How Focus and Position Affect the Interpretation of Demonstrative Pronouns

Focus, along with the notion of topic, is part of information structure, which describes how the information in a message is ‘packaged’ (Chafe, 1976; Krifka, 2008). Topical information is typically given in previous discourse, while focused constituents represent new information within the ongoing discourse. Thus, although topic and focus should not be equated with the given–new distinction (which represents an information structural notion on its own, see Krifka, 2008; Lambrecht, 1994), it suffices for the present purposes to assume that an entity that is in focus represents new (or newly important) information to the discourse. For example, in the question–answer pair in (1), the musician is understood to be in focus because it provides the answer to the question posed by speaker A, and the information conveyed by the expression “the musician” cannot be retrieved from the context.¹

Focus can be signalled linguistically in various different ways, for example prosodically (via pitch accent), syntactically (via clefts or scrambling to focus positions), morphologically via focus markers, via particular focus constructions, or via focus particles, although there is no one-to-one mapping between focus and its expression (Féry, 2008; see also Zimmermann & Onea, 2011 for a cross-linguistic review of the expression of focus). Two realizations of focus that are important for the current study are questions and clefts. In questions such as (1), the presence of an explicit question (posed by speaker A) triggers a focus–background partition in speaker B’s response (Féry, 2008; Frey, 2004), with the background being the proposition already expressed by A, and the focus being the person denoted by the expression “the musician”. In so-called cleft sentences such as (2), the meaning expressed is equivalent to “The artist bequeathed the ring to the musician”, but placing “the musician” in the it-cleft typically marks it as being in focus in some languages, including English and German (Lambrecht, 2001).²

Another characteristic of focus is that focused information often occurs sentence-finally. The relationship between focus and position is unclear, but focused elements often end up at the right edge of the phrase, perhaps for independent reasons (Féry, 2008). In German ditransitives (relevant for the current study), patients (the direct object) must appear in final position when they are in focus, while recipients (the indirect object) are free to appear in initial, medial or final position when in focus (Frey, 2015). Final position is not an explicit signal of focus in the same way as e.g. clefting or pitch accent, but in the absence of more explicit signals, final position could be considered a default focus position.

Almor (1999) used items such as those shown in (3) to demonstrate that noun phrases (NPs) referring to clefted referents elicited faster reading times than those referring to non-clefted referents.

In both (3a) and (3b) the NP anaphor “the bird” refers to the robin, but in (3a) the robin is the focus of an it-cleft, whereas in (3b) the apple is the focus of the wh-cleft, so the robin is not in focus. Using a similar design, Foraker and McElree (2007) tested the impact of clefts on reading times for a subsequent pronoun, see (4):

In both cases, the pronoun “she” in the second sentence refers to “the cheerful waitress”, which in (4a) is in cleft focus, while in (4b) “the decaffeinated coffee” is the cleft focus. The pronoun is read faster in (4a), when the referent is in cleft focus compared to (4b) when the referent of the pronoun is not in cleft focus.

Several studies have compared the effect of focus on pronoun resolution with the effect of topicality. This is an interesting comparison because topic and focus have different functions in information structure. Nonetheless, most studies have found that both these factors affect pronoun resolution in a similar way. For example, Cowles et al. (2007) used auditory stimuli in a cross-modal priming task to assess the effect of topicality and focus marking (cleft and pitch accent in conjunction) on naming times during the presentation of a pronoun, and found that topicality and focus both shortened naming times. Kaiser (2011a) carried out a visual-world eye tracking study to look at the impact of topic and contrastive focus while controlling for effects of grammatical role. Both offline interpretations and online looking behaviour were influenced by all three factors, and Kaiser agrees with the conclusion in Cowles et al. (2007) that both topicality and focus enhance the prominence of a referent, making pronoun resolution more felicitous.³ The effect on pronoun resolution can be seen in terms of a memory effect (e.g. Foraker & McElree, 2007), but is more often couched in terms of discourse prominence – both topicality and focus mark the importance of the entity, making it a better candidate referent for pronoun resolution.⁴

A recent set of studies sheds some further light on why focused entities have an advantage in pronoun resolution. Interestingly, they do so by investigating the opposite effect, the so-called “anti-focus effect” in pronoun resolution (Colonna et al., 2012, 2015; de la Fuente, 2015; de la Fuente & Hemforth, 2013; Patterson et al., 2017; Patterson & Felser, 2020). Focused entities have a clear advantage when the pronoun appears in the following sentence. However, when the pronoun appears in the same sentence as the focused entity, as in (5), there is no such advantage, and the likelihood of pronoun resolution to the focused entity even decreases in some cases.⁵ In (5b), for example, Peter is the focus of the it-cleft, and the pronoun he is less likely to be interpreted as Peter than in (5a) when there is no explicit focus marking on Peter.

This contrast in the role of focus between within- and between-sentence pronoun resolution highlights the fact that it is not focus per se that enhances pronoun resolution. Colonna et al. (2015) propose instead that focus signals an upcoming shift of topic.⁷ For example, when Peter is placed in a cleft in (5b), the addressee can infer that Peter is likely to be talked about further in the next sentence, i.e. Peter is put in focus in the first sentence to signal that he will most probably be the topic of the following sentence. On this account, it is really the topic status of the referent in the next sentence that gives rise to the focus advantage in pronoun resolution. This proposal explains why previous studies have found similar effects for both topicality and focus on pronoun resolution; the focus effect is in fact a (shifted) topicality effect. This proposal raises an interesting question with respect to demonstrative pronouns, as we will see below.

It is yet to be tested whether focus affects the resolution of demonstrative pronouns to the same extent as personal pronouns. German has both personal pronouns and a rich set of demonstrative pronouns (such as der, dieser, jener, derjenige) that can be used to refer to animate referents (including people). The demonstrative forms can very roughly be translated into English as “this one” or “that one”, and tend to refer to a non-topical or less expected referent.⁸ For instance, the second sentence in (6a) starts with a personal pronoun er which is the equivalent of English “he”, and in this context is quite likely to refer to the musician. In (6b) the second sentence starts with the demonstrative dieser, which is much more likely to refer to the artist than the musician.

Several factors have been claimed to affect the resolution preferences for demonstratives in German. Demonstratives have been characterized as having an anti-topic and/or anti-subject preference (Bosch et al., 2003, 2007; Bosch & Hinterwimmer, 2016; Bosch & Umbach, 2007; Wilson, 2009), an anti-agent preference (Schumacher et al., 2015, 2016, 2017), and as referring to referents with lower discourse prominence (Bader & Portele, 2019; Kaiser, 2011b). The demonstrative dieser⁹ is less flexible and has stronger preferences than the personal pronoun, but is nonetheless somewhat sensitive to the next-mention bias (Patterson et al., 2022). When more than two potential referents are available, a gradient pattern emerges, with reference to agents/first-mentioned entities the least felicitous and reference to entities in final position the most felicitous; thus dieser has a strong preference for referents in final position (Patterson & Schumacher, 2021).

It has also been claimed that demonstrative pronouns have a topic-shift function (Abraham, 2002; Fuchs & Schumacher, 2020; Schumacher et al., 2015; Weinrich, 1993). The role of focus in determining resolution preferences for demonstratives has not been tested, but Patterson and Schumacher (2021) considered focus as a possible explanation for the strong preference for resolving demonstratives to referents in final position. They tested the acceptability of resolving personal pronouns and the demonstratives der and dieser to the agent, recipient and patient in ditransitive contexts; see (7).

In the first experiment the final position of the context sentence contained the patient (die Parkfläche “the parking space”), see (7a); in a second experiment the final position of the context sentence contained the recipient (dem Anwohner “the resident”), see (7b). In both experiments dieser received the best scores when its referent was in final position. The authors speculated that the association of final position with focus may have contributed to this preference. However, it was not possible to draw stronger conclusions about the interaction of focus and position on demonstrative pronouns because focus was not explicitly manipulated.

Several interesting questions arise from considering the role of focus in demonstrative pronoun resolution. If focus makes a referent more prominent (as claimed in e.g. Cowles et al., 2007; Kaiser, 2011a), are demonstratives more or less likely to refer to that referent, since they prefer referents with lower discourse prominence? Furthermore, if, as is claimed by Colonna et al. (2015), focus signals an upcoming referential shift, how will this interact with the topic-shift function of the demonstrative pronoun? Secondly, is the strong final position preference of the demonstrative dieser solely attributable to focus?

A related question is how focus influences pronoun resolution in demonstrative (and personal) pronouns. This is investigated by testing the influence of focus on the production likelihood and the next mention bias, as outlined below.

The Bayesian model for pronoun resolution (Kehler et al., 2008; Kehler & Rohde, 2013) proposes that pronoun interpretation is derived from two likelihoods which combine in a Bayesian formula as in (8):

Pronoun interpretation is represented on the left-hand side of the formula and is cast as the probability of a referent, given a pronoun. Repeating example (6a) as (9) below, this can be understood as the probability of the referent the musician, given the pronoun he, and similarly the probability of the referent the artist, given the pronoun he.

The likelihood P(pronoun|referent) is the addressee’s estimate of the likelihood of using a pronoun, given a particular referent. In (9), this is (the estimate of) the probability of producing a pronoun in order to refer to the musician, or conversely, (the estimate of) the probability of producing a pronoun in order to refer to the artist. We call this the pronoun production likelihood. Finally, the likelihood P(referent) represents what we call the next-mention bias. This is the addressee’s estimate of the likelihood of continuing the discourse by talking about a particular referent, using any referential expression (not just pronouns). In (9), at the point of the pronoun, it is (the estimate of) the likelihood of continuing the discourse about the musician, or similarly, continuing the discourse about the artist. The insight of the Bayesian model is that the ‘work’ the pronoun has to do is more limited than has been previously supposed, by separating out the next-mention bias from the production likelihood. In other words, some factors that were previously assumed to influence pronoun resolution directly can actually be attributed to the next-mention bias, which is not specific to pronouns.

The basic Bayesian model proposes simply that the two likelihoods, P(pronoun|referent) and P(referent) are combined in the Bayesian formula in (8) and can be used to estimate the likelihood of a particular pronoun interpretation. An important extension of the model is the proposal of the strong Bayesian model, which goes beyond the original formulation in claiming that the pronoun production likelihood and the next-mention bias are influenced by different factors (Kehler et al., 2008; Kehler & Rohde, 2013). The claim of the strong Bayesian model is that pronoun production likelihood is influenced by structural or information-structural factors such as subjecthood and/or topicality, while the next-mention bias is influenced by semantic factors (e.g. verb type and coherence relations). It should be noted that, while work by Kehler and colleagues shows a clear dissociation between influences on the next-mention bias and the production bias, this strong dissociation has been brought into question (Rosa & Arnold, 2017; Weatherford & Arnold, 2021; Zerkle & Arnold, 2019).

To our knowledge, the influence of focus on the production likelihood or next mention bias has not been tested. Focus poses an interesting question here because of competing claims about how it exerts influence on pronoun resolution. The claim of the strong Bayesian model is that information-structural factors should influence the production bias but not the next-mention bias. Since focus is clearly part of information structure, the prediction of the strong Bayesian model is that focus influences pronoun resolution via the production bias only. On the other hand, the claim from Colonna and colleagues (2015) is that focus influences pronoun resolution by indicating an upcoming change of topic which is usually realised in the next sentence. While not explicitly stated in terms of the Bayesian model, their description clearly fits with the next-mention bias, i.e. a focused referent in sentence n is likely to be mentioned (probably as a subject/topic) in sentence n+1. Our aim is therefore to test these two competing hypotheses about how focus influences pronoun resolution, comparing demonstrative and personal pronouns.

In the current study we carry out two story completion experiments. Story completion experiments have been used in studies testing the Bayesian model for pronoun resolution because the relevant biases for the model can be directly extracted from the data (Patterson et al., 2022; Rohde & Kehler, 2014; Zhan et al., 2020). Participants are presented with prompts consisting of two sentences, where the second sentence is truncated. Participants are asked to complete the second sentence. They see either a pronoun prompt (e.g. “The musician yelled at the artist. He…”), or a free-prompt in which no referential expression is presented (e.g. “The musician yelled at the artist. …”). Annotation of the pronoun prompt data gives the pronoun interpretations; annotation of the first referential expression in the free-prompt data gives the next-mention bias (i.e. which referent is talked about?), as well as the production bias (was a pronoun used to refer to a particular referent or not?).

The materials for both experiments are adapted from the ditransitive contexts from Patterson and Schumacher (2021). In Experiment 1, we test whether the demonstrative pronoun dieser is more likely to be interpreted as the recipient (compared to the patient) when the recipient is in focus via a cleft.¹⁰ We also test whether the degree of influence from focus status is similar in demonstrative and personal pronouns by comparing resolution of dieser and er. Furthermore, we investigate whether focus status influences pronoun resolution via the next-mention bias or the pronoun production likelihood, i.e. are these influences active on pronouns per se, or are they rather influences on the likelihood of further mentioning particular referents.

In Experiment 2 the issue of final position is investigated. As noted above, the association of final position with focus may contribute to the strong preference for interpreting dieser as referring to the referent in final position in the previous sentence. The question we ask here is, is the strong final position preference of the demonstrative dieser solely attributable to focus? We use question–answer pairs like (1), in which the recipient is always the target of the question so that it is always in focus in the answer. We then manipulate whether the recipient is presented in final or medial position in the answer. As such, the dependent variable is likelihood of recipient interpretations, and the independent variable is position (final vs. medial). If the final-position preference is solely attributable to focus, interpretation preferences for dieser should shift to the medially-presented referent when that referent is in focus. If, however, there is also an independent final-position preference, finally-presented recipients should have an advantage over medially-presented referents. We also test to what extent the influence of position is specific to dieser by comparing resolution of dieser and er in this respect. As in Experiment 1, we further test whether position influences pronoun resolution via the next-mention bias or the pronoun production likelihood.

Data from a pilot study of Experiment 2 (N = 12) was used to estimate power via simulation, using the mixedpower package (Kumle & Draschkow, 2020) in R (R Core Team, 2021), following guidance in Kumle et al. (2021). It was determined that >90% power to detect the five effects in Experiment 2 could be achieved by testing 100 participants (power tables S1-S8; R scripts for the power analysis and pilot data are available in Supplementary Materials). We expected the effects in Experiment 2 to be smaller than those in Experiment 1, because of the more equivocal effect of position (Exp. 2) as opposed to focus (Exp. 1). The power analysis therefore focused on Experiment 2, whose estimates we expected to be more conservative, and the same participant numbers were used also for Experiment 1. To avoid wasting resources on an overpowered experiment, however, planned interim statistical analyses were carried out at 50 participants and again at 75 participants for Experiment 1; the adjusted alpha level was 0.012. Data collection would have been stopped at an interim stage if the p-value for all theoretically meaningful effects were to fall below the adjusted alpha level.¹¹ This turned out not to be the case, however.

Power was estimated for each of the following effects:

• the effect of Position (medial; final) on the interpretation bias for demonstratives (model 1);

• the interaction of Position x Prompt (dieser versus er) in the interpretation biases for personal versus demonstrative pronouns (model 2);

• the effect of Position on the likelihood of mentioning the recipient (model 3);

• the effects of Position and Referent (recipient; patient) on the production bias for personal pronouns (model 4) and demonstratives (model 5).

For all the effects listed above, power > 0.90 could be achieved with 100 participants (see Tables S1-S8 in Supplementary Materials). For model 1 power > 0.90 could already be achieved with 40 participants (Table S1). Power simulations for models 1, 3 and 4 were based on the pilot data. While we acknowledge that effect sizes based on pilot data may be under- or overestimates, we see this as the best alternative for estimating power given the lack of directly comparable published data. It was not possible to use the pilot data for models 2 or 5 because of model non-covergence (model 2) and rank-deficiency (model 5).¹² Power simulation for model 2 was based on data from Patterson et al. (2022), Experiment 1.¹³ The interaction of interest was simulated using the command R2power from the mixedpower package (Kumle & Draschkow, 2020) in R (R Core Team, 2021). The simulated variable was participant, in order to inspect power at varying numbers of participants. The R2 variable items was held constant at 36, in order to reflect the number of items in the planned experiment. The R2 power simulation showed that, with 36 items, even at 40 participants power was already at 100% (Table S2). Because this simulation was based on published data, power for this model was simulated again, this time with a 20% reduction in the estimated beta values (to mimic a smaller effect size); this also resulted in 100% power at 40 participants (Table S3), showing that the power estimate was robust to some variation in the effect size.

For model 5 a simulated dataset was generated, using the pilot data to inform the data structure and the production bias model (Patterson et al., 2022, Experiment 1) for the model parameters.^14,15 Power was estimated to be >90% at 100 participants for the effects of both Position and Referent (Table S6). Reducing the beta values by 20% resulted in a drop of power to 79% (Position) and 89% (Referent); reducing by 30% dropped power to 69% (Position) and 77% (Referent) (Tables S7 and S8). This indicated some sensitivity to the size of the effect. For our planned experiments, we had enough power to detect an effect of a similar size to that in the demonstrative production bias in Patterson et al. (2022), but would not have had enough power to detect effects of a much smaller size. Effects for the demonstrative production bias were therefore interpreted with caution.

Experiment 1 was a story completion experiment, in which participants were presented with a context followed by a prompt. Their task was to continue the story by completing the prompt sentence. Responses were annotated to obtain the likelihood of referring to particular referents in the context sentence, and the likelihood of using a personal or demonstrative pronoun to do so.

In total data from 124 self-reported adult native German speakers was collected.^16,17 All participants gave informed consent before participating. Eighteen participants who reported being bi- or multilingual from a young age were excluded. A further two participants were excluded because they reported language or speech disorders. No participant took less than 5 minutes to complete the task, or gave more than 10 nonsensical, incomplete or excessively long completions. This left 104 viable participants for the analysis, who were aged between 19 and 70 years (M = 34, SD = 12), 55 male, 49 female.

Each of the 36 items consisted of one context sentence followed by a prompt sentence which was completed by the participants. The context sentence always started with a syntactic it-cleft as shown in (10). This placed the NP within the cleft in focus. The critical sentence always contained a ditransitive predicate with an agent as the subject, a dative object (the recipient) and an accusative object (the patient). It-clefts focused either the recipient or the patient. The pronoun prompts (er/dieser) were always masculine.¹⁸ The recipient and patient were always masculine (and were thus potential referents for the pronoun) and the agent feminine. The agent and recipient were always animate person characters, while the patient was always an inanimate object; this is the most common constellation of animate and inanimate arguments for German ditransitive verbs (Røreng, 2011). Accordingly, our experiment was confounded by animacy, since recipients were always animate and patients inanimate. As a reviewer points out, what if there is a form-specific preference for demonstratives to resolve to inanimate referents? Practical reasons for not manipulating animacy were: to avoid testing a large number of conditions (n=12) in one experiment; the low number of verbs available in which the patient can be animate; the fact that some ditransitive contexts with three animate referents tend to sound far less natural than those with inanimate patients. More importantly, however, we think that the issue under investigation (how focus affects the resolution of demonstratives) can be considered somewhat independent of the role of animacy. If there is a weak bias for resolving demonstratives to inanimates (or, indeed, to animates), this would affect the proportion of resolutions to the patient and recipient, but focus could still nonetheless shift those proportions - this is in fact what we were measuring. If the bias towards inanimates is so strong that focus does not shift it, we would not find an effect of focus. Our interpretation in the case of finding no effect of focus would be that we have no evidence regarding the effect of focus on demonstratives - while we may speculate about the reasons for this, it would require further investigation in new experiments which are outside the scope of the current paper. Finally, we take into consideration the findings in Patterson and Schumacher (2021), in which demonstratives with referents in final position were clearly judged to be the most felicitous, whether this was an (inanimate) patient or an (animate) recipient. Also, earlier experiments by Uzun (2015a, 2015b), in which demonstratives were tested in ditransitive contexts, showed a preference for the patient argument, whether or not this was animate. These findings are not indicative of a strong form-specific preference for resolving demonstratives to inanimate referents.

Moreover, care was taken to avoid patients that were event NPs (e.g. “the party”) or complex nouns that introduce an additional argument (e.g. das Katzenvideo “the cat video”). Care was also taken to ensure that the patient arguments in particular were sufficiently interesting that a story could be constructed about them; banal inanimate entities could introduce a bias towards writing continuations about the animate entities, i.e., the recipient in the pronoun prompt conditions and recipient or agent in the free prompt conditions. Within each item the factors Focus (recipient; patient) and Prompt (er; dieser; free prompt) were manipulated such that there were six versions of each experimental item, see (10). A simplified gloss has been provided for 10a as well as a translation. For the remaining conditions only a translation has been provided.

Fillers. 36 fillers were also presented in order to mask the design of the target experimental items. Fillers contained a mix of cleft (n=9), pseudo-cleft (n=3) and non-cleft (n=24) context sentences. A mix of transitive, ditransitive and stative verbs were used. The prompts were as follows: ten personal pronouns, ten demonstrative pronouns, ten free prompts, six adverbial or connector prompts. The demonstrative dieser can be used as a pronoun or as an adnominal modifier (e.g. dieser Mann, “this man”); since we were only interested in the pronominal use, excessive adnominal interpretations would have reduced the available dieser cases for analysis. For this reason, four of the demonstrative prompts in the first half of the experiment were followed directly by an auxiliary verb (e.g. Dieser ist…/Dieser hat…) in order to force a pronominal interpretation, so that participants were biased towards the pronominal reading of the demonstrative prompts in the experimental items.²⁰ Of the pronoun prompt fillers, half were ambiguous with respect to the referent of the pronoun. Ten feminine pronoun prompts were used in the fillers.

Experimental items were distributed over six lists in a Latin-square design and interspersed with the fillers. The six lists were hosted on the Qualtrics platform (Qualtrics, Provo, UT) in the form of questionnaires. Each questionnaire was divided into two parts (both of which were completed by the participant in the same session). Part 1 contained 12 experimental free-prompt items plus 12 fillers, and Part 2 contained 24 pronoun-prompt items and 24 fillers.²¹ This was to avoid the free-prompt continuations being influenced by a large number of pronoun prompts (H. Rohde and J. Hoek, personal communication, 8^th July 2021). Participants were invited to participate based on a short task description on the recruitment platform Prolific (prolific.co). Participants gave their consent to participate and were asked a short series of biographical questions. They were then given the task instructions. Participants were asked to complete every item by completing/supplying the next sentence, without making changes to the text presented. They were told that they should make the most obvious completion and not the most creative or humorous one, and that completions should be kept short and precise. On starting the questionnaire participants were randomly assigned to one of the lists. Participants received monetary compensation via the Prolific platform for taking part.

Data was coded by two student assistants. Both annotators made independent judgments about the responses. Missing, incomplete, ungrammatical and nonsensical continuations were excluded. All remaining responses (excluding fillers) were coded in the following way: the first referential expression in each continuation was coded for its intended referent (agent; recipient; patient; both; other; ambiguous). The first referential expression in the pronoun prompt conditions was the prompt itself (er or dieser). The first referential expression in the free-prompt conditions was identified by the annotators, and the type of referential expression was categorised as follows: personal pronoun (singular, including possessive pronouns); demonstrative pronoun (singular, including dieser, der, jener, derjenige); definite full noun phrase (singular); name; other. The two annotators agreed in 93% of observations, with a Cohen’s (unweighted) Kappa of .88 (z = 78.3, p < .001). Subsequently, observations where annotators disagreed were resolved via discussion in order to retain the maximum number of data points. For the data analysis, trials were retained if the first referential expression refers to the patient or recipient; all other responses were dropped. In the free prompt conditions it was possible that the referential expression referred to the agent in the context sentence (this would have been ungrammatical in the pronoun prompt conditions because the agent did not match in gender with the pronoun prompt). The number of references to the agent in the free prompt conditions was reported but these cases were not included in the analysis.

All data was analysed in R (R Core Team, 2021) using RStudio (RStudio Team, 2021) using generalized linear mixed-effects models (family = binomial) in the package lmerTest (Kuznetsova et al., 2017). For all models described below, fixed effects were sum-coded and random intercepts for participant and item were included (as long as model convergence was achieved).²² The appropriateness of including random slopes for the fixed effects was determined using the rePCA function in the package RePsychLing (Baayen et al., 2015).²³ We calculated five models to test for the following effects:

1. The effect of Focus (recipient; patient) on the rate of recipient interpretations of dieser, using the dieser-prompt data;

2. The interaction of Focus x Prompt (dieser; er) on the rate of recipient interpretations of er and dieser, using the dieser- and er-prompt data;

3. The effect of Focus on the rate of continuations about the recipient (the next-mention bias), using the free-prompt data;

4. The effect of Focus and Referent (recipient; patient) on the rate of personal pronoun use (er production bias), using the free-prompt data;

5. The effect of Focus and Referent on the rate of demonstrative pronoun use (dieser production bias), using the free-prompt data.

In Experiment 1, we investigated the following questions:

1. Whether the demonstrative pronoun dieser was more likely to be interpreted as the recipient (compared to the patient) when the recipient was in focus via a cleft;

2. Whether dieser differed from the personal pronoun er in the degree to which it was influenced by cleft focus;

3. Whether focus influenced pronoun resolution via the next-mention bias or the pronoun-production likelihood, separately for dieser and er.

A total of 3744 continuations were recorded. 166 continuations were excluded for being ungrammatical or nonsensical; 3 continuations were blank and counted as missing data; 18 continuations were excluded because the dieser-prompt was interpreted as a determiner instead of a pronoun and a further 13 were excluded because they contained no referential expression (total exclusion rate of 5.3%). The first referential expression in each of the remaining 3544 continuations was annotated (1210 dieser-prompt, 1223 er-prompt; 1111 free-prompt). The overall distribution of the referents in the three prompt conditions is shown in Table 1 and the distribution of referential expressions in the free-prompt condition is shown in Table 2.

Figure 1 shows the mean proportion of recipient interpretations, split by prompt and condition. Model outputs for Models 1-5 are summarised below. Full tabulated model outputs can be found in Supplementary Materials, Tables S9-S13. Effect size estimates were calculated using the effectsize package (Ben-Shachar et al., 2020, version 0.8.2) in R (R Core Team, 2021) using RStudio (RStudio Team, 2021). Note that for this experiment, the adjusted alpha level for testing significance is 0.012 (see Prior Power Analysis section for details).

Model 1. We analysed the effect of Focus (recipient; patient) on the rate of recipient interpretations of dieser, using the dieser-prompt data. The model showed that the recipient focus condition elicited significantly more recipient interpretations for dieser than the patient focus condition (ß = 1.626, z = 2.985, p = .003, d = 0.29 [0.10, 0.48]).

Model 2. We analysed the interaction between Focus and Prompt on the rate of recipient interpretations, using the dieser- and er-prompt data. The model showed that, while there were significant main effects of both Focus and Prompt, there was no evidence for a significant interaction between Focus and Prompt (ß = 0.136, z = 1.536, p = .124).

Model 3. We analysed the effect of Focus on the rate of recipient continuations (next-mention bias), using the free-prompt data. There was a significant effect of Focus; the recipient focus condition elicited significantly more recipient continuations than the patient focus condition (ß = 0.780, z = 3.733, p = .0002, d = 0.37 [0.18, 0.56]).

Models 4 and 5 tested the production likelihoods for personal and demonstrative pronouns respectively. Due to the nature of the dependent variable in the production likelihoods (proportion of personal or demonstrative pronouns produced), the crucial test for an effect of focus in Models 4 and 5 was in fact the interaction of Focus and Referent: i.e. testing whether the production likelihood of the chosen referent (recipient or patient) is modulated by Focus (recipient focus versus patient focus). A main effect of Focus without the interaction would indicate only that the production likelihoods were affected by thematic role of the focused referent. The interaction of Focus and Referent for Models 4 and 5 was mistakenly omitted from the Stage 1 manuscript but is reported below.

Model 4. We analysed the main effects of Focus and Referent and their interaction on the rate of personal pronoun use, using the free-prompt data. The mean proportions are plotted in Figure 2. There was no significant main effect of Focus (ß = -0.121, z = -0.701, p = .483). There was a significant main effect of Referent, with a higher likelihood of using a personal pronoun when the referent is a recipient compared to a patient (ß = 0.730, z = 4.152, p = 3.30x10^-5, d = 0.41 [0.22, 0.61]). There was a significant interaction of Focus and Referent (ß = 0.701, z = 3.997, p = 6.43x10^-5, d = 0.40 [0.20, 0.59]).

Exploratory analysis. The significant interaction of Referent and Focus in Model 4 was followed up pairwise by comparing the effect of Referent on each level of Focus separately. When the recipient was in focus, there was a significant effect of Referent such that the personal pronoun production bias was higher for recipient referents than for patient referents (ß = 1.573, z = 4.893, p = 9.92x10^-7, d = 0.49 [0.29, 0.69]). When the patient was in focus, there was no significant effect of Referent (ß = 0.033, z = 0.204, p = .839).

Model 5. We analysed the main effects of Focus and Referent and their interaction on the rate of demonstrative pronoun use, using the free-prompt data. The mean proportions are plotted in Figure 3. There was a significant main effect of Focus; the likelihood of using a demonstrative pronoun decreased in the recipient focus condition compared to the patient focus condition (ß = -0.543, z = -2.659, p = .008, d = -0.26 [-0.46, -0.07]). There was no significant effect of Referent (ß = 0.061, z = 0.355, p = .722) nor was there evidence for a significant interaction of Focus and Referent (ß = 0.042, z = 0.249, p = .803).

In Experiment 1 we investigated the effects of cleft focus on the interpretation of demonstrative pronouns. The overall proportion of recipient interpretations was very high compared to patient interpretations, for both the demonstrative and personal pronoun. Nonetheless, the interpretation preferences for the demonstrative dieser examined in Model 1 show a clear influence of cleft focus, such that the recipient focus condition (where the recipient is in cleft focus) significantly increased the likelihood of a recipient interpretation of the demonstrative pronoun relative to the patient interpretation. The interpretation of the personal pronoun er was also influenced by cleft focus in the same way, as can be seen in Figure 1. The output of Model 2 shows no evidence for a significant interaction between prompt (demonstrative versus personal pronoun) and focus. We therefore have no evidence that the interpretation of dieser and er are influenced to a different extent by cleft focus.

We then turned to the free-prompt data to investigate how cleft focus influences the interpretations: via the next-mention bias or via the production likelihood. Model 3 showed a clear effect of focus on the next-mention bias; that is, the likelihood of producing a referential expression, regardless of form, that refers to the recipient, is significantly higher when the recipient is in focus compared to when the patient is in focus. It is therefore likely that the interpretations of the demonstrative and personal pronouns are influenced by cleft focus via the next-mention bias. This could certainly be the case for the demonstrative pronouns, for which the production likelihood (i.e. the likelihood of producing a demonstrative pronoun) does not show evidence for a Focus by Referent interaction in Model 5. The production likelihood for demonstratives is nevertheless affected by whether the recipient or the patient is in focus, regardless of the referent of the demonstrative, with fewer demonstratives produced when the recipient is in focus. This suggests that demonstrative production is influenced by the thematic role of the focused entity. For personal pronouns, on the other hand, the production bias (i.e. the likelihood of producing a personal pronoun) is strongly affected by focus, as demonstrated by the significant Focus by Referent interaction in Model 4. When the recipient is in focus, personal pronouns are more likely to be produced when the referent is a recipient compared to a patient. When the patient is in focus, conversely, there is no evidence for a difference in production likelihoods between recipient and patient referents, as shown in the exploratory analysis. Overall, from Experiment 1 we can see that, while cleft focus affects the interpretation of both personal and demonstrative pronouns, it does so in different ways. For the demonstrative, the influence of focus is due to the next-mention bias rather than the demonstrative production bias. For the personal pronoun, however, interpretation is influenced by cleft focus via both the next-mention bias and the production bias. The implications of these findings are considered in the General Discussion.

Experiment 2 is a story-continuation experiment designed to test whether final position has an independent effect on pronoun resolution for demonstrative pronouns.

In total data from 114 self-reported adult native German speakers was collected.²⁴ Three participants who reported being bi- or multilingual from a young age were excluded. Two further participants were excluded: one participant reported language or speech disorders, and one participant took less than five minutes to complete the task. No participant gave more than 10 nonsensical, incomplete or excessively long completions. This left 109 viable participants for the analysis, who were aged between 18 and 72 years (M = 31, SD = 10), 57 male, 51 female, 1 genderqueer.

The 36 experimental items were adapted from Experiment 1. Each item consisted of two context sentences and a prompt sentence which was completed by the participants. The context sentences were in a question–answer dialogue format: the first context sentence is contributed by speaker A in the form of a focus question. Speaker B answers in a full sentence (the critical sentence) and continues with the prompt sentence. Properties of the prompts, and the gender match between prompts and the referents in the context sentence, were the same as in Experiment 1. The question posed by Speaker A always asked about the recipient, starting with wem “to/for whom”. This question triggers a focus–background partition in the critical sentence, placing the recipient in focus (Féry, 2008; Frey, 2004; Zimmermann & Onea, 2011).

Within each item the factors Position (final; medial) and Prompt (er; dieser; free prompt) were manipulated such that there were six versions of each experimental item, see (11). The factor Position represents the position of the focused entity (i.e. the recipient) in the critical sentence. A simplified gloss has been provided for 11a, as well as a translation. For the remaining conditions only a translation of B’s response has been provided (A’s question is the same in all conditions).

Fillers. 36 fillers (adapted from Experiment 1) were also presented in order to mask the design of the target experimental items. All fillers resembled the experimental items in that they started with a question from speaker A and a full sentence answer from speaker B, followed by the prompt sentence. The questions used a range of interrogatives (e.g. was “what”, warum “why”, wer/wen “who(m)”) and targeted a variety of arguments. A mix of transitive, ditransitive and stative verbs were used. The properties and distribution of the prompts were the same as in Experiment 1.

The procedure was the same as for Experiment 1.

As per Experiment 1.²⁵ The two annotators agreed in 96% of observations, with a Cohen’s (unweighted) Kappa of .94 (z = 75.9, p < .001).

Data analysis procedures were the same as in Experiment 1. Models were specified as for Experiment 1. We calculated five models to test for the following effects:

1. The effect of Position (medial; final) on the rate of recipient interpretations of dieser, using the dieser-prompt data;

2. The interaction of Position by Prompt (er; dieser) on the rate of recipient interpretations of er and dieser, using the dieser- and er-prompt data;

3. The effect of Position on the rate of continuations about the recipient (the next-mention bias), using the free-prompt data;

4. The effect of Position and Referent (recipient; patient) on the rate of personal pronoun use (er production bias), using the free-prompt data;

5. The effect of Position and Referent on the rate of demonstrative pronoun use (dieser production bias), using the free-prompt data.

In Experiment 2, we investigated the following questions:

1. Whether the demonstrative pronoun dieser is more likely to be interpreted as the recipient than the patient when the recipient is in final (versus medial) position;

2. Whether dieser differs from the personal pronoun er in the degree to which it is influenced by final position;

3. Whether position and referent type influences pronoun resolution via the next-mention bias or the pronoun-production likelihood, separately for dieser and er.

A total of 3924 continuations were recorded. 65 continuations were excluded for being ungrammatical or nonsensical; 14 continuations were excluded because the dieser-prompt was interpreted as a determiner instead of a pronoun; a further 20 were excluded because they contained no referential expression (total exclusion rate of 2.5%). The first referential expression in each of the remaining 3825 continuations was annotated (1287 dieser-prompt, 1303 er-prompt; 1235 free-prompt). The overall distribution of the referents in the three prompt conditions is shown in Table 3 and the distribution of referential expressions in the free-prompt condition is shown in Table 4.

Figure 4 shows the mean proportion of recipient interpretations, split by prompt and condition. Model outputs for Models 1-5 are summarised below. Full tabulated model outputs can be found in Supplementary Materials, Tables S14-S18. Effect size estimates were calculated using the effectsize package (Ben-Shachar et al., 2020, version 0.8.2) in R (R Core Team, 2021) using RStudio (RStudio Team, 2021).

Model 1. We analysed the effect of Position (final; medial) on the rate of recipient interpretations of dieser, using the dieser-prompt data. The model showed that the medial condition elicited significantly fewer recipient interpretations for dieser than the final condition (ß = -1.105, z = -2.577, p = .010, d = -0.25 [-0.43, -0.06]).

Model 2. We analysed the interaction between Position and Prompt on the rate of recipient interpretations, using the dieser- and er-prompt data. The model showed that, while there was a significant main effect of Position, there was no evidence for a significant interaction between Position and Prompt (ß = 0.100, z = 0.819, p = .413).

Model 3. We analysed the effect of Position on the rate of recipient continuations (next-mention bias), using the free-prompt data. There was only a marginally significant effect of Position; the medial condition elicited marginally fewer recipient continuations than the final condition (ß = -0.242, z = -1.690, p = .091, d = -0.16 [-0.35, 0.03]).

Models 4 and 5 test the production likelihoods for personal and demonstrative pronouns respectively. Due to the nature of the dependent variable in the production likelihoods (proportion of personal or demonstrative pronouns produced), the crucial test for an effect of position in Models 4 and 5 is in fact the interaction of Position and Referent: i.e. to test whether the pronoun production likelihood of the chosen referent (medial or final) is modulated by Position (focused referent in medial or final position). The interaction of Focus and Referent for Models 4 and 5 was mistakenly omitted from the Stage 1 manuscript but is reported below.

Model 4. We analysed the main effects of Position and Referent and their interaction on the rate of personal pronoun use, using the free-prompt data. The mean proportions are plotted in Figure 5. There was no significant effect of Position (ß =0.091, z = 0.494, p = .621), nor for Referent (ß = 0.219, z = 1.181, p = .238), nor was there evidence for a significant Position by Referent interaction (ß = 0.178, z = 0.908, p = .364).

Model 5. We analysed the main effects of Position and Referent and their interaction on the rate of demonstrative pronoun use, using the free-prompt data. The mean proportions are plotted in Figure 6. There was no significant effect of Position (ß = 0.119, z = 0.523, p = .601), but a significant effect of Referent (ß = -0.540, z = -2.449, p = .014, d = -0.23 [-0.42, -0.05]) and a significant interaction between Position and Referent (ß = 0.815, z = 3.620, p = .0003, d = 0.35 [0.16, 0.53]).

Exploratory analysis. The significant interaction of Position and Referent in Model 5 was followed up pairwise by comparing the effect of Referent on each level of Position separately. When the Position (of the focused recipient) was medial, there was no evidence for a difference in demonstrative production bias between medial and final referents (ß = 0.274, z = 0.986, p = .324). When the position (of the focused recipient) was final, the demonstrative production bias was significantly higher for final than for medial referents (ß = -6.037, z = -2.742, p = .006, d = -0.26 [-0.45, -0.07]).

In Experiment 2 we investigated the effect of final position on the interpretation of demonstrative pronouns. The overall proportion of recipient interpretations was very high compared to patient interpretations, for both the demonstrative and personal pronoun. There was a higher overall proportion of recipient continuations in Experiment 2 compared to Experiment 1, but this is not surprising considering that in Experiment 2, the recipient was always in focus. Controlling the focus in this way allowed us to explore whether position influenced pronoun interpretation independently of focus. The interpretation preferences for the demonstrative dieser examined in Model 1 show a clear influence of position, with a higher proportion of recipient interpretations when the recipient was in final position compared to medial position. Somewhat surprisingly, the interpretation of the personal pronoun er was also influenced by position in the same way, as can be seen in Figure 4. The output of Model 2 showed no evidence for a significant interaction between Prompt (demonstrative versus personal pronoun) and Position. We therefore have no evidence that the interpretation of dieser and er are influenced to a different extent by Position.

Looking at the free-prompt data allowed us to investigate how position influences the interpretations. Model 3 showed only a negligible effect of position on the next-mention bias; that is, the likelihood of producing a referential expression, regardless of form, that refers to the recipient, is only slightly (and not reliably) higher when the recipient is in final compared to medial position. The production bias for the personal pronoun was not affected by the position of the recipient (medial or final) nor by the position of the pronoun’s referent (medial or final). The demonstrative production bias, on the other hand, was affected by the position of the pronoun’s referent; a demonstrative was more likely to be produced when referring to referents in final position compared to medial position, and this effect was modulated by Position (i.e. the position of the focused recipient): the advantage in demonstrative production for referents in final position was only apparent when the position of the focused recipient was final. To restate this in a different way: demonstratives were more likely to be produced when the focused recipient was in final position and the demonstrative referred to the recipient compared to the patient; when the focused recipient was in medial position, demonstratives were equally likely to be produced for referring to the recipient or the patient. This pattern will be considered further in the General Discussion.

In this study we set out to investigate how focus and position affect the interpretation of demonstrative pronouns. The demonstrative dieser has previously been shown to have a strong preference for referents in final position of the previous clause or sentence involving ditransitive predicates (Patterson & Schumacher, 2021), and the association of focus with final position has been considered a possible explanation for this preference. This study is, to our knowledge, the first to explicitly test the effect of focus on German demonstrative pronouns. While the influence of focus on personal pronouns has been widely attested (Cowles et al., 2007; Foraker & McElree, 2007; Kaiser, 2011a, among others), it is not a matter of course that demonstratives will also be sensitive to focus. A second important and novel aspect of this study is that we investigated how focus and position may influence pronoun interpretation by evaluating their influence on the next-mention bias and the pronoun production bias for personal and demonstrative pronouns respectively, following the framework of the Bayesian model for pronoun interpretation (Kehler et al., 2008; Kehler & Rohde, 2013).

In Experiment 1, we investigated the influence of focus. We first tested whether focus influences the interpretation of demonstratives by manipulating whether the recipient or the patient was in focus and measuring the proportion of recipient interpretations. We used cleft constructions of the type “It was the musician to whom the artist bequeathed…” where the focused element was the first argument in the clause. We expressed some doubt as to whether demonstratives would be interpreted towards the focused referent, for two reasons. First, focus makes a referent more prominent (Cowles et al., 2007; Kaiser, 2011a), and demonstratives tend in general to refer to less prominent referents (e.g. Bader & Portele, 2019). Second, demonstratives have a topic-shift function (Abraham, 2002; Schumacher et al., 2015), and focus, too, has the function of indicating an upcoming change of topic (Colonna et al., 2015), so the two similar discourse functions of focus and demonstratives may conflict with each other. However, the results showed a clear effect of focus on the interpretation of the demonstrative dieser, with a higher proportion of recipient interpretations when the recipient was in focus. In fact, the sensitivity of demonstrative pronouns to cleft focus was indistinguishable from that of personal pronouns, which are well known to refer to focused referents.

It seems, then, that rather than being in conflict, the similar discourse functions of focus and demonstrative pronouns fit together well: the focus in the first sentence signals an upcoming change of topic to the focused referent, and the demonstrative in the second sentence can be interpreted as that focused referent at the point of the topic shift. But the finding that demonstratives refer felicitously to focused referents may at first seem difficult to reconcile to the fact that, in general, demonstratives refer to less prominent referents. Looking at how focus influences pronoun interpretation may shed light on this apparent contradiction.

First, there was a strong influence of focus on the next-mention bias; if a recipient was in focus in the first sentence, it was likely to be the referent of the first referential expression in the second sentence. This finding makes sense intuitively: if a referent is in focus, we expect to hear more about that referent in the following sentence. Our expectations about the upcoming referent in turn influence how we interpret pronouns if they are encountered, with interpretation towards the expected referent.

Second, the production bias for demonstrative pronouns was affected by our experimental manipulation of Focus, such that a demonstrative was less likely to be produced when the recipient was in focus compared to the patient. But we have to be careful in interpreting this effect. A Focus effect in the production bias, unlike in the interpretation and next-mention biases, does not tell us anything about the influence of focus per se²⁶: rather, it suggests that properties of the focused referent were moderating the production bias here. Crucially, there was no evidence for a Focus by Referent interaction in the demonstrative production bias, so no evidence that the rate of demonstrative production for a particular referent was modulated by focus. Instead, fewer demonstratives were produced when the referent with a higher thematic role (recipient) was in focus, compared to when the referent with a lower thematic role (patient) was in focus (following the thematic role hierarchy in Primus (1999))²⁷. The demonstrative production bias, then, seems to reflect the previous literature that associates demonstratives with a less prominent referent. However, because demonstrative interpretation is simultaneously influenced by the next-mention bias, which in this experiment was strongly biased towards focused referents, it was the next-mention bias (and not the production bias) that leads to the demonstrative interpretation towards focused referents.

A different picture emerges for the personal pronoun. The production bias for personal pronouns is influenced by focus in the following way: the likelihood of producing a personal pronoun for recipients compared to patients varies according to which referent is in focus. More personal pronouns are produced when referring to the recipient compared to the patient, but only when the recipient is in focus. For personal pronouns, then, referring to a referent with a higher thematic role leads to a boost in pronoun production, but only when that referent is in focus. Combined with the next-mention bias, which is also influenced by focus, this leads to a high interpretation bias towards focused referents for the personal pronouns.

A key insight from this experiment is that claims in previous literature which associate personal pronouns with more prominent and demonstratives with less prominent referents appear to relate to the production biases and not the next-mention bias. This makes sense because next-mention biases in terms of strong semantic or pragmatic biases may not have been present in previous studies, and may have instead been deliberately avoided in experimental settings. When the next-mention bias does not strongly point to one particular referent, the influence of the pronoun’s production bias will have a stronger effect on pronoun interpretation, according to the Bayesian model of pronoun interpretation. However, in our Experiment 1 the production biases are overlaid with a strong next-mention bias towards focused referents. Our findings then add a level of precision to existing claims about personal and demonstrative pronouns, not only about how focus affects pronoun resolution, but also about how referent prominence may affect pronoun interpretations.

Given that the next-mention bias was the same for personal and demonstrative pronouns, the differences in the production biases between personal and demonstrative pronouns should have led to slightly different interpretation biases. This was not statistically confirmed (Model 2, Experiment 1). It may be that ceiling effects in recipient interpretations prevented us from detecting possible differences between er and dieser here. A further caveat is that, according to our prior power analysis, the demonstrative production bias was somewhat sensitive to fluctuations in power. It is therefore possible that the demonstrative production bias would look different with a higher-powered experiment, but this is purely speculative.

In the Introduction we considered the competing claims about how focus influences pronoun interpretation. According to Colonna et al. (2015), focus indicates an upcoming change of topic which is realised in the next sentence. This claim can be cast in terms of the next-mention bias: a focused referent in sentence n is likely to be mentioned in sentence n+1. On the other hand, the strong Bayesian model for pronouns claims that information-structural factors (therefore including focus) influence pronoun interpretation via the production bias and not the next-mention bias. In Experiment 1, we found that focus not only influences the next-mention bias, in accordance with Colonna et al. (2015), but also the production bias. Our findings seem to be partially in keeping with both predictions. Colonna et al. make predictions about the next-mention bias, which are confirmed by Experiment 1, but do not say anything about production biases. The strong Bayesian model, on the other hand, claims a separation between semantic factors that influence the next-mention bias and structural factors (including information-structural factors) that influence the production bias (Rohde & Kehler, 2014), but the claims about information structural factors have only been tested for topichood and not focus. At least for the configurations tested here, the strong separation between structural and semantic factors is not confirmed. Focus affects not only the next-mention bias (unexpected according to the strong Bayesian model), but also the production bias for the personal pronoun.

So why should focus differ from topichood in this respect? Here we speculate about two possibilities. First, it might be that the particular configurations that have been tested give rise to different effects for focus and topic. In Experiment 1 we used cleft sentences, which bring the focused constituent into initial position in the sentence. This position may have conferred an extra degree of importance or a strong memory trace that increased the likelihood of mentioning that constituent in the next sentence. As far as topic is concerned, this was tested by Rohde and Kehler (2014) by manipulating passive versus active sentences, but a more explicit topic-marking device may have an effect on the next mention bias. In fact, Zhan et al (2020) could not entirely confirm the Rohde and Kehler findings with respect to topichood when they tested a passive construction and a secondary topic marker in Mandarin Chinese, and they suggest that in the end grammatical role might be more important for the production bias than topichood.

A second possibility is that focus and topic have different effects on the discourse structure and do indeed affect next-mention bias and production biases differently. If focus signals an upcoming topic shift as claimed by Colonna et al (2015), then this will certainly have an effect on which entity is mentioned next in the discourse. But the topic status of a discourse referent does not necessarily signal anything about the ensuing discourse. Topics do tend to stay constant over a stretch of discourse, but this is a tendency rather than an explicit signal. In addition, focus may confer more semantic content than topic does. One way to reconcile the current data with the strong Bayesian model would be to consider the semantic function of focus (choosing from a set of alternatives) as one of the semantic cues contributing to the next-mention bias. So it may be that the different characteristics of topic and focus lead to contrasting effects on the next-mention and production biases.

In either case, the claim from the strong version of the Bayesian model about information-structural factors affecting only the production biases needs to be re-examined, which would require further testing in a variety of structures and languages.

In Experiment 2 we controlled focus and looked at how position affects the interpretation of demonstrative pronouns. We asked whether the attested strong final-position preference for dieser interpretation was independent of focus, given the association between focus and final position in otherwise unmarked structures. We investigated this by keeping the recipient in focus in all conditions and manipulating the position in which the recipient appeared, between medial and final. The prediction was that if position is in any way independent of focus, there should be an effect of position on recipient interpretations.

The results of Experiment 2 showed an effect of position for demonstrative interpretations, such that the proportion of recipient interpretations increased when the recipient was in the final position in the previous sentence. This already indicates that position operates to some extent independently of focus as far as dieser is concerned: the recipient was always in focus, and there were more recipient interpretations when the recipient was in the final position than in the medial position. So for dieser, referents in final position are more accessible than referents in medial position, independent of focus status.

The interpretation of personal pronouns was also affected by final position. This finding was unexpected because in general, personal pronouns tend to exhibit a preference for referents in initial position (which was ruled out in Experiment 2 because the first NP always mismatched the pronoun in gender) and for German ditransitive structures, the personal pronoun er does not display strong preferences but is instead rather flexible in its interpretation in comparison to demonstratives (Patterson & Schumacher, 2021). Thus it is surprising that er interpretations would tend towards referents in final position as much as demonstratives. As in Experiment 1, it is possible that ceiling effects in recipient interpretations prevented us from detecting possible differences between er and dieser here.

The next-mention bias is not strongly influenced by position (only a marginal effect was found). The production bias for personal pronouns was also not affected by position. This leaves us with a puzzle for the personal pronouns: when neither the next-mention nor the production biases are strongly affected by position, where is the final-position bias in pronoun interpretation coming from? We do not have an explanation for this finding.

The demonstrative production bias, on the other hand, was affected by position in the following way: demonstratives are more likely to be produced when the focused recipient is in final position and the demonstrative refers to the recipient, rather than the patient; when the focused recipient is in medial position, demonstratives are equally likely to be produced for referring to the recipient or the patient. One way to unpack this rather complicated finding is as follows: recall that in Experiment 2 (unlike Experiment 1), the recipient was always in focus, and the factor Position denotes the position of the (in-focus) recipient. Demonstrative production is boosted when the referent is in final position, and when the referent is in focus (i.e. the recipient). Both factors point in the same direction in the Position = final condition, but point in opposite directions in the Position = medial condition, as illustrated in Table 5.

Let us compare this finding with the demonstrative production bias in Experiment 1. In that case, focus did not influence the production of demonstratives, rather, the properties of the focused referent modulated the production bias, with fewer demonstratives produced for focused recipients compared to focused patients. Why should focus boost the production bias in Experiment 2, but not in Experiment 1? It seems that the production bias (in Experiment 2) is affected when two factors are at play simultaneously: the combination of final position and focus gives a boost to the demonstrative production bias, but each factor on its own is not enough to modulate the production bias.

In sum, our experiments have shown that the interpretation of the demonstrative dieser can be sensitive to focus to a similar extent as er, but that er and dieser differ in how this comes about; dieser interpretation with respect to focus is largely influenced by the next-mention bias, while for er the production bias additionally plays a role. We have also shown that the final-position preference for dieser interpretation is, at least to some extent, independent of focus and that the position influences dieser interpretations largely via the production bias and not the next-mention bias. But the evidence about the independence of final position from focus is conflicting: the demonstrative interpretations show a final-position advantage that is independent of focus, but the demonstrative production bias shows an interdependence of focus and final position.

Our study provides further evidence that demonstratives should be considered separately from personal pronouns. Furthermore, it highlights the importance of considering not only which factors influence pronoun interpretation, but also how these factors exert their influence. Claims in the literature which associate personal pronouns with more prominent and demonstratives with less prominent referents appear to relate to the production biases and not the next-mention bias. Looking at interpretation data alone may gloss over information about the mechanics underlying pronoun resolution, and may miss important differences between different pronoun types.

The studies involving human participants were reviewed and approved by the Ethikkommission der Deutschen Gesellschaft für Sprachwissenschaft. The participants provided their written informed consent to participate in this study.

Contributed to conception and design: CP, PS

Contributed to acquisition of data: CP

Contributed to analysis and interpretation of data: CP, PS

Drafted and/or revised the article: CP, PS

Approved the submitted version for publication: CP, PS

Thanks to Anne Lützeler and Julia Wenzel for their help in setting up the experiments and for assistance in preparing the materials. Thanks to the annotators Claudia Kilter, Anne Lützeler, Brita Rietdorf, Clarissa Selegrad, and Robert Voigt. We thank Sophie Repp for helpful initial discussions about the experimental design. Finally, we are grateful to three anonymous reviewers and to the editor for their thorough and insightful comments, and for their encouragement.

The research for this paper has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project ID 281511265 – SFB 1252 “Prominence in Language” in the project C07 “Forward and backward functions of discourse anaphora” at the University of Cologne, Department of German Language and Literature I, Linguistics.

No competing interests exist.

Supplementary materials can be found on OSF: https://osf.io/a3dsm/

All the stimuli, anonymized participant data and analysis scripts can be found on this paper’s project page on OSF (https://osf.io/a3dsm/). The approved Stage 1 protocol can be found here: https://osf.io/auvmz