2.1 The mechanism: epistemic reinforcement under asymmetry

Many studies have associated the positioning of constituents in the left periphery, which has been defined as “the most leftward positional slot of the utterance, outside the dependency structure of the verb” (Degand 2014: 154), with specific pragmatic effects such as frame-setting, topicalization and focusing. In cartographic approaches, these different effects are mapped onto specific syntactic positions within the left periphery; frame-setting functions are restricted to the “higher” periphery and information-structural functions to the “lower” periphery (see Wolfe 2022: 236–240 for details and references). These studies have shown that Romance languages vary in terms of the degree to which they allow for fronting. Modern colloquial French shows the most restrictions regarding fronting of constituents in comparison to other Romance languages such as Italian (Lahousse & Lamiroy 2012), Venetan (Wolfe 2022: 260) or Spanish (Leonetti 2011: 344). Moreover, word order variability has decreased in French over time (Wolfe 2022), possibly as the result of a grammaticalization process affecting the relation between word order and information structure (Lahousse & Lamiroy 2012). To our knowledge, however, the fronting of iterative adverbials has not yet been studied in detail.

Leonetti (2017: 907–915) argues that fronting of adverbials in examples such as (3) is an instance of resumptive preposing (Cinque 1990: 86–94). In resumptive preposing, “the fronted constituent […] contains an anaphoric element that creates a textual connection with a discourse antecedent” (Leonetti 2017: 908). In (3), for example, ainsi ‘in this way’ refers to the preceding sentence, in which Alexandre’s manner of writing is described in detail.

Rosemeyer (2022) proposes to treat fronted de nouveau as an instance of resumptive preposing. He argues that in example (4), the fronted adverbial links the sentence to a previous utterance by the author.

We accept this characterization as a starting point, but argue that it does not exhaust the pragmatic work performed by fronted de nouveau. By fronting de nouveau in example (4), the author does more than link the sentence to a previous utterance, namely La vague de froid qui frappe le Sud-Est a tué par deux fois la semaine dernière ‘The cold snap hitting the Southeast claimed two lives last week’ (https://www.humanite.fr/societe/-/misere-en-quelques-jours-huit-sdf-ont-ete-victimes-des-temperatures-glaciales-en-france, accessed 1 July 2026). Rather, they foreground the repetition itself. This is the epistemic reinforcement function: the speaker uses fronting to strengthen their epistemic authority over the construal of the situation as recurrent.

Crucially, the insistence on the truth of the repetition leads to a specific rhetorical effect. In (4), fronting of de nouveau frames the explanation (c’est sans doute le mélange létal d’alcool et de froid) as another instantiation of an already established pattern. In this sense, the construction has an epistemic effect at two levels: it reinforces the validity of the author’s inference by presenting the current explanation not as an isolated conjecture, but as the recurrence of a recognizable fatal configuration.

As another example of this effect, consider example (5) below, taken from the FRANTEXT corpus (see Section 3). The penultimate sentence, with postverbal de nouveau, presents the repetition at a relatively abstract level: the speaker once again invents a parallel love story. The following sentence, introduced by fronted de nouveau, does not merely repeat this information. It substantiates it by describing the whole recurrent experiential pattern: silent love, suffering, self-hatred. The fronted adverbial thus gives discourse prominence to the recurrence and retrospectively reinforces the validity of the preceding characterization.

The inferential mechanism underlying epistemic reinforcement can therefore be formulated in two steps. First, utterance-initial de nouveau foregrounds the recurrence and presents it as a salient interpretive frame. Second, this foregrounded recurrence is used to validate a broader proposition that is at issue in the discourse. In (4), the repeated fatal configuration supports the causal inference that the combination of alcohol and cold is responsible; in (5), the repeated affective sequence supports the speaker’s preceding self-characterization as having once again invented a parallel love story. Thus, the construction appears to strengthen the truth of the repetition itself, but its rhetorical force lies in using that repetition to reinforce the truth, plausibility, or experiential validity of a larger proposition.

2.2 Epistemic access and the interactional cost of fronting

Having described the epistemic reinforcement strategy in 2.1, we can now proceed to ask ourselves in which interactional contexts such a strategy is called for. In our view, the concept of at-issueness (Koev 2018) is key to answering this question. In particular, epistemic reinforcement is necessary in contexts in which the truth of the proposition of the de nouveau clause is at issue. Crucially, however, this strategy may come at an interpersonal risk. It depends critically on whether the hearer is in a position to challenge the truth of the proposition at all. Consider two scenarios. In the first, the speaker fronts de nouveau in a context where the repeated event is publicly observable: both participants have, or could have, perceptual access to the evidence. Here, the hearer is not only in a position to evaluate the proposition, but may well have done so already. Treating such a proposition as unchallengeable common ground is therefore an interactionally risky move: the hearer can push back, and the fronting may be perceived as insistent, corrective, or even confrontational. In the second scenario, the repeated event concerns an internal state or evaluation accessible only to the speaker, such as a feeling, an attitude, or a subjective experience. In this case, the hearer simply has no independent means of verifying whether the state holds, let alone whether it has held before. The proposition is, in this sense, structurally unchallengeable, and fronting carries no comparable interpersonal cost.

This asymmetry in epistemic access is precisely what we capture with the distinction between objective and subjective contexts. Objective contexts are those in which the relevant state of affairs can be assessed through shared perceptual or epistemic means: rain falls visibly, a door closes audibly, a fact is verifiable (Wiebe 1994: 235). Subjective contexts are those in which the relevant state of affairs is epistemically private to the speaker: how one feels, what one finds significant, how one evaluates a situation.

Consider the minimal pairs in (6) and (7). A context such as je suis heureux ‘I am happy’ is subjective in the sense that it conveys the speaker’s inner perspective (Traugott & Dasher 2002: 20). The subjectivity of the context can be derived from at least two manifest linguistic indicators: (a) the adjective heureux, which refers to a (speaker-internal) mental state, and (b) first-person reference. Consequently, when a hearer encounters an utterance like (6a), they have no or little independent means of verifying its truth and must simply take the speaker’s word for it. The situation is different for a context such as je suis mouillé ‘I am wet’ (7). Its truth is assessable through shared sensory perception: both speaker and hearer are, or could be, in a position to observe whether the speaker is wet.

This difference in epistemic access determines the interactional cost of fronting, at least partially. In (6b), where the predicate describes an internal state accessible only to the speaker, the proposition is structurally unchallengeable if the speaker has not yet communicated that they are happy: the hearer cannot independently verify whether the speaker was happy before, let alone dispute the claim. Fronting here is therefore interactionally safe in the sense that the speaker reinforces a proposition the hearer is in no position to contest, and no reproachful reading arises.

In (7b), by contrast, the proposition concerns a publicly observable state of affairs. The hearer either has, or could have, independent access to the relevant evidence independently of whether the speaker has previously communicated that they are wet. Treating iterativity as unchallengeable common ground in this context is an interactionally risky move: the hearer may not have acknowledged that the speaker is wet, or may actively disagree, and fronting will then be read as insistent or reproachful.

2.3 From content-level to context-level readings

Our description of the contrast between fronted and non-fronted de nouveau in Section 2.2 carries a further implication: if fronting is systematically associated with subjective contexts (i.e. contexts in which the speaker’s epistemic access is private and the proposition structurally unchallengeable), then fronted de nouveau is, in a precise sense, the more subjective of the two variants. This observation connects to a well-established trajectory in pragmatic change. Traugott (1995; 1999 et passim) has argued that grammaticalizing expressions tend to shift from content-level meanings, which pertain directly to states of affairs in the world, toward context-level meanings, which encode the speaker’s commentary on or attitude toward those states of affairs, a process she terms subjectification. If the fronted variant of de nouveau has come to be associated with subjective contexts, it is therefore to be expected that this association reflects, or is the product of, a broader diachronic shift from content-level to context-level readings.

To give but one example that exemplifies the relevance of these considerations for the analysis of de nouveau, consider Hansen’s (2018) analysis of the diachrony of Middle French or ‘now’. Or originally expresses temporal uses (Hansen 2018: 134–135). Over time, however, it developed discourse-marking functions and in particular, topic shifts. Crucially for our purposes, Hansen explicitly states that this change went hand in hand with an increase in the frequency of or-fronting: “There is little doubt that position is relevant, given that […] increasing pragmaticalization results in or becoming fixed in initial position” (Hansen 2018: 136).

The parallel with or is instructive for the analysis of de nouveau in two respects. First, it suggests that the association of a lexical item with utterance-initial position need not be a syntactic accident, but may itself be the product of pragmaticalization: as an expression develops context-level functions, the left periphery becomes its preferred, and eventually its fixed, position. Second, and more directly, it provides a diachronic motivation for the synchronic pattern described in Section 2.2. If fronted de nouveau has developed a context-level reading through pragmaticalization, then its systematic association with subjective contexts is not merely a distributional preference but reflects an ongoing semantic-pragmatic shift: the construction is in the process of becoming a marker of epistemic reinforcement in its own right, rather than a positional variant of the adverbial. From this perspective, the corpus finding that fronting of de nouveau is more frequent in subjective contexts (Section 3) and the experimental finding that it triggers marked interpretations in objective contexts (Section 4) can be read not only as synchronic evidence for the epistemic reinforcement function, but also as a snapshot of a construction undergoing pragmaticalization.

2.4 Epistemic reinforcement and the Lexical Operator Thesis

If fronted de nouveau is developing into a marker of epistemic reinforcement, a natural question arises: how does this function relate to other, better-described realizations of the same communicative strategy? The Lexical Operator Thesis (LOT; Gutzmann et al. 2020) provides a useful theoretical comparator here, as it offers a cross-linguistically grounded account of constructions whose primary function is precisely to reinforce the truth of a contested proposition.

Under the LOT, verum is not a special case of focus but an independent use-conditional operator. Its function is to intervene in the current Question Under Discussion (Roberts 1996) by explicitly asserting the truth of a proposition in contexts where its denial is salient or epistemically at stake. Crucially, verum marking does not affect the propositional content of the utterance; it operates at the level of use conditions, signaling the speaker’s commitment to the truth of a contested proposition. Gutzmann et al. (2020) argue that this operator is cross-linguistically realized through a range of prosodic and morphological strategies, including nuclear stress shift, particles, and inflectional categories.

We propose that fronting of de nouveau constitutes a syntactic realization of a verum-like epistemic reinforcement function. Like prosodic or morphological verum markers, fronted de nouveau is non-focal, does not introduce new information, and does not alter the truth conditions of the utterance. Instead, it signals the speaker’s insistence on the truth of a proposition, in contexts where that proposition is or could be at issue. The parallel is not coincidental: both verum marking and fronting of de nouveau are interactionally costly when the addressee has independent access to the relevant evidence, and interactionally neutral when they do not. This parallel motivates treating fronted de nouveau as part of a broader, cross-linguistically relevant class of epistemic reinforcement constructions, rather than as an idiosyncratic property of French word order. We return to this connection and the empirical evidence bearing on it in the General Discussion (Section 5).

2.5 Predictions

The theoretical framework developed in this section generates two testable predictions. The first concerns production: if fronting of de nouveau functions as an epistemic reinforcement strategy whose interactional cost depends on the distribution of epistemic access, then speakers should prefer fronting in subjective contexts, where the proposition is structurally unchallengeable, and avoid it in objective contexts, where it carries interpersonal risk. The second concerns interpretation: if listeners are sensitive to this asymmetry, then fronting in objective contexts, where reinforcement appears unwarranted, should give rise to marked, potentially negative inferences about the speaker’s stance, while fronting in subjective contexts should pass largely unremarked. These predictions are formalized in (8) and tested, respectively, in the corpus study (Section 3) and the psycholinguistic experiment (Section 4).

3.1 Determining the variable context

For a variationist approach of the alternation between fronted and non-fronted de nouveau to succeed, it was necessary to determine the variable context, i.e. the contexts in which there is flexibility as regards the positioning of de nouveau. Including contexts where fronting was impossible would artificially overrepresent non-fronted tokens and bias the analysis (Labov 1972: 72). To achieve this, we carefully inspected the syntactic configurations in which de nouveau occurred in our data.

We began with a broad exclusion of environments in which the left periphery was already occupied by other material. In French, sentence-initial de nouveau can co-occur with locatives or discourse markers, as in (9).

The exact positioning of de nouveau in such contexts may be subject to concerns such as the processing cost and length of constituents in the left periphery, leading to interactions that are difficult to control for in the variationist paradigm. In particular, postverbal use of de nouveau sounds more natural in the sentences in (9) due to the presence of other elements in the left periphery. This complexity effect escapes our generalization about the effect of subjectivity on fronting de nouveau, potentially obscuring the factors we aimed to study. These cases were therefore excluded to avoid conflating fronting preferences with constituent-length or processing effects.

From there, we narrowed the operational definition of fronted de nouveau to tokens that were strictly sentence-initial (marked in big caps), involved a finite verb, and to main clause uses. Embedded clauses or clauses following conjunctions, commas, or colons were excluded, since in cases such as (10), fronted de nouveau could never be sentence-initial.

With this definition in place, we applied a series of fine-grained restrictions to the non-fronted tokens, retaining only those that could plausibly occur in a fronted configuration. First, de nouveau had to be syntactically anchored to a finite verb or auxiliary in the non-fronted version or to a participle fulfilling the same syntactic role. Non-finite clauses, as in example (11), were excluded because fronting de nouveau is impossible in this syntactic context.

We also restricted the dataset to affirmative declarative clauses, since sentence-initial de nouveau was unattested in negative polarity contexts such as the invented example (12). In addition, we limited the dataset to declarative clauses only, omitting interrogatives, exclamatives, and imperatives, in order to maintain functional comparability across tokens.

A further restriction concerned the material preceding the subject in fronted contexts. We allowed at most one adverb or prepositional phrase to intervene, provided there was a clear clause boundary (such as a full stop) before de nouveau. For example, (13a) could correspond to the fronted variant (13b).

By contrast, clause-internal insertions without a boundary, as in example (14), were excluded because fronting de nouveau is impossible in such contexts.

As a final step, we also controlled for text-level characteristics that could otherwise distort the comparison between fronted and non-fronted tokens. To avoid genre effects, we retained only non-fronted de nouveau tokens drawn from the same texts in which at least one instance of sentence-initial de nouveau occurred. This ensured that both variants were compared within equivalent stylistic and discourse contexts. In addition, the small number of tokens from poetry (n = 8) was removed. Poetic syntax licenses word orders unavailable in prose, which may artificially inflate fronting rates. Finally, we eliminated cases from texts whose dating was unreliable, particularly historical editions of journals, memoirs, or similar works, where the recorded date referred to the edition rather than the original text production.

In sum, the variable context was defined as environments in which de nouveau could occur either fronted or non-fronted without syntactic or genre-based constraints. This was achieved by excluding tokens in occupied left peripheries, embedded or non-finite clauses, negative polarity, and non-declarative sentence types, and by matching non-fronted tokens to texts containing at least one fronted counterpart. Additional exclusions targeted poetry and texts with unreliable dating. The final dataset contained 461 tokens drawn exclusively from two genres: (short) novels (n = 390) and personal writing (n = 71). The dataset, as well as the corpus analysis presented in this section, are available for replicability at https://osf.io/et2uh/.

3.2 Operationalizing subjectivity

Having established the dataset for the corpus study, we proceeded to operationalize the distinction between subjective and objective usage contexts established in Section 2. In our description of the minimal pairs in (6) and (7), we hinted at the fact that subjectivity is not a criterion that can be defined on the basis of one contextual parameter alone. Rather, different contextual parameters such as predicate type and personal morphology, but also tense, mood and text genre can be seen as contributing to the subjectivity-objectivity distinction. From this perspective, subjectivity can be defined as a latent construct, i.e. a concept that is not directly observable, but rather inferable from linguistic choices in usage contexts. Subjectivity can thus be approached by analyzing associations between the aforementioned contextual parameters in sentences involving use of de nouveau. These parameters are summarized in Table 1.

Latent Class Analysis (LCA) is the most suitable statistical framework for modeling unobservable constructs such as subjectivity. As a variant of structural equation modeling tailored to categorical data, LCA enables the identification of unobserved subgroups or categories within a population (here, variation in subjectivity) by examining their associations with a set of observed variables, in this case, the subjectivity-triggering features listed in Table 1 (Lazarsfeld 1950; Lazarsfeld & Henry 1969; Hagenaars & McCutcheon 2002; Nylund-Gibson & Choi 2018). This method computes the probability that each individual in the dataset belongs to a specific latent class, based on indicators relevant to that class. A key assumption of the model is local independence: once the latent class variable is controlled for, correlations between indicators are assumed to vanish, as any association is attributed to membership in the same latent class (Nylund-Gibson & Choi 2018: 442). For recent applications of this method in linguistics, see Rosemeyer (2024a; 2025), Groß & Rosemeyer (2026), and Gipper (2026).²

The present study adopts the LCA-based approach to identifying situated meanings in corpus data using the published R code in Rosemeyer (2024b). Two contextual profiles of fronted and non-fronted de nouveau were distinguished using the variables outlined in Table 1, implemented via the poLCA package (Linzer & Lewis 2011) in R (R Core Team 2024, version 4.4.1). The analysis generated posterior probabilities of class membership for each token, and class interpretation was guided by the average probability of each contextual predictor within a given class. Models with up to five classes were tested without including additional covariates. Following the recommendations in Linzer & Lewis (2011: 6–7), a two-class solution—associated with the lowest Bayesian information criterion (BIC) value—was selected. For descriptive purposes, the discrete class assignments were employed to illustrate class composition.

To interpret the outcomes of the LCA, it is useful to examine the probabilities of membership in each of the two latent classes (LCs), as determined by the design-related and contextual variables outlined in Table 1. Figure 1 presents these probabilities, showing how tokens in the dataset are allocated to the respective LCs according to the indicators. This representation facilitates an interpretation of the two LCs within our population of de nouveau tokens in terms of their relative subjectivity.

Figure 1: Probability of assignment of de nouveau tokens to one of the two latent classes, by contextual features.

With n = 306 cases, a greater number of cases of de nouveau was assigned to LC 1, whereas only n = 155 cases were assigned to LC 2. We interpret the results from the LCA in terms of the distinction between subjective and objective contexts, where LC 1 is more likely to represent objective contexts and LC 2 is more likely to represent subjective contexts. With the exception of mood, the parameters listed in Table 1 are clearly relevant to the distinction of the two latent classes in our data. LC 1 was typically assigned to sentences in third person involving objective predicates and perfective past in novels. For instance, example (15) has the maximum probability (=1) of assignment to LC 1. In contrast, LC 2 tokens of de nouveau typically involve first person morphology, subjective predicates, present perfect, present or future tense, and occur in personal writing. As a prototypical example for LC 2, consider (16), which has the maximum probability (=1) of assignment of LC 2. Whereas in (15) the repetition is clearly evident for all discourse participants (the female character is likely aware of the fact that she looked at the protagonist before), in (16) it is only accessible to the protagonist, whose thoughts are reported.

To summarize, LC 2 captures the type of subjective contexts discussed in Section 2: contexts in which the hearer lacks independent access to the truth of the repeated event and must rely on the speaker’s internal perspective, whereas LC 1 captures objective contexts.

3.3 Subjectivity and fronting

The relationship between subjectivity, epistemic asymmetry and fronting of de nouveau is easily visible in examples such as (17–19), drawn from our dataset. In (17), avoir quelqu’un à soi describes an intimate, emotionally private state of affairs that is epistemically accessible only to the narrator. The addressee and the reader are in no position to challenge or verify this evaluation. Fronting of de nouveau in this context is therefore interactionally safe: the speaker reinforces a proposition that the addressee cannot contest, and accordingly no reproachful or corrective reading arises. The effect is one of affective intensification rather than epistemic imposition.³

In (18), by contrast, clearing one’s throat is a publicly observable physical action to which all discourse participants have equal perceptual access. With the relevant state of affairs symmetrically accessible, de nouveau appears in-situ and the sentence reads as a neutral narrative description. Together, (17) and (18) are in line with P1: fronting tends to occur where epistemic access is asymmetric, while in-situ placement is the default in objective contexts.

The examples also offer a preliminary illustration of P2, which we will test experimentally in Section 4. P2 predicts that when the relevant evidence is publicly accessible, fronted de nouveau should be perceived as more marked or negative than non-fronted de nouveau, because reinforcement of an independently verifiable recurrence may suggest that the addressee has failed to recognize something they should have recognized. Like (18), example (19) involves an objective context: the narrator has been trying to reason with his father throughout the passage, making the iterativity fully salient to both parties. Because the father is the co-present addressee of these reasoning attempts and consequently has independent access to the same shared situation, the use of fronted de nouveau presents the iterativity as something he should already have acknowledged, producing exactly the corrective force that P2 predicts will arise when epistemic access is symmetric.

While these three examples are necessarily illustrative rather than systematic, they motivate the quantitative analyses in the remainder of this section and in Section 4, which test our two predictions respectively across the full dataset. We turn first to test the prediction that fronting of de nouveau is relatively more likely in subjective contexts than in objective contexts (P1). We applied linear mixed-effects regression modeling using the lme4 and lmerTest packages in R (Bates et al. 2015; Kuznetsova et al. 2017) to calculate the effect of four predictor variables on the probability for a token to be assigned to LC 2, representing subjective contexts (dependent variable LC 2). Using the posterior probability of LC 2 membership as the dependent variable preserves the graded nature of subjectivity as estimated by the LCA, rather than forcing a discrete classification. This specification follows the approach described in Rosemeyer (2024a) and reflects our theoretical interest in the association between fronting and subjective contexts rather than in the prediction of fronting per se: the model asks whether fronted tokens are more subjective than non-fronted tokens, not whether subjective contexts predict the choice to front.

The predictor variables are described in Table 2. To test our prediction, we included the predictor variable DNPos, which operationalizes fronting vs. non-fronting of de nouveau. To increase reliability of the results, we also included three control variables: Adverbial to control for whether or not another adverbial was present in the sentence, SentenceLength to control for possible processing effects and Year to control for a possible diachronic trend. Given that we sampled only a random number of verb lemmas on the basis of their occurrence with de nouveau, we included verb lemma as a random effect. Including random effects has been shown to substantially increase model quality (see Gries 2015).

Table 3 summarizes the results from the linear mixed-effects model. The model explained 21% of the variance in LC 2 assignment probability via fixed effects (marginal R² = 0.21), and 35% when including random effects (conditional R² = 0.35).

While no significant effect was found for the control variables Adverbial and SentenceLength, both the effect of DNPos and Year were significant at p<.05*. As predicted, fronting of de nouveau significantly increased the likelihood for a token to be assigned to LC 2, representing subjective contexts. We visualize this effect in Figure 2. The effect of fronting is modest in size, which is expected given that subjectivity is modeled as a multifactorial latent construct rather than as a deterministic property of any single constructional choice.

Figure 2: Predicted probability of LC 2 assignment by fronting of de nouveau. Plot was generated using the sjPlot package (Lüdecke 2018).

Finally, the model predicted assignment to LC 2 to increase over the 22-year time span covered by our dataset. While the specific reasons for this change need to be explored in more detail, note that the directionality of the change would seem to be compatible with the hypothesis of an ongoing pragmaticalization of de nouveau to context-level readings.

To summarize, the corpus analysis shows that fronting of de nouveau is systematically associated with subjective contexts, characterized by asymmetric access to evidence between speaker and addressee. This distributional pattern supports Prediction 1, according to which fronting is relatively more likely in contexts where the proposition cannot be independently verified by the addressee. Taken together, these findings suggest that fronting is not a purely syntactic option, but a pragmatically constrained choice sensitive to epistemic asymmetries in discourse.

4.1 Pre-study

Prior to the main study, a pre-study was conducted to match items from the objective and subjective conditions on their average valence while ensuring sufficient contrast in their degree of subjectivity. 26 potentially objective and 21 potentially subjective items that did not exhibit strong positive or negative valence (e.g., habillé – dressed, stoïque – stoic) were initially pre-selected. In addition, three reference items (triste – sad, joyeux – happy, furieux – furious) were included to provide positive and negative baselines for valence judgments.

Pre-selected items were subsequently evaluated by 25 native French speakers (age 18 to 39, M = 28.88 years, SD = 5.96; 16 female, 9 male) with respect to both valence and objectivity. For valence, participants judged the extent to which a state could be considered positive or negative for a person. The valence rating scale, ranging from 1 (very positive) to 5 (very negative), followed the valence scale of the Self-Assessment Manikin (SAM; see Figure 3) as originally introduced by Bradley & Lang (1994). To assess objectivity, participants were instructed to evaluate for each pre-selected item how far external observers would be able to assess the respective states of a person on a scale from 1 (impossible to assess) to 5 (very easy to assess). The objectivity rating scale additionally included an option to indicate that the expression is not sensible.

Figure 3: Valence rating scale used in the pre-study, based on Bradley & Lang’s (1994) Self-Assessment Manikin (SAM).

The survey was programmed on PsyToolkit (Stoet 2010; 2017) and distributed via Prolific. Based on these judgments, 14 items were selected for each condition (subjective vs. objective) of the Main Experiment, resulting in 28 items overall. None of the selected items were judged as not sensible by more than one participant. The two sets of items were on average comparable in valence ratings (objective mean = 2.75, SD = 0.71; subjective mean = 2.83, SD = 1.03) but differed in their mean objectivity ratings (objective mean = 4.55, SD = 0.22; subjective mean = 2.97, SD = 0.33; see Figure 4).

Figure 4: Rating properties of items selected for the main experiment. Horizontal and vertical lines correspond to the mean valence (horizontal) and objectivity (vertical) ratings for selected objective (teal) and subjective (salmon) items.

Note that whereas items’ objectivity ratings differed between conditions, objective items were frequently judged as mid-scale items, as the internal states they described could still be inferred by external observers to some extent through facial expressions. An overview of all selected items and their rating properties is provided in the appendix (Table 6).

4.2 Main experiment

4.2.2 Procedure

Preceding the Main Experiment, participants read the digital information sheet, provided informed consent, and completed a demographic questionnaire. The survey and the subsequent experiment were programmed on PsyToolkit.

For each trial of the experiment, participants were presented with a fixation cross at the centre of the screen for 500 ms, followed by a screen showing one stimulus sentence at a time, with the question Le locuteur veut-il dire cela de façon positive ou négative? (Does the speaker mean this in a positive or a negative way?) and the rating scale below the question. As in the pre-study, the rating scale corresponded to the Self-Assessment Manikin (SAM; Bradley & Lang 1994); in addition, the endpoints of the scale were labeled très positive and très négative (see Figure 3). Participants were instructed to respond within five seconds by clicking on one of the five figures, which corresponded to an underlying scale from 1 (very positive) to 5 (very negative) in the output data. Following each response, an “x” was displayed at the selected scale section for 300 ms, followed by the fixation cross of the next trial. If no response was made, the next trial’s fixation cross was presented after 5 s. All participants completed the task on a desktop computer or laptop; participation via other devices (e.g., smartphones) was not permitted.

To implement a within-subjects manipulation of the variable Position (of de nouveau) without repeating items for the same participant, two stimulus lists were created and items were rotated across these lists. Consequently, all subjective and objective items appeared in sentences with and without fronting of de nouveau. For instance, the item stoïque was presented as De nouveau, je suis stoïque in the fronted condition, and as Je suis de nouveau stoïque in the non-fronted condition. All stimulus sentences followed the same structure (je suis + item, with fronted or non-fronted de nouveau) to make sentences comparable between conditions of the main experiment, and to allow for maximal subjectivity in the subjective condition that involved unobservable traits.

Each list additionally contained an equal number of filler items, including the four control items (see Table 7 in the appendix for an overview of all stimuli including filler and control items). Participants were randomly assigned to one of the two lists, and trials were presented in random order.

After the experiment, participants were asked about the suspected purpose of the study and any problems encountered during the experiment. The experimental setup is illustrated in Figure 5.

Figure 5: Experimental setup of the main experiment.

4.2.3 Results

Data were analyzed in R (R Core Team 2024, version 4.4.1). After excluding two participants based on the pre-registered outlier procedure (see Section 4.2.1.), rating data were visualized and subsequently analyzed in a linear mixed-effects model.

A total of 2.06% of trials were excluded, as they were not responded to within the five-second window. Visual inspection of the graphs (see Figure 6) suggested that, consistent with the prediction, fronting of de nouveau elicited more negative interpretations, with a significantly more pronounced effect for objective items.

Figure 6: Valence ratings for subjective (red) and objective (green) items in the main experiment.

To test the significance of this descriptive finding, a linear mixed-effects model was implemented using the lme4 package (Bates et al. 2015). All categorical predictor variables were sum-to-zero coded. To balance type I and type II error risks, a balance between a data-driven and a hypothesis-driven approach to model selection was sought (e.g., Cunnings 2012; Barr et al. 2013; Matuschek et al. 2017). Accordingly, the best-fitting random effects structure was selected first by establishing an intercept only model with the maximal random structure justified by the design. This initial model was subsequently optimized through the step function (lmerTest package; Kuznetsova et al. 2017). Finally, the fixed effects Objectivity and Position (of de nouveau), including their interaction, were added to directly assess the prediction. The resulting best-fitting model was:

Valence Rating ~ Objectivity * Position + (Position | Participant) + (1 | Item)

Visual inspection of the model’s residual plot did not indicate a violation of the homoscedasticity assumption. P-values were obtained through the lmerTest package, whereas random effects estimates were calculated using the get_variance function (insight package; Lüdecke et al. 2019). Full model results are reported in Table 4.

Table 4: Results table of the main experiment.

Fixed Effects		Estimate (β)	Std. Error	t-value	p-value
Intercept		2.89	.15	19.63	<.001***
Objectivity		–.11	.15	–.73	.474
Position		.03	.01	2.17	.032*
Objectivity: Position		.02	.01	2.09	.037*
Random Effects
σ2	.63
τ00	.02Participant
	.60Item
τ11	.01Participant.Position
ρ01	–.18Participant
* p < .05, ** p < .01, *** p < .001

Items were rated as significantly more negative when they were presented in a sentence in which de nouveau occurred in sentence-initial position, as indicated by a main effect of Position (β = .03, p = .032). Crucially, and in line with the prediction, a significant interaction between Objectivity and Position (β = .02, p = .037) confirmed that this tendency was significantly more pronounced for objective than for subjective items. The effect sizes are small, but they are theoretically informative because the manipulation involved a minimal word-order contrast within otherwise highly controlled stimuli.

None of the participants reported any problems completing the experiment. When asked about the presumed purpose of the study, most participants either reiterated the instructions, referred to a general role of emotions, or suggested that individual differences could influence how specific traits of a person are perceived. Only two participants mentioned a possible role of sentence structure, and two additional participants pointed to a potential influence of adverbials. Moreover, only two out of the 118 participants in the final sample explicitly noted the repeated exposure to de nouveau.