series of sentences similar to (7). Within those items, the author manipulated the pronominal form (o, kendisi, ø) and whether the potential antecedent was referential or quantified (e.g., Burak vs everyone). After each sentence, participants were asked to indicate their interpretation of the embedded subject selecting from three options: coreferential with the matrix subject (e.g., Burak), disjoint reference (i.e., someone else), or both.

Similar items were used in the story-based and picture-based truth-value judgement tasks. However, in those tasks, participants were presented either a short story in English or a picture along with each item. This was done to provide a context in order to force particular interpretations of the pronominal elements. Against these story contexts, participants were asked to judge whether the following Turkish sentence was true or false.

Results from the written interpretation task indicated that both groups almost always selected the disjoint reading and almost never interpreted o as coreferential with the matrix subject. Statistical analyses indicated that the only difference between the two was that the experimental group selected subject interpretation marginally more frequently than the control group in quantified items. No difference was found for referential antecedents. Moreover, no statistically significant differences were found in the rates of ‘both’ (i.e., subject or disjoint) readings.

As for kendisi, the control group preferred ‘both’ interpretations and never selected disjoint readings. The experimental group similarly dispreferred disjoint readings. However, they rarely gave ‘both’ responses and instead strongly preferred subject interpretations. A similar, but less pronounced, shift away from ‘both’ and toward subject readings by the experimental group was also attested for null pronouns. For both pronominal types, these shifts toward more subject readings were significant.

For the story-based truth-value judgement task, both groups preferred disjoint reading for o and subject readings for kendisi and ø. The only significant difference between the groups surfaced with o. While the control group almost never accepted subject readings, the experimental group did so at a higher rate for both referential and quantified antecedents. Similarly, in the picture-based task, the control group categorically rejected bound readings whereas the experimental group accepted such readings in a subset of trials. However, in that task, significant differences were also observed within kendisi and null pronominal items. Although both groups preferred subject readings, this bias was stronger in the experimental group.

Gürel (2004) interprets their results to indicate three things. First, the distinction between the o and the other two pronouns is maintained under attrition. Second, attriters begin to treat o more like ‘he’ under pressure from English, consistent with the ATH. Third, they take the increased bias toward subject readings for kendisi and ø to indicate that the experimental group exhibited a reduced awareness of the ambiguity for those pronouns.

However, integrating the results from their two truth-value judgement tasks we might come to an alternative interpretation. In those tasks where participants were forced to accept or reject a particular reading, participants did not exhibit indeterminacy. Rather, all participants preferentially interpreted ø and kendisi as referring to the matrix subject (values ≥ 74 %). Therefore, it seems reasonable to suggest that Turkish speakers exhibit a matrix-subject bias in the interpretation of the relevant forms. Integrating this with their written task results, we might suggest that the experimental participants are not less aware of the ambiguity, but instead more sensitive to their L1’s matrix-subject bias in that context.

Integrating the Turkish findings with those from Tsimpli et al. (2004) then, there are two parallels worth highlighting. First, the results from Gürel’s (2004) two truth-value judgement tasks indicate that o in Turkish is susceptible to attrition similar to overt pronouns in Italian, with matrix-subject readings becoming more prevalent. Moreover, the effect in Turkish (story-based: referential: 4 % vs 30 %; quantified: 3 % vs 22 %; image-based: referential: 0 % vs 21 %) is similar in magnitude to the effect observed in matrix-first items with overt pronouns in Italian (7.6 % vs 21.2 %). Clearly, in neither case have the attriting participants lost the L1 bias, despite its relaxation. Second, under our re-interpretation, the results for ø and kendisi in Turkish are convergent with Tsimpli et al.’s (2004) results for the matrix-first null pronominal items in Italian and the preverbal items in Greek. In those cases, a bias already present in the L1 appears to have strengthened under attrition. Moreover, the results from Gürel’s (2004) interpretation task are replicable (Gürel and Yılmaz 2011), suggesting we should not consider them spurious.

A convergent pattern was also observed in two studies by Martín-Villena (2023). For a pilot study exploring the role of conjunctions on pronominal reference, Martín-Villena (2023) recruited a large sample of native (Iberian) Spanish speakers (N = 131) who had grown up monolingually and were still residing in Spain (Martín-Villena, p.c.). However, seventy-six considered themselves ‘highly proficient’ in their L2-English, while fifty-five considered themselves as ‘not [being] proficient enough in English’.

Their experimental method and items were adapted from the pronominal sentence interpretation task in Tsimpli et al. (2004). In this Spanish version, the matrix clause always preceded the embedded clause and the type of conjunction (mientras – ‘while’, cuando – ‘when’) was manipulated. Items were presented with images corresponding to subject, object, and other interpretations of the ambiguous pronoun. Participants were instructed to select their preferred interpretation.

Focusing on their most relevant results, Martín-Villena (2023) reported a significant L2-level by pronoun interaction which was driven by more object interpretations for the highly proficient L2ers in the overt conditions, i.e., participants who deemed themselves more proficient in their L2 exhibited a stronger – not weaker – TS bias in their interpretation of overt pronouns. This is unexpected under a feature-based account like that in Tsimpli et al. (2004) or the ATH but fits with the novel pattern noted in the Greek, Italian, and Turkish results above. Of course, one might reasonably object that even though these are ‘highly proficient’ L2ers, they were still residing in their L1 community and as such may not be representative of prototypical attriting populations in migration contexts. To that end, we turn to Martín-Villena’s (2023) main study.

For their main study, Martín-Villena (2023) conducted an almost identical picture selection task to explore potential attrition effects using three groups of native (Iberian) Spanish speakers who had grown up monolingually: a control group, an immersed experimental group, and an instructed experimental group. The control group consisted of ‘functionally monolingual’ speakers who had never lived abroad and had minimal contact with English. The immersed bilingual group consisted of a large number of participants (N = 94) who were living in the UK or Ireland and had emigrated after the age of 15 years. The instructed experimental group consisted of a large number of participants (N = 80) still residing in Spain who were studying for English Studies degrees and received non-trivial English input. Both experimental groups were assessed to be highly proficient in their L2-English.

For that main task, Martín-Villena (2023) based their predictions on both the IH and the ATH. Starting with the IH, they predicted a relaxation of the TS bias for overt pronouns. However, they follow a more recent instantiation which focuses on potential processing-based explanations (Sorace 2011, 2016) rather than representational change. This shift toward a processing-based account is based on two ideas: (i) that the felicitous use of pronominal forms requires rapid, real-time integration of contextual and linguistic information and (ii) that bilinguals may be less efficient in this process (potentially due to cognitive load related to language control or a trade-off between increased inhibitory control and updating). Under this account, the overexertion of overt pronouns compensates for bilinguals’ difficulties with real-time integration as overt pronouns have been suggested to function as a processing default in languages like Italian (Sorace 2011). Following the ATH, Martín-Villena (2023) additionally predicted the attrition effect to be more pronounced in the immersed group relative to the instructed one.

Focusing on their most germane results, Martín-Villena (2023) observed significant group-by-pronoun interactions. Convergent with their pilot results, these interactions were driven by a significantly stronger TS bias for overt pronouns for both experimental groups relative to the control group, with no difference between the experimental groups.

Complicating the picture, however, Martín-Villena (2023) used the Bilingual Language Profile questionnaire (BLP, Birdsong et al. 2012) to assess participants’ language dominance. This indicated that all groups were still L1 dominant with the difference between the two languages most pronounced for the control group and least pronounced for the immersed group. Modelling revealed that in addition to the group-by-pronoun interaction, there was also a significant BLP score by pronoun interaction, surfacing as a weaker TS bias for overt pronouns in participants who were less L1 dominant. On the surface, this seems to converge with previous work that has reported a weaker TS bias for overt pronouns under attrition. However, this conflicts with the Martín-Villena’s (2023) own findings (pilot and main study). Martín-Villena interprets this conflict as evidence that we should move beyond dichotomous groups and instead investigate attrition with continuous measures such as dominance (2023: xxi, 292). However, the interaction with BLP score does not negate the interaction with group. Moreover, if dominance (as measured by an index like the BLP) is the factor that we should be interested in when exploring attrition, rather than the traditional factor of group, we might predict that the interaction with BLP score should survive in a simplified model which does not include group. This is not the case (Martín-Villena p.c.); the significance of the interaction with BLP score depends on the inclusion of the interaction with group, but not vice versa. That observation in turn raises the question of what relevant aspect of the participants the BLP is tapping.

More recent work on attrition and pronominal resolution in Italian, however, does not fit nicely into the emerging pattern noted in the studies discussed above. For their study, Gargiulo and van de Weijer (2020) recruited two groups of native Italian speakers: a control group in Italy and an experimental group in Sweden. Although that group had grown up in Italy, they had migrated to Sweden at least seven years prior to testing (mean = 11.83) and rated themselves as highly proficient (4.75/5) in Swedish, a non-null-subject language. Both groups completed a sentence interpretation task using items as in (10).

However, as Gargiulo and van de Weijer (2020) were interested in how re-immersion in a speaker’s L1 community affects attrition (on which see Chamorro et al. 2016b), they had experimental participants complete their task twice, once before their summer holidays in Italy (min 11 days, mean = 23.2) and a second time as soon as possible after their return to Sweden (mean = 2.95 days). For consistency, the control group also participated twice with a minimum of twenty days between test sessions (mean = 22 days).

Results indicated that both groups exhibited the expected interpretation biases for null and overt pronouns; they preferentially selected subject readings for null pronouns and non-subject readings for overt pronouns. However, the effect of pronoun interacted with session. The authors interpreted this to be driven by an increased non-TS bias for null pronouns in the second testing session. The effect of pronoun also interacted with group (i.e., an effect of attrition). Based on visual inspection, this was interpreted as being driven by a weaker non-TS bias for null pronouns in the experimental group contra Tsimpli et al. (2004) despite a similar trend in the data for overt pronouns. The authors suggest that the change in the interpretation of null pronouns is due to a locality effect (Table 1)

Thus, summarizing the above, while there is evidence that attrition can lead to the relaxation of the TS bias for overt pronouns (e.g., Italian: Tsimpli et al. 2004; Turkish: Gürel 2004; Greek: Kaltsa et al. 2015; Spanish: Martín-Villena 2023), there is also evidence that attrition can lead to a strengthening of the TS bias instead (Spanish: Martín-Villena 2023). Moreover, the latter of these two effects appears to be part of a more general pattern as strengthening of interpretive biases is observed in a wider range of contexts. This sometimes appears where there is no (surface) overlap between the L1 and the L2 (e.g., ø: Italian: Tsimpli et al. 2004; Turkish: Gürel 2004; Gürel and Yılmaz 2011; reflexives: Turkish: Gürel 2004; Gürel and Yılmaz 2011), and sometimes surfaces in instances of overlap (e.g., overt pronouns: Spanish: Martín-Villena 2023; preverbal subjects: Greek: Tsimpli et al. 2004). Given the different structures involved, it is clear we cannot maintain Tsimpli et al.’s (2004) suggestion that attriters sometimes misanalyse sentences with null pronouns as non-finite resulting in obligatory control. Nor can this pattern be captured by the ATH or a processing-based version of the IH that assumes overt pronouns function as a processing default. Rather we suggest that the increased reliance on interpretive biases might be a general (cognitive) effect under attrition. Nonetheless, a sceptical reader might object (i) that our interpretation of the data leads to a radically different view of attrition (ii) that it is based on a small number of studies, and (iii) that there are some potentially contradictory results for null pronouns in Italian (Gargiulo and van de Weijer 2020). Therefore, we might worry whether the effects are spurious. To that end, the present paper presents an experiment to further investigate how attrition affects the interpretation of pronouns – both null and overt – in Italian. For this, we drew the following research questions:

3.1 Participants

For this experiment, we recruited two groups of native Italian speakers who had grown up monolingually^[8] in Italy until at least the age of sixteen and reported no diagnosed language-related disorders.

The control group comprised thirty-three participants who were living in Italy. As part of a background questionnaire, participants were asked to list the languages they spoke at or above an ‘intermediate’ level. They were additionally asked to indicate how often they used each of these languages in a typical day. That revealed that these participants were not idealized monolinguals, with thirty reporting speaking some language other than Italian. Of those participants who spoke another language, all reported speaking English. To quantify language use in a typical day, we calculated percentages for each language. Table 2 presents group means and standard deviations for the use of Italian, English, and other. At the time of testing, the mean age of these participants was 40.91 years (SD = 6.40; range = 35–57)

The experimental group comprised twenty-nine participants who had been living in a majority English-speaking country for a minimum of six years (cf. Tsimpli et al. 2004) at the time of testing (mean = 14.27 years; SD = 7.89; range = 6.17–40.58).^[9] To avoid possible re-exposure effects (Chamorro et al. 2016b), experimental participants were pre-screened to ensure they had not travelled back to the L1 community in the three months prior to testing. Group means and standard deviations for the use of Italian, English, and other languages as percentages of a typical day are presented in Table 2. As some previous work that had restricted itself to ‘near-natives’ (e.g., Chamorro et al. 2016a, 2016b; Tsimpli et al. 2004), experimental participants took the Cambridge Assessment General English quick placement test. On average, experimental participants scored 21.76/25 (SD = 1.66; range = 18–24), suggesting that they are likely upper-intermediate/advanced L2 speakers. This group was comparable in age to the group still residing in Italy (mean age in years = 43.48; SD = 7.18; range = 36–66).

3.2 Stimuli

Critical items (N = 20) were adapted from Tsimpli et al.’s (2004) picture selection task. These sentences vary by pronoun (null or overt) and order (matrix-first^[10] vs embedded-first). As it was unclear whether the original items were fully counterbalanced, we adapted the items by counterbalancing the pronominal subject in each sentence. This was done to ensure that any effect of pronominal type in later modelling was not due to differences in the items used in the null and overt conditions. An example of the four conditions for the pronominal items is presented in (11). Sentences were distributed across two lists such that each participant saw each item once, five in each condition.

The experiment also included distractor sentences (N = 24) and filler items (N = 26), both of which were biclausal. Distractor sentences contained pseudorelative-relative clause parsing ambiguities and are reported in Cairncross et al. (2024). Filler items consisted of unambiguous co-ordinated sentences as in (12).

3.3 Procedure

Participants were recruited via Prolific Academic, and the study was implemented using PCIbex (Zehr and Schwarz 2018). After providing informed consent, participants completed a short language background questionnaire. Experimental participants then completed the English-placement test. Finally, all participants completed the sentence interpretation task during which they were presented sentences in isolation. After reading each sentence participants pressed the space bar, causing the sentence to be replaced by a comprehension question.^[11] For critical items, this always asked about the participants’ interpretation of the ambiguous pronoun.^[12] Below the comprehension question, three potential answers were always written, one for each of the NPs mentioned in the item as well as the option qualcun altro (‘someone else’). Potential responses were labelled ‘F’, ‘J’, and ‘B’. While the positions of the two textually given possible responses were counterbalanced such that the NP which was mentioned first appeared as the ‘F’ option in half of the trials (and as the ’J’ option in the other half). Qualcun altro (‘someone else’) always appeared as the ‘B’ option. To submit their preferred interpretation, participants used their keyboards. The study lasted approximately 30 minutes and participants were paid £4.50. Prior to testing, ethical approval was obtained from the Ethics Committee of the Faculty of Modern and Medieval Languages and Linguistics at the University of Cambridge.

3.4 Hypotheses and predictions

The following hypotheses were drawn in relation to RQ1 and RQ2 respectively:

Therefore, within the matrix-first items we expected a main effect of group as well as a main effect of pronoun. Specifically, based on the findings in Tsimpli et al. (2004) (see Section 2.2), we expected the control group to select subject interpretations around chance level for null pronouns while the experimental group would display a non-TS bias. For overt pronouns in the same items, both groups were expected to exhibit a TS bias, although this is expected to be weaker in the experimental group.

3.5 Data cleaning and analysis

We coded as missing any trial for which the sentence reading time or the question response time were implausibly fast (implemented as < 1500 ms and < 500 ms respectively), affecting 0.72 % of the data.

Prior to data collection, we made two decisions about our planned analyses that are worth highlighting. First, we decided that the matrix-first items should be modelled separately from the embedded-first ones. This was motivated by the fact that our research questions are not about whether attrition significantly affects pronoun-antecedent biases in one of the orders more than the other. Rather, we are specifically interested in how the pronoun-antecedent biases are affected within the matrix-first items. Second, we decided to collapse the distinction between non-subject and other readings (original responses available on OSF). Although we included both as they are natural responses to the target questions and increased the comparability to Tsimpli et al. (2004), we are only theoretically interested in the TS/non-TS distinction.

For expositional ease, only the results from the matrix-first items (i.e., those items relevant for our research questions) are presented below. Results for the embedded-first items are available on the associated OSF page.

To analyse the data, responses were coded as ±subject and were entered into a mixed effect logistic regression using the lme4 package (Bates et al. 2015) in R (R Core Team 2022). Using an empty model, we selected the best fitting random-effects structure for our within-participant fixed effect (i.e., pronoun, Matuschek et al. 2017) using the Akaike Information Criterion. We then entered pronoun (negative level = null) and group (negative level = control) as fixed effects using sum coding.