Effects of goals and strategies on predictive processing: a visual world eye-tracking study on honorific agreement in Korean

5.1 Methods

5.1.2 Materials

The linguistic stimuli were 20 sets of items, each of which consisted of a critical sentence and a question, as illustrated in (6). These items were counterbalanced for the type of the progressive marker attached to the relative clause predicate in the critical sentence, with half of the items containing the honorific -kyeysi (honorific condition) and half containing the non-honorific -iss (non-honorific condition), yielding 10 tokens per condition (see Appendix A for a list of all experimental items).

(6)

Critical sentence
Ecey	tosekwan-eyse	kongpwu-lul
yesterday	library-at	study-acc

ha-ko kyeysi-ess-ten/ha-ko iss-ess-ten	palo	ce
do-prog hon -past-adn/do-prog-past-adn	very	that
sensayngnim-un	sengkyek-i	coh-ta.
teacher-top	character-nom	good-decl
‘The teacher who was studying at the library yesterday has a good character.’
Question

Nwu-ka	sengkyek-i	cohna-yo?
who-nom	character-nom	good-q
‘Who has a good character?’

Following the relative clause predicate, we inserted the determiner phrase palo ce ‘very that’ ahead of the target referent (e.g., sensayngnim ‘teacher’), which functioned as a padding phrase to allow for capturing participants’ predictive processing. Any evidence of participants’ preactivation of honorific information of the following target referent will be indicated by their increased looks to the target image compared to its competitor in the time window including this padding phrase. We thus analyzed this time frame as the critical region (see Data treatment and analysis below).

For the target referent, we selected characters that can receive honor and thus be modified by an honorific marker, including nouns denoting professions (e.g., sensayngnym ‘teacher’, kyoswunim ‘professor’, uysasensayngnim ‘doctor’, pyenhosanim ‘lawyer’, kamtoknim ‘director or supervisor’), kinship terms (e.g., hyengnim ‘older brother’, apenim ‘father’, emenim ‘mother’, halapeci ‘grandfather or male elder person’, halmeni ‘grandmother or female elder person’, samchon ‘uncle’), and nouns with a certain social standing (samonim ‘madam’, sacangnim ‘boss or sir’, acwumeni ‘madam or ma’am’, senpaynim ‘senior’, elusin ‘elder’). Because the target referent in the critical sentence was always associated with an honorific character, we included 10 fillers that were structured analogously to the critical sentences in the experimental stimuli but included a non-honorific predicate and a non-honorific referent, such as chengnyen ‘young man’, congepwen ‘employee or worker’, elini ‘child’, kkomaai ‘kid’, namcaai ‘boy’, namhaksayng ‘male student’, yehaksayng ‘female student’, cwunghaksayng ‘middle-school student’, yetaysayng ‘female college student’, and sinipsayng ‘freshman’.

Following the critical sentence, a question appeared asking participants to identify the referent mentioned in the preceding sentence by clicking on one of the characters on the display. Participants’ responses to the questions were taken to estimate their understanding of the critical sentence, and thus we excluded trials with incorrect responses (see results).

The experimental items were intermixed with 46 fillers, including those containing a non-honorific referent modified by a non-honorific predicate in contexts like (2) and sentences denoting various events where two characters interact with each other. The linguistic stimuli were recorded by a female native speaker of Korean at a natural speech rate using standard intonation. The audio recordings of the critical sentences in the experimental stimuli were edited using Praat (Boersma and Weenink 2017) to control for the length of the recording frames of interest across the honorific and non-honorific conditions. Consequently, the critical sentences between the two conditions did not significantly differ in the overall sentence length, the offset of the relative clause predicate, the onset of the padding phrase, and the onset of the target referent. Detailed descriptive and inferential statistics for these analyses are presented in Table 1.

Table 1:

Information about the acoustic duration (milliseconds) of relevant regions of interest.

	Honorific condition		Non-honorific condition		t	p	Cohen’s d
	M	SD	M	SD
Overall sentence length	12,022.0	464.0	11,906.5	503.3	0.755	0.455	0.239
Offset of the relative clause predicate	3,187.9	224.3	3,191.9	217.9	–0.058	0.954	–0.018
Onset of the padding phrase	4,550.3	2,668.9	4,544.5	230.6	0.073	0.942	0.023
Onset of the target referent	5,454.8	300.9	5,466.1	281.5	–0.123	0.902	–0.039

Each set of experimental stimuli in the honorific and non-honorific conditions was paired with a visual scene that contained two pictures of human faces with the words printed in Korean below (see Figure 1). One of the faces corresponded to the target referent (e.g., teacher) mentioned in the critical sentence, and the other face served as a competitor (e.g., student). The gender of the referents within each visual stimulus was kept identical to forestall any unexpected effects arising from gender stereotypes. The position of the target picture was counterbalanced across items.

Figure 1:

Example of a visual stimulus. The images represent a teacher (left) and a student (right) with the corresponding Korean words printed below.

5.2 Results

Participants’ fixations over the course of the critical time windows are presented in Figure 2 for the Prediction group and in Figure 3 for the Comprehension group. Outcomes of the mixed-effects models conducted for the three critical windows are presented in Table 3.

Figure 2:

The Prediction group’s proportion of looks to the competitor (red line) and the target (green line) in the honorific (left) and non-honorific (right) conditions in Experiment 1. Predictive region = W1 + W2; W1 = relative-clause boundary region (from the offset of the predicate in the relative clause to the onset of the padding phrase); W2 = main-clause region (from the onset of the padding phrase to target onset).

Figure 3:

The Comprehension group’s proportion of looks to the competitor (red line) and the target (green line) in the honorific (left) and non-honorific (right) conditions in Experiment 1. Predictive region = W1 + W2; W1 = relative-clause boundary region (from the offset of the predicate in the relative clause to the onset of the padding phrase); W2 = main-clause region (from the onset of the padding phrase to target onset).

In the analysis of the predictive region (W1 + W2, from the relative clause predicate offset to target onset), we found a main effect of Condition, with a significantly higher TargetBias score (i.e., an increased proportion of looks to the target referent relative to the competitor) in the honorific than in the non-honorific condition. While there was no significant effect of Group, we observed a significant interaction between Group and Condition. To further inspect this interaction, we conducted separate analyses for each group using linear mixed-effects models, with Condition (deviation-coded and centered around the mean) as a fixed factor and the random effects of participant and item. We applied an adjusted alpha level of 0.025 (0.05 divided by 2) for multiple comparisons. The results of the by-group analyses showed a significant effect of Condition for both the Prediction (b = –0.114, SE = 0.011, p < 0.001) and the Comprehension group (b = –0.045, SE = 0.015, p = 0.002). These results indicate that while both groups were more likely to look to the target referent following the honorific marker compared to the non-honorific marker in the predictive region, the magnitude of this effect was larger for the Prediction group than for the Comprehension group.

We obtained similar findings in the analysis of the relative-clause boundary region (W1, from the relative clause predicate offset to padding phrase onset). There was a significant effect of Condition, qualified by its interaction with Group. Subsequent by-group analyses showed that the effect of Condition was significant only in the Prediction group (b = –0.138, SE = 0.015, p < 0.001), but not in the Comprehension group (b = –0.005, SE = 0.018, p = 0.789). These results suggest that only the Prediction group exhibited predictive looks to target referents in this early time window.

In the following main-clause region (W2, from padding phrase onset to target onset), we observed a main effect of Condition, with a significant interaction with Group. By-group analyses revealed a significant effect of Condition for both the Prediction (b = –0.069, SE = 0.018, p < 0.001) and the Comprehension group (b = –0.111, SE = 0.023, p < 0.001). These results indicate that both groups engaged in predictive processing in this time window.

5.3 Discussion

In Experiment 1, two groups of participants, the Prediction and Comprehension groups, listened to Korean sentences involving honorific agreement under different task instructions. The analysis of the eye movement data in the global predictive region showed that both groups used the honorific marker to generate expectations about the honorific status of the following referent. However, the magnitude of prediction was greater for the Prediction group than for the Comprehension group. Moreover, in the more narrowly defined analysis regions, we observed a group difference in terms of the timing of predictive processing. Participants in the Prediction group demonstrated faster engagement in predictive processing, as they were quicker to detect the honorific information encoded in the honorific marker. These findings suggest that the participants’ predictive processing was not solely driven by unconscious automatic associative mechanisms but also modulated by top-down strategic control (Brothers et al. 2017; Huettig 2015).

However, caution is warranted in the interpretation of the current findings due to some potential methodological limitations. The evidence showing the Prediction group’s facilitated prediction might be due to the explicit task instruction provided to this group, which may have deliberately induced strategic processing more than is expected under other circumstances without explicit instructions. Although the participants who were asked to predict the target referent during the task showed the early signs of predictive processing, such an explicit instruction is hardly encountered in normal comprehension situations. This issue raises a question as to whether comprehenders’ predictive processing is still sensitive to top-down goals in contexts without explicit manipulation of task instructions. We attempted to address this concern in Experiment 2 where we manipulated the validity of predictive cues in linguistic stimuli instead of providing different types of instructions.

6.1 Methods

6.2 Results

Participants’ fixations over the course of the critical time windows are presented in Figure 4 for the Higher-validity group and in Figure 5 for the Lower-validity group. Model outcomes are presented in Table 4.

Figure 4:

The Higher-validity group’s proportion of looks to the competitor (red line) and the target (green line) in the honorific (left) and non-honorific (right) conditions in Experiment 2. Predictive region = W1 + W2; W1 = relative-clause boundary region (from the offset of the predicate in the relative clause to the onset of the padding phrase); W2 = main-clause region (from the onset of the padding phrase to target onset).

Figure 5:

The Lower-validity group’s proportion of looks to the competitor (red line) and the target (green line) in the honorific (left) and non-honorific (right) conditions in Experiment 2. Predictive region = W1 + W2; W1 = relative-clause boundary region (from the offset of the predicate in the relative clause to the onset of the padding phrase); W2 = main-clause region (from the onset of the padding phrase to target onset).

In the predictive region (W1 + W2), there was a main effect of Condition, with a significantly higher TargetBias score in the honorific than in the non-honorific condition, and a marginal interaction between Condition and Group. Separate analyses for each group revealed the significant effect of Condition in both the Higher-validity group (b = –0.070, SE = 0.011, p < 0.001) and the Lower-validity group (b = –0.036, SE = 0.015, p = 0.014) at the adjusted alpha level of 0.025, with a larger effect in the former group. These results suggest that while both groups predictively used the honorific cue, the magnitude of the prediction was greater in the High-validity than in the Lower-validity group in this overall prediction region.

Focusing on the narrowly defined regions, we found a significant interaction between Group and Condition, without a main effect of each factor, in the relative-clause boundary region (W1). This interaction was driven by the significant effect of Condition emerging in both groups yet pulled in the opposite direction. For the High-validity group, the TargetBias score was significantly higher in the honorific than the non-honorific condition (b = –0.047, SE = 0.014, p = 0.001). In contrast, the Lower-validity group had a significantly lower TargetBias score in the honorific than the non-honorific condition (b = 0.054, SE = 0.018, p = 0.003). These results indicate that only the Higher-validity group engaged in predictive processing in this time window. In the following main-clause region (W2), there was a main effect of Condition and its significant interaction with Group. By-group analyses showed a significant effect of Condition both for the High-validity group (b = –0.098, SE = 0.018, p < 0.001) and the Lower-validity group (b = –0.164, SE = 0.023, p < 0.001), suggesting both groups’ predictive processing, with the larger effect size in the Lower-validity group.

6.3 Discussion

In Experiment 2, we manipulated the overall validity of predictive cues in the stimuli by presenting participants with different types of fillers. The analyses of participants’ eye movements showed distinct processing patterns between the two groups. The Higher-validity group, who encountered implicit consequentiality sentences with 100 % bias-consistent endings, exhibited early signs of predictive use of honorific cues, whereas the effect of prediction was delayed for the Lower-validity group for whom the prediction effect emerged only in the later time window (i.e., main-clause region). Moreover, the magnitude of prediction in the entire critical time window (i.e., predictive region) was greater for the Higher-validity than the Lower-validity group. Taken together, these results provide confirming evidence of the modulation of comprehenders’ top-down goals in the speed and degree of predictive processing.

Unlike in Experiment 1, we manipulated predictive validity through fillers, and despite this implicit strategic manipulation, we obtained outcomes supporting the modulating role of top-down goals. Participants in the Higher-validity group, who were exposed to bias-consistent filler endings, likely recognized the utility of prediction as a resource-optimizing mechanism. The continuous confirmation of their initial predictions through bottom-up input (e.g., honorific cues aligning with referent status) may have enabled them to strategically enhance anticipatory processing. Conversely, the Lower-validity group faced frequent prediction failures due to bias-inconsistent filler endings, where the cognitive cost of revising initial expectations may have outweighed the benefits of prediction. This asymmetry in adaptive strategies highlights how comprehenders dynamically weigh the utility of prediction against its computational demands.

These behavioral patterns align with Kuperberg and Jaeger’s (2016) utility function framework, which posits that prediction is dynamically optimized based on the comprehender’s ongoing assessment of prediction validity under specific circumstances. This framework integrates the notion that prediction is a flexible cognitive strategy adjusted to maximize efficiency under varying environmental conditions (see also Kuperberg 2013; Lupyan and Clark 2015). More generally, the modulating effect of top-down goals observed here can be interpreted as part of a broader adaptive mechanism whereby listeners continuously update their mental models to “better reflect the broader statistical structure of the current environment” (Kuperberg and Jaeger 2016: 46). This process ensures that predictive mechanisms remain sensitive to changes in linguistic context, thereby optimizing comprehension efficiency.

We further investigated whether participants in the Higher-validity group adapted their predictive processing based on the linguistic cues presented in the experiment. To this end, we conducted additional exploratory analyses that included Trial number as an interactive fixed factor in the models for this group across the three time windows. The results revealed significant interactions between Condition and Trial number in all three regions of interest: W1 + W2 (b = 0.003, SE = 0.001, p < 0.001), W1 (b = 0.002, SE = 0.001, p = 0.004), and W2 (b = 0.003, SE = 0.001, p = 0.005). These findings indicate that as participants in the Higher-validity group encountered a greater number of items, they demonstrated enhanced predictive processing in the honorific condition. These results provide further compelling evidence of adaptive predictive processing in this group.

It is worth noting that the effect of prediction also emerged in the Lower-validity group, albeit delayed until the main-clause region. This suggests that this group did not completely shift away from their predictive strategy for processing the sentences involving honorific agreement despite the challenge of revising their initial expectations when processing the implicit consequentiality cues. One possible account for this outcome may be found in the nature of honorific agreement in Korean. Honorific agreement is a ubiquitous feature in Korean (Brown 2015; Sohn 1999), allowing Korean speakers numerous opportunities to practice honorific agreement in various situations, particularly in the social milieu that emphasizes the expression of social superiority in language use (Kim and Sells 2007). The prevalent and preponderant occurrences of Korean honorification thus may have promoted the learning and entrenchment of this knowledge in the speaker’s mind, rendering honorific marking a highly informative precursor to the status of an upcoming referent in the current experiment. However, this account is largely speculative, and it remains to be determined whether the facilitative prediction witnessed among our participants is specific to the processing of honorific agreement or can be generalized to other linguistic features. To address this question, future research needs to test the effect of top-down factors in prediction using diverse linguistic phenomena.

7.1 Summary of findings

The purpose of this study was to assess the effect of top-down goals and strategies in Korean speakers’ predictive processing of Korean honorific agreement. Across two visual-world eye-tracking experiments, we found that top-down strategies and predictive validity modulated the timing and degree of comprehenders’ predictive processing. In Experiment 1, participants who received explicit task instruction for prediction were faster in the use of honorific cues for generating prediction than those who processed sentences simply for comprehension. In Experiment 2, an earlier and stronger effect of prediction was observed among the group who constantly encountered linguistic stimuli that promoted prediction, compared to the group who listened to stimuli with low prediction validity. These findings suggest that comprehenders engage in strategic predictive processing that depends on the current goals and the validity of predictive cues in the input (Huettig 2015; Kahneman 2011).

7.2 Theoretical implications for predictive processing

The robust effect of the top-down factors observed in this study is consistent with the findings from Brothers et al. (2017), which demonstrated the modulation of task instruction and prediction validity in English speakers’ lexical prediction. The present study extended this research by showing that the strategic manipulations also had measurable impacts on the Korean speakers’ predictive eye movements during their processing of honorific agreement in Korean. In addition, our findings go further in demonstrating that the top-down factors influenced comprehenders’ processing of honorific agreement not just in the magnitude of prediction but also in terms of the timing of prediction, as indicated by the early effect of prediction that emerged when prediction was promoted. Our study thus provides additional empirical support for the assertion that predictive pre-activation is a general processing mechanism subject to top-down goals and strategies (Huettig 2015; Kuperberg and Jaeger 2016).

As a reviewer pointed out, a question arises as to whether the manipulation employed in Experiment 2 reflects a genuinely context-driven adaptation or whether it imposes a form of top-down guidance. While our use of biased fillers was intended to implicitly encourage participants to adjust their predictive strategies over time, we acknowledge that this manipulation may also be viewed as introducing a structured influence that shapes predictive behavior. However, as Brothers et al. (2017) suggest, such context-sensitive adjustments reflect the flexible deployment of anticipatory strategies without requiring modifications to the underlying architecture of the comprehension system. In other words, as comprehenders repeatedly encounter prediction errors through bottom-up input, they strategically reduce reliance on predictive strategies due to the associated processing costs. Although the experimental setting does not fully replicate naturalistic communication, it plausibly approximates how comprehenders adapt to varying predictive environments, where comprehenders implicitly optimize their predictive behavior in response to the output of a speaker. Consistent with this perspective, we interpret the adaptation observed in our study not as an artificial effect imposed by top-down instruction, but rather as a reflection of the natural sensitivity of the language processing system to distributional regularities in the input. In this sense, participants’ predictive behavior in Experiment 2 likely emerged from their ongoing evaluation of the utility of prediction in the given context, closely mirroring the adaptive nature of prediction in everyday language comprehension.

The variable nature of predictive processing depending on the particular processing requirements is consistent with the view of prediction as a utility function, according to which comprehenders adjust the degree of prediction in response to the current need for optimizing cognitive resources for efficient comprehension (Kuperberg and Jaeger 2016). The utility of predictive processing is characterized as the resource-bound comprehender’s effortful behavior to “maximise the utility of her predictions and rationally allocate resources” (Kuperberg and Jaeger 2016: 44). This utility function has been identified as a crucial feature of the human processing mechanism. Studies have shown that the speed and accuracy of comprehenders’ speech perception and word recognition are influenced by specific task goals and statistical regularities of co-occurrence between linguistic items (e.g., Bicknell and Levy 2012; Hayhoe and Ballard 2005; Kleinschmidt and Jaeger 2015). In keeping with the view that extends this processing mechanism to the domain of prediction (Kuperberg and Jaeger 2016), our participants exhibited substantial variability in the timing and extent of prediction depending on the task goals and strategies, reflecting their flexible adaptation of prediction to the requirements of current processing environments. Our findings thus suggest that predictive processing as a utility function applies to Korean speakers’ predictive use of morphosyntactic and semantic features during their processing of honorific agreement (see Fine et al. 2013, for adaption of prediction during syntactic processing).

The present findings also support Huettig’s (2015) account of the multiple pathways of prediction whereby comprehenders engage in predictive processing via an interconnected system that incorporates both automatic mechanisms and conscious strategic mechanisms. According to this account, comprehenders unconsciously pre-activate linguistic information of upcoming materials through simple associative mechanisms, but at the same time, when there arise demands for allocation of memory resources to optimize processing efficiency, they often engage in strategic, heuristic prediction through the combinatorial system that includes the active reasoning mechanism sensitive to multiple linguistic and nonlinguistic constraints. Huettig (2015) proposed that the mechanism of heuristic processing is interconnected with other prediction systems such as the production-based and simulation-based prediction, constantly influencing each other. For instance, automatic prediction through the simple spreading activation of items may add to the effect of strategic prediction. Consistent with this account, our participants showed predictive behaviors even without external factors promoting prediction under a normal comprehension situation (i.e., the Comprehension group in Experiment 1) and even with linguistic cues biasing against their initial prediction (i.e., the Lower-validity group in Experiment 2), an indication of the automaticity of their predictive processing. Concurrently, the magnitude and speed of prediction scaled up when participants engaged in strategic prediction in response to the task goal and prediction validity of linguistic cues. While our findings suggest some possible association between these two systems, further research is required to investigate the interconnectivity of the multiple predictive processing mechanisms more thoroughly by examining how different mechanisms operate in concert or in conflict to influence the way that comprehenders engage in prediction under diverse processing contexts (e.g., Hintz et al. 2017).

7.3 Limitations and future directions

We acknowledge several limitations of our study. One of the primary limitations is the relatively small sample sizes. The limited number of participants may have resulted in reduced statistical power, which can impact the generalizability of our findings. Each experiment comprised only 32 participants divided into two groups. Despite the limited number of participants, we were able to observe significant effects resulting from our manipulations (instruction manipulation in Experiment 1 and predictive validity manipulation in Experiment 2). This might be because our study had a larger number of items per condition (k = 10) compared to typical psycholinguistic research. This resulted in an overall sample size per condition comparable to previous studies with larger participant pools but fewer items per experimental condition (e.g., k = 5). Still, a larger sample size would have provided greater statistical power and increased the reliability of our findings. Future research should address this problem by involving larger sample sizes to obtain more reliable and generalizable results.

Another limitation, as pointed out by an anonymous reviewer, is the low proportion of critical items (k = 20) to filler items (k = 70), along with the overall large number of stimuli (k = 90), in Experiment 2. We admit that this design choice might have weakened our experimental manipulation, and that the resulting large number of stimuli may have led to attentional problems or fatigue among participants. In further research, we will take this factor into account in our design.

Furthermore, we only included object-biased implicit consequentiality verbs in Experiment 2. A null referent in sentences like (4) is preferentially associated with the subject in the preceding clause in Korean (Hwang 2023), which introduces a confound to the object bias induced by the verb semantics. In order to disentangle the effects of these factors, it is important to include both subject- and object-biased verbs in future studies.

We also acknowledge that our models did not provide a good fit to the data. In both experiments, the models for each region of interest demonstrated only a marginal fit, suggesting that they may not have accurately captured the processing patterns among participants. This issue could have contributed to some unexpected patterns observed among participants in both experiments. For example, in Experiment 1, the Comprehension group showed reduced fixation on the target even after it was mentioned, which may be attributed to increased processing difficulty with the integration of information towards the end of the sentence. However, it is important to note that this unexpected pattern could also be influenced by the limitations of the model fit. Similarly, in Experiment 2, the Lower-validity group looked at target objects more often in the non-honorific than the honorific condition in W1, which may be due to the limitations of the model fit. To address these limitations, future studies should increase the sample size and the number of items, allowing for a larger number of data points that produce more robust models with a good fit to the observed data.