Glossa Psycholinguistics

3.1 Method and stimuli

The experimental arrangement, overviewed in Table 1, consisted of four subexperiments, which constituted mutual fillers. Experiments 1 and 2 were designed as parallel in Czech and Russian (with only minor differences in particular lexical realizations).⁸ Experiment 1 conceptually replicates the gist of Tian et al. (2016) and the design of Experiment 2 follows Tian et al. (2021), facilitating cross-linguistic comparison. Experiment 3 is an exploration of the effect of bias-inducing particles. Filler stimuli involved small-scale exploratory experiments (not reported here). Items from all experiments were combined, amounting to a total of 80 stimuli for each participant. For each experimental list, item order was pseudo-randomized, so that two items from the same experiment could not appear one after another and two conditions from the same experiment could not be presented one after another, even when separated by an item from another experiment. We use the standard Latin Square design, so participants heard each condition with eight items, but only one condition per given item.

The main task was to listen to prerecorded mini-dialogues between a man (M) and a woman (W) and simultaneously look at a visual scene containing four pictures – two target pictures and two distractors – as in Figure 1. An English translation of an actual stimulus for Figure 1 is available in (10). The auditory stimulus went off 1 second after the visual part appeared. After listening to the mini-dialogue, participants selected, by pressing the corresponding button, the picture they thought constituted the best match to the dialogue’s conclusion; for the example in (10), it is the bottom left picture of the dirty boots in Figure 1. The general shape for all experimental items was the same, as shown in (11). The auditory stimuli were recorded by four native speakers in the phonetic laboratory at the Faculty of Arts, Charles University.

Figure 1: Visual stimulus for (10).

Our picture materials were taken from Tian et al. (2021) and only slightly adapted. Contrary to Tian and colleagues’ design, our visual scene did not involve a visual representation of the agent of the activity described in M’s utterance. We were concerned that looks to the agent would attract gaze and thereby reduce the amount of looks to the targets, which could lead to unnecessary information loss, esp. in the interrogative (VS) orders, where the effect of interrogativity on processing is expected to step in very early on.

The whole dialogue was divided into seven interest periods (IPs): four in the first utterance, a one-second gap, and two in the response (see Table 2). The first utterance consisted of the subject (a common male or female name), a verb in the past tense, a temporal or locative adverbial introduced to provide more processing time,⁹ and the object, which corresponded to the entity depicted in the targets of the visual scene. The relative positions of subject and verb were varied in the PQ experiments. The response was split into ‘I think that’ and either ‘yes’ or ‘no’.

3.2 Participants, procedure, and equipment

We tested 52 participants per language and provided them with a compensation of €11 (Berlin) / 350 CZK (Prague). All participants had uncorrected or corrected-to-normal vision. The reliability of the participants was checked via the picture selection task in the positive V1 question condition. A participant was considered reliable if, in 6 or more trials out of 8, they chose the picture corresponding to the polarity of the response particle.

The Russian part of the study included 51 participants (one was excluded as unreliable; 38 females, mean age: 30.5 years). They were Russian native speakers who had recently relocated, recruited via social networks. The experiment took place in Zentrallabor of the Faculty of Linguistics and Literature at Humboldt-Universität zu Berlin. The Czech study covers 48 participants (four excluded; 37 females) who were native Czech speakers and mostly students recruited via an online university psycholinguistics pool. It was carried out in Prague, in the LABELS lab (affiliated with the Faculty of Arts, Charles University, and the Institute of Psychology of the Czech Academy of Sciences).

Upon arrival at the laboratories, participants filled in a short questionnaire and signed a consent form. After they received all instructions and underwent calibration, all participants tried out two test items to make sure they understood the task. A short break was enforced after completing 40 stimuli (i.e., in the middle of the session). The whole procedure took approximately one hour.

The experiment was performed on the EyeLink Portable Duo eye-tracker in Berlin and on the EyeLink 1000 Plus eye-tracker in Prague. Experiment Builder (SR Research Ltd., 2020) and R (R Core Team, 2022) were employed for data collection and statistical analysis. We have obtained full approval from the Ethics Committees at Humboldt-Universität zu Berlin and Charles University for all experiments.

3.3 Dependent variable

Independent variables were different for each experiment, and will be described below. The dependent variable, following Tian et al. (2021), was the natural logarithm of the ratio of the probability of looks to the $p$ picture over the probability of looks to the $\overline{p}$ picture:

$\text{ln}\left(\frac{P_p}{P_{\overline{p}}}\right)$

This metric is commonly used in visual world analyses that compare fixations on two areas of interest (see Tian et al., 2016, and references therein). To compute the probabilities of looks, total fixation durations to the $p$ and $\overline{p}$ pictures during each interest period were divided by the duration of the interest period. Time boundaries of individual interest periods were obtained by annotation of the auditory stimuli, using Praat software (Boersma & Weenink, 2009). Czech stimuli were annotated manually. For Russian stimuli, WebMAUS Basic service (Kisler et al., 2017) was applied to generate automatic annotations that were then manually corrected. To prevent division by zero in cases where there were no fixations on the $\overline{p}$ picture, 0.5 ms was added to all values (total fixation durations and interest period durations). The natural logarithm transformation aims at making the distribution of the data less skewed and, therefore, more suitable for the statistical method we applied (Baayen, 2008, p. 31).

The dependent variable (from now on referred to as log-ratio) is centered around 0. Since the natural logarithm of 1 is 0, a zero log-ratio indicates an equal probability of looks to the $p$ and $\overline{p}$ pictures. A positive log-ratio corresponds to a higher probability of looks to the $p$ picture than to the $\overline{p}$ picture. The higher the positive value, the larger the preference. A negative value indicates a preference for the negative alternative, and lower values represent a larger difference.

This derivation of the dependent variable implies that only data for fixations on the pictures $p$ and $\overline{p}$ were used for the analysis. Although fixations on the distractors were also measured, they are not of interest to our research hypotheses and will not be reported.

Picture selection was an auxiliary dependent variable.

3.4 Data manipulation and statistical analysis

Data preprocessing was performed in R¹⁰ using the eyelinkReader (Pastukhov, 2024), vctrs (Wickham et al., 2023), plyr (Wickham, 2011), here (Müller, 2020), and tidyverse (Wickham et al., 2019) packages. During this step, fixations that lasted less than 80 ms were excluded, because short fixations are commonly attributed to false saccade planning (Bosker et al., 2014; Tian et al., 2021). All interest periods were offset by 200 ms to account for the time needed to plan an eye movement (Hallett, 1986; Tian et al., 2021). Trials in which no fixations on $p$ or $\overline{p}$ pictures were registered during the whole auditory stimulus (a question or an assertion, the gap, and the response) were excluded from the analysis.

The rationale of our statistical analysis largely follows Tian et al. (2016, 2021). We fitted linear mixed-effects (LME) models using lme4 (Bates et al., 2015), with statistical inference performed using lmerTest (Kuznetsova et al., 2017). The analysis consisted of two parts. First, log-ratio was compared to 0 for each interest period in each separate condition. The dependent variable being significantly different from 0 would imply a significant difference between the probabilities of looks to the $p$ and $\overline{p}$ pictures for that condition and interest period. This would allow for an indirect comparison between conditions, based on when a significant preference for one of the alternatives emerged. For this analysis, we followed the guidelines in Ito & Knoeferle (2022). For each interest period, LME models were fitted to data from each condition. Fixed effects were, therefore, excluded from these models, leaving only the intercept and the random intercepts, as shown in the formula below.¹¹ Since each condition was analyzed separately, fitting random slopes would not be justified. If a model did not converge or returned a singularity warning, one of the random intercepts was removed.

$\text{log\_ratio\_transf}\hspace{0.17em}\hspace{0.17em}\sim \hspace{0.17em}\hspace{0.17em}\text{(1|item)+(1|participant)}$

Second, the impact of fixed predictors on log-ratio was tested. LME models were fitted to each interest period. All predictors for fixed effects were categorical and binary and were effect-coded (exact values for each analysis are given in the corresponding subexperiment sections). When fitting random effects of item and participant, we started from the maximal model that was supported by the design (Barr et al., 2013). To prevent conversion and singularity issues, correlation parameters were excluded from all the models. This is justified by the fact that the research hypotheses only concerned fixed effects and there were no specific expectations about variance between participants or items (Bates et al., 2018). Furthermore, the control parameter was set to control = lmerControl(calc.derivs = FALSE) (Brown, 2021). If a model still did not converge without a singularity warning, the number of random slopes and, if needed, intercepts was gradually reduced up until successful convergence. Plots were generated based on model predictions obtained with help of the ggeffects package (Lüdecke, 2018).¹²

4.1 Design and hypotheses

The goal for this experiment was to investigate the processing of positive and negative assertions, so polarity was the only manipulated variable. An example of a Russian item (its linguistic part) is provided in (12). Czech items were fully analogous.

We hypothesized that the gaze would be affected by the polarity of the assertion: positive assertions would instigate fixations on the picture representing the positive alternative (also called positive fixations/picture, or simply $p$, below), in this case, an open book, and negative assertions would lead to fixations on the picture representing the negative alternative (negative fixations/picture, or simply $\overline{p}$), i.e., a closed book.

4.2 Descriptive findings

Figure 2 shows how the proportion of fixations on $p$ and $\overline{p}$ evolved over time during the dialogue (the beginning of all interest periods was offset by 200 ms; see 3.4). To generate this and the corresponding visualizations in the following subsections, each trial was split into 1000 subsequent fragments of equal duration. For each fragment, it was calculated if there was a fixation on $p$ or on $\overline{p}$. The proportion of fixations, thus, corresponds to the proportion of experimental trials in which the positive or negative picture was fixated on in each fragment.

Figure 2: Proportions of looks to $p$ and $\overline{p}$ in Czech and Russian assertions, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

The proportions of looks both to $p$ and $\overline{p}$ grow as information about the target object is revealed to, and processed by, the participants. However, this increase is not equal for the positive alternative and the negative alternative, and is conditioned by the polarity of the assertion. In both Czech and Russian, positive assertions are characterized by a higher proportion of looks to $p$, and negative assertions, to $\overline{p}$. This tendency seems to establish itself in late periods of the assertion (adverbial in Czech negative assertions, object in the other cases), and to become most apparent in the gap or the first part of the answer. The next subsections report on the statistical analysis that estimates the significance of those trends.

4.3 Models

In this experiment, the sole predictor was polarity, a categorical variable with two levels: positive (coded as 1) and negative (coded as –1).

As a starting point, the maximal reduced LME models were fitted to each interest period.¹³ Aside from the fixed predictor, each model included random intercepts and random slopes for items and participants. If needed, the random effects structure of individual models was gradually simplified to resolve convergence and singularity issues (see 3.4).

4.4 Results

Figure 3 shows mean log-ratio (on the y-axis) over the interest periods (on the x-axis) across languages. On each facet, the black horizontal line indicates zero log-ratio, which would correspond to equal mean proportions of fixations on $p$ and $\overline{p}$. Green dots mark combinations of interest period and condition in which log-ratio was significantly different from 0, according to the LME analysis of the intercept described in 3.4.

Figure 3: Average progression of log-ratio of proportions of fixations on $p$ over $\overline{p}$ by interest period in Czech and Russian assertions. Error bars indicate 95% confidence intervals. Green dots show where log-ratio was significantly different from 0.

It is evident that assertions tend to trigger a preference for the pictures which correspond to their polarity. In the positive condition (represented by the violet line), mean log-ratio is above 0 from the adverbial to the first part of the response in Czech, and from the object on in Russian. Negative assertions (orange line) show the contrary tendency. From the adverbial on, mean log-ratio in the negative condition is below 0 in both languages, indicating longer average fixation times on the negative picture. However, the statistical comparison of the dependent variable to 0 rendered only few significant differences, all after the target utterance in Russian.

Although the preference for the picture with the corresponding polarity may be characterized as rather moderate, the difference between positive and negative assertions is robust enough to be supported by the statistical models. In both languages, a significant effect of polarity was found in the gap (Russian: t = 3.381, p < 0.01; Czech: t = 2.873, p < 0.01) and in the first part of the answer (Russian: t = 3.692, p < 0.001; Czech: t = 2.059, p = 0.04). Figure 4 shows predicted values of the dependent variables, based on the LME models fitted for these interest periods.

Figure 4: Log-ratio predicted by LME models in selected interest periods of Czech and Russian assertions.

It is worth mentioning that Russian and Czech exhibit different patterns in the second part of the response (‘yes’ or ‘no’). In Czech, mean log-ratio in this interest period is relatively close to 0 in both the positive and the negative conditions. This indicates roughly balanced proportions of looks to $p$ and $\overline{p}$, which is in accordance with the equal amount of stimuli containing ‘yes’ and ‘no’ answer particles in each experimental list. A similar tendency can be observed for Russian positive assertions. In contrast, the negative condition in Russian shows a clear preference for the negative picture during the answer particle period, log-ratio being estimated as significantly lower than 0 by the model. A more complete picture can be achieved if we observe the outcome for ‘yes’ and ‘no’ response particles separately (see Figure 5). In Czech, ‘yes’ seems to be associated with the $p$ picture and ‘no’ with the $\overline{p}$ picture, after both positive and negative assertions. In Russian, however, both response particles seem to trigger fixations on $p$ after a negative assertion. These fixation patterns are in accordance with the results of the picture selection task (see Figure 6).

Figure 5: Log-ratio in the second part of the response (‘yes’/‘no’) in Czech and Russian assertions.

Figure 6: Picture choices in assertions.

4.5 Discussion

In both languages, positive assertions instigated a higher proportion of looks to the $p$ picture over looks to the $\overline{p}$ picture than negative assertions. The difference became significant in the gap period, i.e., right after the assertion had been fully heard, and continued through the first part of the response (‘I think that’). Comparisons of the log-ratio measure to the 0 line showed that looks to the picture corresponding to the polarity of the assertion diverged significantly from a balanced distribution in this same period for Russian positive assertions, and for Russian negative assertions upon hearing the answer particle.

This result is in line with Tian et al. (2016), who found that simple English positive vs. negative assertions significantly favored looks to the picture corresponding to their polarity from about 900 ms after the end of the verb period on.¹⁴

The result is consistent with the idea that the semantic interpretation of the utterance is built up incrementally during processing, with a time lag between the cue and the zooming in on the corresponding visual stimulus. Even though the visual stimuli represented only the objects to which the action had (not) been applied, not the agents, verb meaning itself already allowed for a prediction as to which object might fit. Thus, the somewhat delayed fixation on the expected visual stimulus may come as a surprise. As argued in some detail in Tian et al. (2016), there is good reason to assume that the positive proposition remains active in processing for a rather long time in both positive and negative utterances.

In addition to testing our main hypotheses, the findings obtained also provide some insights into the processing of the response particles. Previous linguistic analyses have shown that the status of the response particle in reactions to a negated assertion differs between Czech and Russian. In Czech, the negative particle ne ‘no’ encodes the negative polarity of the response and ano ‘yes’, agreement with the positive part of the negative prejacent (Hrdinková & Šimík, 2025). In Russian, both the positive da ‘yes’ and the negative net ‘no’ can mark agreement with a negative prejacent, while only the combination of a negative response particle and a following positive verb can indicate rejection (Esipova, 2021, ex. (5)).

Picture choices confirm this analysis (see Figure 6): In Russian, both the positive and the negative response particles were associated with $\overline{p}$. This reasoning explains the difference in fixation patterns we observe during the second part of the response to the negative assertions. While in Czech, ‘yes’ can agree with the positive part of the prejacent and instigate positive looks, in Russian, it conveyed agreement with the negative statement, triggering negative fixations.

5.1 Design and hypotheses

We used a 2 × 2 fully-crossed factorial design for the PQs experiment. The variable verb position represents two major interrogative strategies in Czech and Russian – verb-initial order (V1) and verb-second order (V2) combined with question-specific intonation. The second variable is polarity; PQs were either positive or negative. Both variables were manipulated within 32 items and within subjects. The linguistic stimuli are illustrated in (13) by one Czech item in all conditions. Russian stimuli were fully parallel and differed only in condition V1, in that the verb is followed by the particle li.

Factoring in the results of Tian et al. (2021), we hypothesized that the processing of PQs would be accompanied by looks to the $p$ picture, independently of their polarity, reflecting the interpretation of positive bias.

We further assume that this tendency will be modulated by our experimental and cross-linguistic manipulations. First, because V1 questions in Russian are accompanied by the li particle (attached to the verb), a salient question marker morphologically related to the disjunction ili ‘or’ (see Section 1), we expect the looks to be more evenly distributed between $p$ and $\overline{p}$ than in the V2 condition, where question meaning is conveyed by means of intonation. This expectation also follows from the assumption that V1 questions in Russian are unbiased (Šimík, to appear; Korotkova, submitted). The effect of the li particle can be set against the background of Czech, where we expect to observe a “pure” effect of interrogative word order; the results can, thus, reasonably be expected to be similar to what Tian et al. (2021) observed for English, i.e., dominant looks to $p$. V2 questions in Czech are less neutral than in Russian, in the sense that they are (evidentially) biased (Staňková & Šimík, 2025), which, in turn, can be expected to lead to increased fixations on the proffered alternative.

The effect of polarity is expected to differ depending on word order. Negation in V1 questions in both Russian and Czech arguably maps to positive epistemic bias (Šimík, to appear). The expectations for V2 questions are not straightforward. On the one hand, negation could instigate more looks to $\overline{p}$ (esp. in Czech, where V2 is accompanied by negative evidential bias); on the other, negation triggers positive epistemic bias even in V2 questions (possibly leading to a conflict of both biases), which might neutralize the effect of negation.

The word order manipulation in both languages is further expected to be reflected in the different timing of where question marking affects the fixations: the preference for $p$ developing earlier in V1 than in V2. More specifically, in Russian V1, it becomes evident that the processed utterance is a question in the initial (verb) period; the effect is reinforced by the li particle. However, li as a marker of unbiased PQs (see above) may also lead to the opposite effect on timing, causing a more balanced distribution of fixations in the early periods. The situation in Czech is different, because Czech (being pro-drop) allows for non-interrogative V1 utterances; but as soon as the postverbal overt subject is encountered, the likelihood of question interpretation reaches near-certainty (see Šimík, to appear). In Russian V2 questions, interpretation is conveyed by the dedicated pitch accent on the verb; question interpretation should, therefore, step in soon after that. In Czech V2, it is the utterance-final rising tone that is considered to be a reliable question cue. Question interpretation is, thus, expected to step in fairly late.

5.2 Descriptive findings

Figures 7 and 8 display the temporal progression of fixation proportions on $p$ and $\overline{p}$ for Czech and Russian PQs, respectively. Almost all conditions seem to exhibit a clear preference for the positive alternative which manifests itself starting from the second to fourth period of the question. Only in Russian negative V2 questions do the looks tend to be distributed in a more balanced way.

Figure 7: Proportions of looks to $p$ and $\overline{p}$ in Czech questions, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

Figure 8: Proportions of looks to $p$ and $\overline{p}$ in Russian questions, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

5.3 Models

The statistical analysis included two categorical predictors: polarity and verb position. Positive polarity and V1 verb position were coded as 1, negative polarity and V2 verb position as –1. The random effects structure includes random intercepts for item and participant and random slopes for polarity, verb position, and their interaction both for item and participant.¹⁵

To explore potential differences between Russian and Czech in PQs processing, additional models were fitted to selected interest periods of V1 (verb, subject, and adverbial) and V2 (verb, adverbial, object, and gap) questions separately. Those models included polarity and language as predictors, with positive polarity and Russian language coded as 1, and negative polarity and Czech language as –1. The maximal models were fitted to the data subsets for V1 and V2 questions that included data from both languages.¹⁶ As each participant was exposed to stimuli only from one language, the random slope of language by participant was not included. Concerning items, they were parallel in Russian and Czech. Therefore, it was assumed that effects of both predictors as well as their interaction could vary by item. In both analyses, random effects structure was gradually simplified, as described in 3.4, if confronted with conversion and singularity issues.

5.4 Results

Figure 9 displays the progression of mean log-ratio values for positive and negative PQs with verb-initial (V1) and verb-medial (V2) word orders in Czech and Russian. Fixations on the $p$ picture dominate across the board, most of the values being above 0 (except for the first and the last interest periods in some conditions). This overall trend seems to be consistent across all conditions, apart from negative Russian V2 questions. However, an indirect comparison of statistical intercept-based models (see 3.4) reveals differences in temporal emergence of a significant preference of the $p$ picture over the $\overline{p}$ picture.¹⁷

Figure 9: Average progression of log-ratio of proportions of fixations on $p$ over $\overline{p}$ by interest period in Czech and Russian questions. Error bars indicate 95% confidence intervals. Green dots show where log-ratio was significantly different from 0. Red dots show where no LME model could be fitted.

5.4.2 Russian

In the positive V1 condition, a significant preference for the $p$ picture was found in the object region and in the first part of the response; no reliable estimation was obtained for the gap period. Negative V1 questions exhibit the earliest significant preference for $p$ throughout the experiment, namely starting from the subject and continuing until the object. In positive V2 questions, fixations on $p$ significantly prevail from the adverbial to the first part of the response on. The negative V2 condition in Russian is the only one for which no significant preference for the $p$ picture was found in any interest period.

Based on the statistical models that included verb position and polarity as predictors, significant effects were obtained in several interest periods for Russian. These are illustrated in Figure 10, which shows predicted values of the dependent variable across conditions and relevant interest periods.¹⁸ In the adverbial period, an interaction of polarity and verb position (t = –2.173, $p$ = 0.035) was found. In the V2 condition, positive questions were estimated to have more fixations on $p$ during this period than negative questions. For V1 questions, the model predictions are reversed, with negative questions showing a stronger preference for $p$. In the object period, the model identified a main effect of verb position (t = 2.074, p = 0.0497), V1 questions being characterized by more looks to $p$ than V2. Auxiliary LME models fitted to the data subsets divided by polarity reveal that the main effect is due to the simple effect of verb position in the negative condition (t = 2.320, p = 0.029); the corresponding simple effect in the positive condition is absent (t = 0.573, p = 0.57). Finally, a main effect of polarity was discovered in both parts of the response (t = 2.997, p < 0.01 for ‘I think that’ and t = 3.547, p < 0.001 for ‘yes/no’).

Figure 10: Log-ratio predicted by LME models in selected interest periods of Russian questions.

5.4.3 Language comparison

We found an interaction of polarity and language in the adverbial period of V1 questions (t = –2.184, p = 0.037). The corresponding model predictions are shown in Figure 11. The statistical inference indicates different directions of the polarity effect in the two languages. While in Czech, positive questions are characterized by more looks to $p$ than negative questions, in Russian it is the negative ones that instigate more fixations on $p$. No significant effects were found in the periods of V2 questions that were investigated.

Figure 11: Log-ratio predicted by the LME model for the adverbial period of Czech and Russian V1 questions.

5.5 Discussion

The results of Experiment 2 largely corroborate previous findings of Tian et al. (2021) for English and French: PQs generally trigger looks to the positive picture, even if they have negative polarity. We see this in most of our experimental conditions (see Figure 9): in Czech, the positive picture received more fixations than the negative one at the latest in the adverbial or the object period, independently of polarity and verb position; in Russian, the same holds, with the exception of V2 negative questions, where neither fixations on the $p$ picture nor on the $\overline{p}$ picture dominate significantly. We interpret this result as evidence for the salience of positive epistemic bias in PQs.¹⁹ That positive epistemic bias can be present in both positive and negative PQs, although in somewhat different flavors, is illustrated by the following pair of questions, which are formulated in English for simplicity; an analogous generalization applies to Russian and Czech (see also Section 2 for discussion).²⁰

If our reasoning is right, then epistemic bias was more salient for our participants than evidential bias. One reason why this might be so could be that while epistemic bias is private to the speaker and is being conveyed conventionally – possibly in the form of a conventional implicature and, thus, potentially as a new piece of information – evidential bias is based on the common ground of both discourse participants and, thus, has the status of a presupposition.²¹ While our design did not particularly favor the salience of one type of bias over another, it is likely that evidential bias relies on explicit contextual or situational cues to become salient. The absence of such cues in our design might have led to the increased activation of epistemic bias, which arguably does not require any contextual triggering.

We further hypothesized that Russian V1 questions – due to the presence of li – would align with the expectations drawn from Hamblin’s (1973) semantics. The results are consistent with this hypothesis, at least in the early periods in the positive V1 condition, where a significantly increased fixation on the positive picture (over the negative picture) steps in rather late, namely, in the object period. This is later than in Russian positive V2 questions as well as in Czech V1 questions, where the positive picture is fixated on already in the adverbial period.²² This is consistent with the expectation that li forces the participants to consider both alternatives for a longer time. The competing hypothesis that the preference for the positive picture, apparently characteristic of PQ processing, would show itself earlier in the Russian V1 condition could, therefore, not be confirmed. A similar delay is not attested in Russian negative V1 questions (which also involve li), where positive fixations step in very early – already in the subject period (persisting until the object period). This can be ascribed to a salient positive epistemic bias triggered by negation, which – by hypothesis – overrides the assumed effect of li. An anonymous reviewer rightly wonders why the effect of positive epistemic bias should be more pronounced in negative V1 questions than in their positive counterparts. We hypothesize that this is due to the different flavor of the epistemic bias in positive and negative questions: while the former induce what we called a “default” bias, corresponding to the speaker’s expectations or hopes, the latter induce a stronger type of bias, based on the speaker’s prior belief that the positive prejacent was true. Czech V1 questions exhibit a different pattern. Both positive and negative questions exhibit relatively early positive fixations (considering that the utterance is reliably disambiguated as a question only in the subject period), slightly earlier in positive questions (adverbial) than in negative questions (object). We do not know why the positive bias steps in slightly later in negative questions than in positive ones.

If we factor in the assumption that PQ processing corresponds to positive fixations (and set aside the interfering effect of li in Russian positive V1 questions), we see that the results are only partly consistent with our expectations about the timing of question interpretation. Russian negative V1 questions align well with the expectation: positive fixations step in immediately after processing the negative verb+li. Czech positive V1 questions are recognized as questions (i.e., instigate positive picture fixations) as soon as the VS combination has been uttered, as expected. This process appears to be somewhat delayed in negative V1 questions, for unclear reasons. Russian positive V2 questions also follow the expected pattern: positive fixations occur right after the pitch-accented verb has been processed. In Czech V2 questions, question interpretation was expected to become apparent only on the object, which realizes the final rising tone – a distinctive feature of Czech PQs. It is, therefore, unexpected that positive fixations occur already on the object in positive questions and on the adverbial in negative questions (in the latter case, even earlier than in V1 questions). We hypothesize that the intonation of declarative (V2) PQs in Czech is more complex than previously assumed (Romportl, 1951; Veroňková, 2002). In particular, it is likely that the question-related intonation contour is not confined just to the last/nuclear prosodic phrase of the utterance (here, the object), but spreads over a larger portion of the utterance, so that language users recognize an utterance as a question relatively early (see Baltazani et al., 2015, for evidence to this effect from Greek). Verifying this post-hoc hypothesis would require a proper acoustic investigation of Czech PQs (in comparison to assertions) – not just their final prosodic phrase (as in Veroňková, 2002), but the whole utterance.

It may seem implausible to claim a positive epistemic bias also for Czech positive V2 questions, since they superficially resemble English rising declaratives, which (in the positive variant) typically involve a positive evidential bias and a negative speaker expectation (Gunlogson, 2002).²³ However, Czech V2 questions differ from English rising declaratives in an important respect: They can contain not only inner, but also outer, negation, compatible with positive polarity items (Staňková, 2023). Moreover, Malamud & Stephenson (2015) argue that even English rising declaratives involve tentative commitments put up by the speaker and included in his expectations about further discourse (projected). Thus, epistemic biases of the default type envisaged here may well be at stake even in that subtype of Czech V2 questions which is closer to rising declaratives.

We now return to the negative V2 condition in Russian, which was the only one in which preference for fixation on the $p$ picture (or $\overline{p}$ picture, as a matter of fact) did not reach significance. The relatively balanced fixation profile in this condition might suggest a Hamblin-style interpretation of the question processed. In light of all our other results and the results of Tian et al. (2021), we consider this interpretation highly unlikely. One possibility is that the observed pattern is brought about by the expected meaning of the responses.²⁴ As is clear from Figure 12, most negative V2 questions yielded the selection of the negative picture. Therefore, the participant’s expectation of choosing the $\overline{p}$ picture, even before the response polarity was revealed, could have affected the fixations.

Figure 12: Picture choices in Russian questions.

The close-to-equal proportions of picture choices after the positive response can be connected to the interpretation of negation. Russian negative V2 PQs are generally ambiguous between a low-negation and a high-negation reading. According to González-Fuente et al. (2015) and Esipova (2021), the Russian response particle da ‘yes’ as a reaction to a negated proposition is mainly understood as agreeing with it, not as conveying the positive alternative.²⁵ This is consistent with low negation which is interpreted at the propositional level. In this case, da would confirm the negative proposition, leading to the selection of the negative picture. High negation, on the other hand, whether analyzed as operating at the level of speech acts (Krifka, 2015; Romero & Han, 2004, a.m.o.) or as a separate semantic operator (Šimík, to appear, a.m.o.; Repp & Geist, 2025), establishes no negative propositional antecedent (Krifka, 2013). In this case, da conveys the positive alternative (Geist & Repp, 2023), triggering the choice of the positive picture.

The reasoning about the ambiguity between low and high negation is comparable to that concerning French PQs in the Tian et al. (2021) study. Although a marginally significant shift of attention to the $p$ picture was also found for negative PQs in French, it was much reduced in comparison to the clear-cut preference for the $p$ picture in positive PQs. The authors hypothesize that the presence vs. absence of a positive epistemic bias may be responsible for the pattern observed in French PQs vs. English low-negation PQs. More research is needed to uncover the bias profile of Russian negated V2 questions and see whether similar considerations apply here.

6.1 Descriptive findings

Figures 13 and 14 illustrate how the proportions of fixations on $p$ and $\overline{p}$ evolved over time for assertions and V2 questions in Czech and Russian, respectively. The visualization suggests that assertions might exhibit a different pattern in comparison to questions. Whereas the picture preference while processing assertions seems to correspond to their polarity, questions show an overall tendency to fixations on $p$. An exception is formed by negative Russian V2 questions, already discussed in 5.5. Notably, questions in all the other conditions apparently display even a stronger preference for $p$ than the positive assertions.

Figure 13: Proportions of looks to $p$ and $\overline{p}$ in Czech assertions and V2 questions, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

Figure 14: Proportions of looks to $p$ and $\overline{p}$ in Russian assertions and V2 questions, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

6.2 Models

Two categorical predictors were included: polarity and force. Positive polarity and question (force) were coded as 1, negative polarity and assertion (force) as –1.

In this analysis as well, the starting point was the maximal reduced model.²⁶ Crucially, no random slope for force by item was included. This term was not supported by the data, as force was manipulated between items; all the items from Experiment 1 were assertions, and all items from Experiment 2 were questions.

6.3 Results

The mean log-ratio values plotted across interest periods for different conditions and languages are shown in Figure 15. These representations technically repeat those given in Figures 3 and 9 for assertions and V2 questions, respectively. However, in these visualizations, the conditions are combined differently, which allows for a comparison from another perspective. It is evident that assertions generally have lower log-ratio values compared to V2 questions.

Figure 15: Average progression of log-ratio of proportions of fixations on $p$ over $\overline{p}$ by interest period in Czech and Russian assertions and V2 questions. Error bars indicate 95% confidence intervals. Green dots show where log-ratio was significantly different from 0. Red dots show where no LME model could be fitted.

6.3.1 Czech

Both positive and negative V2 questions reach a peak where the preference for $p$ becomes significant. For assertions, on the other hand, no significant inclination could be found either for $p$ or for $\overline{p}$. This distinction is supported by the LME models that contained force and polarity as independent variables. A main effect of force was discovered in the object (t = 2.587, p = 0.014) and in the gap (t = 2.130, p = 0.04), higher values being predicted for questions than assertions. Moreover, an interaction of force and polarity was found in the first part of the response (t = –2.362, p = 0.023). While positive assertions instigated fixations on $p$ and negative assertions on $\overline{p}$, predicted values for both positive and negative questions are above 0; interestingly, negative questions had a stronger preference for $p$. Lastly, a main effect of polarity manifested itself in the gap (t = 3.070, p < 0.01). These effects are visualized by the model prediction plots in the three leftmost panels of Figure 16.

Figure 16: Log-ratio predicted by LME models in selected interest periods of Czech and Russian assertions and V2 questions.

6.4 Discussion

The comparison of V2 questions and assertions supports our interpretation of the results of the question experiment (see 5.5). We remind the reader that any differences between questions and assertions we observe here are due to their formal (prosodic) properties and not to their different discourse use: both types of utterances were presented in the very same discourse setting, namely, as lead-in utterances to be responded to by the addressee’s (dis)agreement. In both languages, we see that questions instigate more positive fixations than the corresponding assertions; in Czech, this difference is statistically significant, but in Russian, only a tendency.²⁸ This is particularly striking in Czech, where, judging by the descriptive findings, negative questions appear to lead to stronger positive fixations than the corresponding positive assertions.

These results underline what appears to be an inherent “positivity” of PQs. That PQs – whether positive or negative – convey positive speaker-related bias has been discussed repeatedly in the theoretical literature. Our results, together with those of Tian et al. (2021), suggest that this meaning component is particularly salient in the processing of PQs.

Finally, we would like to point out the relatively small degree of preference of looks to $p$ over $\overline{p}$ (or to $\overline{p}$ over $p$) in assertions (with the exception of the negative V2 condition in Russian; see 5.5 for some discussion). This preference rarely reached significance in assertions, while it did so relatively often in questions. In a sense, and very informally expressed, assertions behave more Hamblin-like than questions in our design: they instigate more balanced looks to $p$ vs. $\overline{p}$ than the corresponding questions. One possibility why this may be so is that assertions (as well as questions) in our design were always followed by ‘I think that yes/no’. The repeated exposure to this type of dialogue may have prompted the participants to expect a response and, therefore, interpret assertions as relatively neutral invitations to update the common ground, open to being challenged by the addressee’s perspective. Compared to that, PQ force is apparently not just a move that sets two alternatives side-by-side, but rather one that conventionally conveys positive bias.

7.1 Design and hypotheses

The Russian particle razve, licensed only in PQs, is dedicated to conveying/strengthening both epistemic and evidential bias, and, more specifically, the fact that they are in conflict. Razve occupies the clause-initial position and is followed by other constituents in a non-interrogative order (SVO, in our case); it is incompatible with V1 and li. Also for this reason, we only manipulate polarity in this experiment. It has been observed and argued (Korotkova, 2023; Restan, 1969) that a positive question with razve, such as (15a), conveys a positive evidential bias and negative epistemic bias. Negation reverses the biases, as in (15b).

Turning to Czech, although it possesses various particles which optionally occur in PQs, it does not have one that would correspond to the Russian razve straightforwardly. The question-initial particle copak (or cožpak) has been argued to have an analogous semantic effect (Nekula, 1996), but it belongs to a literary register, so we did not deem it suitable for our experimental setting. We instead decided to test the particle snad, which has been called “dubitative” in the Czech linguistic literature (Běličová, 1989; Štícha, 1995a, 1995b). Similarly to razve, it usually conveys a conflict of biases (Chodounská et al., 2025). It can occur in both interrogative (V1) and declarative questions, although the former is more frequent, which is why we use the verb-initial form in our design; see (16). Similar to razve and distinct from English really in PQs, Czech snad is compatible with both outer-negation and inner-negation readings.

We hypothesize that the conflict of biases, made salient by razve and snad, will manifest itself in the processing of PQs in that participants will pay attention to both pictures. We assume that the particles encode not only epistemic but also evidential bias conventionally, in contrast to the PQs without razve, where evidential bias is more context-dependent (see 5.5). It is, therefore, to be expected that in this experiment, we will not observe such a strong preference for the positive picture, since both contradicting biases will be reflected in the processing. In the absence of more detailed knowledge of how the epistemic and evidential biases are processed (thus far, we only seem to have observed the effect of epistemic bias), we do not attempt to formulate more specific hypotheses about the direction and timing of fixations. The experiment, thus, has an exploratory component.

7.2 Russian razve: Results

As can be observed in Figure 17, questions with razve exhibit a characteristic pattern, especially in the positive condition. Although, in general, looks to $p$ prevail, there are intervals during which the proportion of looks to $\overline{p}$ appears to be higher, based on the descriptive findings. Those intervals are a substantial part of the gap period as well as the second part of the reply. In the negation condition, the pattern appears more balanced, with a slight preference for $p$ in the object period and parts of the gap and the first part of the reply. Towards the second part of the reply, a mild tendency towards $\overline{p}$ seems to emerge. There is also an interval of apparently balanced looks in the middle of the gap.

Figure 17: Proportions of looks to $p$ and $\overline{p}$ in Russian questions with the particle razve, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

The visualization of the mean log-ratio progression across interest periods in Figure 18 suggests that the differences between the proportions of fixations to $p$ and $\overline{p}$ in razve questions are rather moderate. A significant preference for $p$ is only present in the first part of the response in the positive condition. The analysis comparing positive razve questions to negative ones revealed no significant effect of polarity in any interest period.

Figure 18: Average progression of log-ratio of proportions of fixations on $p$ over $\overline{p}$ by interest period in Russian V2 questions with particle razve. Error bars indicate 95% confidence intervals. Green dots show where log-ratio was significantly different from 0. Red dots show where no LME model could be fitted.

To test if razve questions differ reliably from simple V2 questions, a cross-experimental analysis was conducted on the relevant data set. LME models with polarity and particle as independent variables were fitted to each interest period, starting from the subject.²⁹ Positive polarity and razve particle were coded as 1, negative polarity and no particle as –1. The particle period in razve questions was removed from the analysis, since no corresponding interval was present in the simple V2 condition and no information on the target object has been revealed up to this period. The statistical analysis did not reveal significant effects for interest periods during the question and the gap.

Moreover, a post-hoc test was conducted to investigate the temporary preference for $\overline{p}$ in the gap after the positive razve questions, as noticed above. A LME model was fitted to compare this interest period in questions with and without razve inside the positive condition. Razve particle was coded as 1, no particle as –1. A significant effect of particle was found (t = –2.045, p = 0.0498). It might indicate a reliable distinction between simple positive V2 questions, which show a clear preference for $p$ during the gap (see, e.g., Figure 8) and positive razve questions, where an inclination towards $\overline{p}$ emerges in the same period.

7.3 Czech snad: Results

The data visualization in Figure 19 indicates patterns somewhat similar to those observed for Russian razve questions (see Figure 17). Also in Czech snad questions, there seem to be fluctuations of fixation proportions, especially in the negative condition. Short intervals of a more balanced distribution of looks can be seen at the beginning of the gap and in some parts of the reply. In positive snad questions, there might also be a slight tendency towards a similar proportion of looks to $p$ and $\overline{p}$ at the beginning of the gap. However, no preference for $\overline{p}$ comparable to the one found for Russian positive razve questions can be identified. A difference between the positive and the negative conditions seems to emerge during the response. In positive snad questions, the preference for $p$ apparently grows towards the end of the dialogue, while in the negative condition, the fixation proportions become more balanced.

Figure 19: Proportions of looks to $p$ and $\overline{p}$ in Czech questions with the particle snad, normalized for sound duration across conditions (IP boundaries also normalized) with geom_smooth.

The results of the dependent variable analysis, depicted in Figure 20, are in accordance with these observations. The majority of the mean log-ratio values are above 0, indicating more fixations on $p$ than on $\overline{p}$. However, in no combination of interest period and condition was this preference estimated as significant by the corresponding LME model. A statistical analysis comparing positive and negative snad questions revealed a main effect of polarity in the second part of the response (t = 2.194, p = 0.044). In the positive condition, the dependent variable was estimated to be significantly higher than in the negative one.

Figure 20: Average progression of log-ratio of proportions of fixations on $p$ over $\overline{p}$ by interest period in Czech V1 questions with particle snad. Error bars indicate 95% confidence intervals. Green dots show where log-ratio was significantly different from 0. Red dots show where no LME model could be fitted.

7.4 Discussion

Despite a lack of decisive statistical support (arguably due to the relatively low statistical power of the experiment), the pattern we observe in the processing of razve questions – particularly in the positive condition – can be regarded as distinct from those we found for other PQs. Although certain conditions without razve also exhibit short intervals of apparently balanced looks or even preference of $\overline{p}$ (see Figure 8), we argue that this tendency is more pronounced at least in positive razve questions. Potential evidence for this is the effect of particle in the post-hoc comparison of positive V2 PQs with and without razve. For ease of exposition, we will use the ad hoc label seesaw pattern to describe the observed alternations in picture preference.

We believe that the seesaw pattern can be considered a preliminary validation of our hypothesis that bias plays a central role in the processing of PQs. The manner in which the participants’ attention alternates between the positive and negative pictures is clearly systematic. If interpreted against the background of the corresponding questions without razve, we see significantly increased fixations on the negative picture in the positive condition of razve questions (and, correspondingly, decreased fixations on the positive picture). This can be understood as a reflection of the processing of the negative epistemic bias conveyed by the particle. This effect is preceded and followed by looks to the positive picture, suggesting an alternating consideration of the two conflicting biases.

Although we also observe a slight tendency toward the seesaw pattern in negative razve questions, the pattern there is much less prominent and less systematic. It is likely that the combination of razve and negation has led to a processing overload and, subsequently, to failure to yield a systematically alternating fixation pattern.

Albeit somewhat inconclusive, we consider the results of the razve experiment promising. Besides providing circumstantial support to our general claim that bias is essential for PQ processing, they also pave the way to a better understanding of how epistemic and evidential bias could be represented in the semantics of PQs and in an interlocutor’s mind.

The results of the snad experiment are qualitatively similar to those of the razve experiment: we found a rather balanced pattern of fixations. We also observe a tendency towards the seesaw pattern, which we assume reflects the processing of both types of biases – epistemic (contrasting with the polarity of the question) and evidential (in accord with the question polarity). The results of the snad experiment are not statistically robust, but, together with the razve experiment results, provide a promising point of departure for a deeper investigation of the processing of question bias.