2.1. Participants

Sixty children, from a monolingual kindergarten in Zhengzhou, China, were invited to participate in the experiment. Written parental consent was obtained prior to the experiment. All the participants were screened to exclude health and intellectual impairments and speak mandarin Chinese as their mother tongue. The children were assigned to three age groups (See Table 3). All the participants managed to complete the trials; however, data from two participants in group 3 were excluded from the final analysis due to background noise and interruptions.

Table 3. Participants

Prior to the experiment, a pilot study was conducted with two three-year-old children to make sure that children at this age and above can understand the words uttered in the recorded stories and are able to recognise and name the items that appear in the pictures.

2.2. Stimuli categorisation standard

Stimuli in the present study are initially classified based on the degree of novelty into two distinct categories: novel metonymy and conventional metonymy.

A novel metonymy (e.g., ‘the ham sandwich left without paying’) is constructed out of an accidental connection between concepts that have no established or stable metonymic reading yet (Schumacher et al., Reference Schumacher, Weiland-Breckle, Reul and Brilmayer2023; Zheng et al., Reference Zheng, Jia and Liang2015). For novel metonymy, although most types of those metonymy construction (e.g., body-part for person; property for person) are universal across language and culture (Slabakova et al., Reference Slabakova, Cabrelli Amaro and Kang2013), the intended reading of specific metonymy is largely contingent upon contextual information. In other words, the link between the “ham sandwich” and the person is salient but ephemeral, susceptible to contextual changes (e.g., “I want the ham sandwich” where the “ham sandwich” refers to the meal set with ham sandwich as the main course).

Conversely, conventional metonymy (e.g., “Dickens” for Dickens’ works; “maoyeye”/“Chairman Mao” for banknotes/money in mandarin Chinese contexts) has undergone a process of conventionalisation and has established a stable association between concepts in commonly used language. The figurative meaning of conventional metonymy is experience-based and is largely retrieved from linguistic and world knowledge (e.g., long-term social and cultural memory), but not as largely cognitive-based as novel metonymy, as predictors such as semantic linguistic ability, compared with cognitive-based measures, are found to better predict participants’ performance and acquisition of conventional metonymy (Vicente & Falkum, Reference Vicente and Falkum2023; Zheng et al., Reference Zheng, Jia and Liang2015).

The conventionality and novelty of metonymy have been measured using various standards and are found to be related to acceptability and effort demanded for processing, with conventional metonymy being less demanding than novel metonymy. Slabakova et al. (Reference Slabakova, Cabrelli Amaro and Kang2013) distinguished conventional (regular) from novel metonymy based on the “noteworthiness and stability” of the link between concepts and rated place-for-event (e.g., “Chernobyl” for the nuclear accident) and producer-for-product (e.g., “Dickens” for Dickens’ works) metonymy as more regular (conventional) than instrument-for-player (e.g., “clarinet” for clarinet player) and loose association metonymy (e.g., the ham sandwich example). Frisson and Pickering (Reference Frisson and Pickering2007) and Schumacher et al. (Reference Schumacher, Weiland-Breckle, Reul and Brilmayer2023) proposed familiarity as a key proxy for determining the conventionality and thus the processing effort required for metonymy. For example, within the construction of producer-for-product metonymy, “I never read Dickens” would be more conventionally metonymic and more readily accessible than “I never read Needham” where “Needham” is not a familiar name for an author (Frisson & Pickering, Reference Frisson and Pickering2007).

To take stock, there is more of a continuum than of a dichotomy from novelty to conventionality. Highly novel metonymy, found towards one end of the continuum, typically features a contextually contingent (unstable) link, lacks idiomaticity or lexicalisation, and lacks central representativeness.Footnote ¹ Moving towards the conventional end, metonymies exhibit figurative readings that are stabler across contexts, especially, for example, the language- or culture-specific idiomatic metonymy constructions (Nunberg, Reference Nunberg1995) (e.g., “one mouth” in a family referring to a member who lives and consumes food in the family).

As “stability of metonymy link” (Slabakova et al., Reference Slabakova, Cabrelli Amaro and Kang2013) is difficult to measure, the present study distinguishes between novel and conventional metonymy by labelling the idiomaticity of metonymy and checking the frequency of metonymy readings of expressions in the CCL (Center for Chinese Linguistics Peking University) corpus (See Table 4). For example, for conventional metonymy, we include idiomatic and cultural-specific metonymy expressions in the Mandarin Chinese context and expressions that are frequently taken metonymically.

Table 4. Frequency of metonymical/literal readings of conventional metonymy in CCL

Taking one step forward, the present study further subcategorizes novel metonymy according to Peirsman and Geeraerts (Reference Peirsman and Geeraerts2006), who provided the prototypical classification model for metonymy, based on their conceptualisation of the relation of contiguity. In their model, the strength of contact (the distance between the salient property and the referred target), boundedness (whether there is an ontological border between the salient property and the referred target), and domain type (spatial, temporal or categorical) together form the metonymy category, where a rich set of metonymies are categorised and linked in terms of the type of contiguity they are motivated by. Refer to Peirsman and Geeraerts’ (Reference Peirsman and Geeraerts2006) model in Figure 1:

Figure 1. Prototypical category of metonymical patterns. (Peirsman & Geeraerts, Reference Peirsman and Geeraerts2006, p. 310)

In Peirsman and Geeraerts’ (Reference Peirsman and Geeraerts2006) model, there are three dimensions indicating three standards by which various metonymies can be linked with the core category, spatial part-whole metonymy. In the utterance “the big beard walked towards him,” the “big beard” itself is a physical part of and spatially linked to the person who wears it.

Along each dimension line, the prototypicality of metonymy becomes weaker. In a sense, various types of metonymy are able to find their places on the proposed three-dimension model, which classifies them by their degrees of prototypicality.

Therefore, the stimuli of this study consist of novel and conventional metonymy; and within the novel type, metonymy stimuli are further categorised into three subtypes based on children’s animacy preference (Piaget, Reference Piaget1964) and Peirsman and Geeraerts’ (Reference Peirsman and Geeraerts2006) model. As we hypothesise that (1) human-related metonymies are easier for children and that (2) prototypical metonymies are easier than non-prototypical ones in metonymy-motivating contexts, the hypothesised relative difficulty of novel metonymy subtypes is: human-related part-whole (e.g., big ears – the person with big ears) < human-related adjacency (e.g., white trousers – the person wearing white trousers) < non-human-related feature (e.g., watermelon – a cup of red juice made from watermelon).

2.3. Stimuli and procedure

The fourteen trials in this experiment consist of one training trial, ten trials with metonymy target utterances, two trials with literal target utterances, and one distracting trial.

The first training trial helps children to familiarise themselves with the experimental procedure. Data collected in this trial are excluded from the final analysis.

The ten metonymy trials are divided into two categories (novel and conventional), with four subcategories (1 sub-category under conventional, 3 sub-categories under novel) in total. Among the 10 trials, 4 trials contain conventional metonymy stimuli, while 6 contain novel stimuli (2 trials under each sub-category).

The two literal trials serve to measure children’s ability to understand the literal meaning of the stories. As it is assumed that three-year-olds should grasp the meaning of these items, they can also signal whether children might select a picture at chance, especially for three-year-olds. The distractor trial presents an utterance that makes no sense. It is put in the middle of the trials to avoid the learning effect.

All the 14 trials are identical in structure and similar in length. In each trial, the participant listens to a pre-recorded story lasting for about 20 seconds, while looking at a set of four pictures (as in Figure 2 below). The background picture on the top of the screen gives participants a general idea of “what” with regards to “whom”. The three option pictures in metonymy trials contain one metonymy picture (the right answer, i.e., protagonist with big ears), one literal picture (the object, i.e., the “big ear” itself), and the irrelevant picture (the other person from the background picture). Each story is composed of one context utterance and one target utterance. Each target utterance contains a referential expression of the metonymy target (“big ear”) and a descriptive expression (“always gets criticized by teachers…”) which helps with the question that follows.

Figure 2. Example stimulus

Following the content of the story, a question for instruction unfolds. The questions in all the trials would use “shenme” (which) instead of “shei” (who) at the beginning, to avoid causing biases to participants’ choices. One example of the whole 20-second audio recording in one trial is shown below.

Example 1:

Then, one follow-up question (“Could you tell me why you chose this?”) is asked to collect data concerning the comprehension process of the participant. Children’s feedback is recorded and transcribed for analysis of the explanation task.

2.4. Coding

The participants’ answers to the forced-choice questions are treated as a categorical variable. All correct selections are coded as “1;” all incorrect selections are coded as “0”.

Concerning the “explanation score” in metonymy trials, children’s explanations of their correct selections fall into three categories (refer to Table 5 and Example 1):

Table 5. Three categories of children’s explanations

To check inter-rater agreement, the two authors coded 100 explanations randomly selected from the children’s explanations. They were unaware of the child’s age when coding the answers (к = .88). According to Landis and Koch (Reference Landis and Koch1977), a к value above .81 indicates almost perfect agreement.

Concerning the “reaction time,” it starts counting once the target expression (e.g., “big ear”) unfolds and stops when the participant presses the button to give feedback. To guarantee the validity and accuracy of the timing, the time span of audio recording (8000 ms) from the target expression to the end of the question is strictly controlled, which means all the figures taken down should be greater than 8000ms, except for cases of false triggering or rushing to answer. Also, data analysis only contains the reaction time from metonymy trials where participants select the correct pictures.

2.5. Results

Data analyses were conducted using the SPSS version 26 statistical package (SPSS 26.0). We used three performance measures in the study: selection accuracy, explanation score, and reaction time. We ran generalised linear mixed models (GLMMs) on metonymy trial accuracy and explanation score, and one-way ANOVAs on literal trial accuracy and reaction time, including age (3 age groups) as the between-subject variable and condition (metonymy types) as the within-subject variable. All models were fitted with fixed effects structure and random effects structure including by-subject and by-item random intercepts. We report relevant F and p-values below, as well as standard errors (SE) and t-statistics for pairwise comparison tests.

The data were analysed as follows.

First, we examined metonymy trial accuracy for each age group in the selection task. Then, in order to avoid the possible over-estimation of metonymy trial performance, three more analyses were carried out: participants’ performance in literal trials; times of unrelated choices in metonymy trials; and participants’ reaction time, which is a relative measuring of time for processing the target metonymy expression.

Second, we investigated the explanation score for each age group in the selection task. To help with solving and explaining the discrepancy between metonymy trial accuracy and explanation score, an ANOVA analysis of explanation-to-choice ratio was conducted.

Third, the within-subjects variable, metonymy type, was added to the analysis. As the most revealing measure between different age groups, the explanation score was chosen to be the measure in the analysis of the influence of metonymy types on children’s metonymy comprehension.

2.5.2 Explanation task

Table 6 and Figure 5 present the data of explanation score (full score = 20) in different age groups. A GLMM analysis on explanation score has been done. We were left with a model including the main effect of Age (as categorical variable).

Table 6. Mean accuracy/score in different types of trials in 3 age groups

Table 7. Average reaction time in 3 age groups

The overall model thus reveals significant differences in explanation score in Age, F(2,577) = 24.248, p < .001. A pairwise test reveals that this effect is due to significant differences between age-3 group and age 4–5 group on the one hand (SE = .10, t = -3.87, p < .001), and between age 4–5 group and age 6–7 group on the other (SE = .10, t = -3.78, p < .001), which forms a discrepancy with the selection results.

To help with solving and explaining the discrepancy between metonymy trial selection accuracy and explanation score, an ANOVA analysis of explanation-to-choice ratio was conducted, with selection accuracy temporarily transformed from categorical data to numerical data. The explanation-to-choice ratio can be revealing for the following reasons: Ideally, if a participant gets all the trials correct in both tasks, the ratio would be 2 (20/10); however, in real situations, if the ratio calculated to be greater than 2, it would mean that there must be cases where the participant selects the incorrect picture in the selection task while gives the correct response in the explanation task, which would suggest that the child might actually understand the metonymy but just does not choose the correct picture for certain reasons; if the ratio is lower than 1, it would mean that there must be cases where the participant selects the correct picture while gives no response or zero-point response in the explanation task, which would suggest that the child might be unable to explain his or her choice or that the child chooses the right answer only by chance.

The boxplot above presents the explanation-to-choice score in different age groups. The mean ratio grows as age increases, but the difference is only significant between two younger groups (F = 16.669, p < .001). As is noteworthy, the ratio of the youngest group is below 1 (the median), which suggests that it might be a common practice for the three-year-old participants to gain points in selection tasks by selecting pictures at chance. Among the data of age 4–5 group, there are figures greater than 2, which suggests that there might be factors masking these children’s metonymy comprehension ability and preventing them from choosing the metonymy option they actually understand in the selection task.

2.5.3. Influence of metonymy type

Since literal trial accuracy, times of unrelated choices, reaction time, and explanation-to-choice ratio jointly suggested that explanation score was the most reliable measure between different age groups, it was chosen in the analysis of the influence of metonymy types on children’s metonymy comprehension. We conducted two analyses of GLMM on explanation score, one for 2-type categorisation of difficulty and one for 4-type categorisation of difficulty. In each analysis, Age was the between-subject variable, and Condition (metonymy types) was the within-subject variable.

Our first analysis with Age (3 age groups) and Condition (novel, conventional) yields a main effect of Age, such that the explanation score increases with age, F(2,576) = 22.173, p < .001; and a main effect of Condition, such that there is a clear drop in explanation score from novel metonymies to conventional metonymies, F(1,576) = 25.190, p < .001 (See Table 8).

Table 8. Explanation score of different metonymy types in three age groups

To have a direct combining view of the influence from age and the influence from metonymy type, a pairwise test was conducted concerning the explanation score between different age groups.

In terms of novel metonymy, the explanation score difference by age is more significant between the two younger groups (SE = .14, t = -3.86, p < .001) than between the two older groups (SE = .13, t = -2.73, p < .05); while in terms of conventional metonymy, the difference is more significant between the two older groups (SE = .13, t = -2.33, p = .021 < .05) than between the two younger groups (SE = .13, t = -2.08, p = .039 < .05).

In the second analysis, metonymy stimuli were divided into four types according to different pre-set difficulty levels. Our final model included as fixed effects Age (3 age groups) and Condition (4 metonymy types) and as random effects the random intercepts for subjects.

The overall model thus reveals differences in explanation score in Age, F(2,574) = 22.173, p < .001, and Condition, F(3,574) = 18.711, p < .001. The interaction of Age by Condition is non-significant, F(6,574) = 1.048, p = .393, which indicates that children of different age groups exhibit similar developmental trajectory in comprehension of four types of metonymy. Explanation scores are calculated for each participant and the mean scores of the four metonymy types are shown in Table 9.

Table 9. Explanation score of different metonymy types in three age groups

At this stage, the previously hypothesised relative difficulty of novel metonymy subtypes turns out to be wrong. These results suggest that the difficulty ranking of the 4 metonymy types should be: novel non-human-related feature metonymy (type 3) < novel human-related adjacency metonymy (type 2) < novel human-related part-hole metonymy (type 1) < conventional metonymy (type 4). Pairwise comparison reveals that the main effect of Condition is due to significant differences between type 1 and 3, type 2 and 3, type 2 and 4, type 3 and 4; p < .001.

3.1. Participants

Thirty-six children ranged from 3 to 7 years old were tested in this experiment (See Table 10). All participants were native mandarin Chinese speakers screened to exclude health, vision and intellectual impairments. Children were recruited from kindergartens and primary schools in Songjiang District, Shanghai. Written parental consent was obtained prior to the experiment.

Table 10. Participants

3.2. Stimuli and procedure

The stimuli contain twelve trails, the first two being training trials. The rest consist of 7 metonymy trials, 2 literal trials, and 1 distracting trial. The reason why the metonymy trials (7 trials) in eye tracking experiment are fewer than those (10 trials) in the picture selection experiment is that, on the one hand, child participants of the eye tracking experiment need to complete the task alone in the lab and thus need more training trials; on the other hand, children doing the task alone easily get distracted or feel bored, so, the time span of the task needs to be shorter. To ensure the comparability of the data from two experiments, we make sure that the 7 metonymy trials in the eye tracking experiment come from the 10 metonymy trials in the picture selection experiment.

Similar to the picture selection experiment, the seven metonymy trials are divided into two categories, novel metonymy (4 trials) and conventional metonymy (3 trials).

In each trial, the stimuli contain four picturesFootnote ² placed on four areas of the screen (See Figure 7). The largest picture on the upper half of the screen is the background picture; the three pictures on the lower half are option pictures. The participants are asked to choose one, by pointing with finger, according to the 20-second audio story.

3.3. Coding

Proportion of fixation is calculated by the number of fixations on target areas divided by the number of all fixations that take place during the selected period of time. The proportions of fixation before and after the target utterance are calculated separately so that they can form a comparison which may show the changes in participants’ fixation patterns triggered by the target utterance.

3.4. Results

At the first step, GLMM was carried out between the three age groups concerning the proportion of fixation on metonymy target areas after the target utterance. The fixation proportion after the target utterance can directly reflect children’s online processing of the target metonymy.

Following this, the variable of metonymy type was entered into the analysis. Two GLMMs were done in each age group concerning the change of fixation proportion before and after the target utterance in two metonymy types (novel vs. conventional). The difference between pre-target fixation proportion and post-target proportion shows how the target utterance works to trigger the changes in participants’ pattern of looks. If there is not an increase in the fixation proportion after the target utterance, it is reasonable to say that the participant’s eye movement is not clearly triggered by the utterance, and thus he or she might not understand the metonymic meaning the utterance carries.

The final analysis took a more detailed look at the data in seven individual trials so as to better explore how metonymy type influences children’s fixation pattern. One-way ANOVA analysis was conducted in each age group concerning the comparison between pre-target and post-target fixation proportion in different trials.

3.4.1. Fixation proportion between age groups

In each trial, as the three option pictures on the screen are equal in area and the background picture is 3 times as large as the option picture, the fixation proportion of random looks should be the ratio of the area of one option picture to the area of all pictures, which is around 16% (See Table 11).

Table 11. Fixation proportion before and after target expression in three age groups

TE: Target expression

No main effect of Age is observed in the fixation proportions before target utterance, F(2,249) = 2.198, p = .113. The means are around 16% in all age groups, which indicates that the participants scanned the pictures on the screen in a random manner before the target utterance unfolded.

However, significant group differences are found in the fixation proportions after target utterance, F(2,249) = 6.029, p < .05. The means in all age groups are higher than 16%, which suggests that all groups show a metonymy bias triggered by the metonymy target utterance. Pairwise comparison reveals that the difference between age groups is statistically significant between the two younger groups (SE = .02, t = -2.67, p < .05), and even greater between the youngest group and the oldest group (SE = .02, t = -3.38, p < .001). As for the two older groups, the difference is not significant (SE = .02, t = -.74, p = .458).

3.4.2. Influence of metonymy type

In this section, we focus on the difference between pre-target fixation proportion and post-target proportion. If the difference is positive (post-target > pre-target) and significant, it suggests that children’s fixation is stimulated by their understanding of the target expression.

Two analyses of GLMM were done, one for novel metonymy and one for conventional metonymy. In each analysis, Age was the between-subject variable. No main effect of Age is found in novel metonymy, F(2,141) = 1.101, p = .335; while in conventional metonymy, age difference is significant, F(2,105) = 3.347, p < .05. Pairwise comparison reveals that the main effect of Age in conventional metonymy comprehension is due to significant differences between the youngest group and the oldest group (SE = .05, t = -2.53, p < .05) and marginally significant differences between the two youngest groups (SE = .05, t = -1.95, p = .054).

However, to better look at how metonymy type influences children’s fixation pattern, we need to take a closer look at the data in different trials (See Table 12).

Table 12. P-value in all trials in three age groups

* Mean Difference = mean post-target proportion − mean pre-target proportion

* The underlined figures are those greater than 0.05 (based on the threshold p<0.05 for significance)

In age-3 group, participants show an increase of fixation proportion after target utterance in all of the four novel metonymy trials and the increases are statistically significant in three of the four trials. On the other hand, in all the conventional metonymy trials, the fixation pattern of age-3 participants does not show a significant difference between pre-target and post-target fixation proportion.

In age 4–5 group, participants show an even clearer and more significant increase of fixation proportion in all the novel metonymy trials. In terms of conventional metonymy trials, the fixation pattern of participants at this age begins to show significant increases in two of the three trials.

In age 6–7 group, participants show significant increase in both types of trials. Compared with the two younger groups, participants of the age 6–7 group show a clear metonymy bias when faced with tasks of conventional metonymy in all the three trials, which might suggest that it is at the age of six that children really begin to show a good command of conventional metonymies.

In general, in novel metonymy trials, although all the three age groups show significant or marginally significant increase in fixation proportion after target expression, there is a difference in the degree of increase between different age groups: the increase is clearer between the two younger groups. On the one hand, the age-3 participants still have problems dealing with novel metonymy stimuli, receiving less triggering influence from target utterances, and thus showing smaller mean difference of fixation proportion and greater p-values; on the other, the age 4–5 children show a clearer increase in all the four trials, which suggests their good comprehension of novel metonymy stimuli, and their fixation pattern is similar to that of the oldest group.

As for the conventional metonymy trials, the clearer increase occurs between the two older groups. The fixation pattern of the age 4–5 participants shows that they still have difficulty comprehending some conventional metonymies, as the youngest group does in all the three trials; while the age 6–7 group’s fixation proportion shows that they have a good command of understanding conventional metonymies.

4.1. Summary of results

In terms of children’s developmental trajectory of metonymy comprehension, the three tasks yield different but mutually explainable results. Our results, regardless of the influence of metonymy type, do not show a U-shape in any clear form or shape as some previous studies did, but show that (1) age-3 children’s performance in the selection task is more random than the other two age groups, selecting unrelated choices and making mistakes even in literal trials; (2) age 4–5 participants’ selection accuracy is not significantly higher than their age-3 counterparts but (3) their performance in the explanation task and eye tracking task is significantly better; (4) in terms of reaction time, age 4–5 children need the most time for processing before choosing a picture in the selection task, which forms a slight U-shaped developmental trend; (5) children older than 6 years perform significantly better under almost all the measurements, compared with the other two groups.

The present study takes into account not only the factor of age but also the possible influence from metonymy type on children’s acquisition of metonymy, as most of the extant empirical studies on metonymy development only focus on the time course (age) of development, using novel metonymy as stimuli. Experimental stimuli in the present study consist of novel metonymy and conventional metonymy, and novel type is further divided into three sub-categories with different difficulty levels. However, in terms of the influence of metonymy type, our results of the three-sub-type metonymy categorisation turn out to go against the hypothesis, as children’s performance in the offline experiment (selection task and explanation task) does not decline as difficulty increases.

We then merged novel sub-types and compared novel metonymy with conventional metonymy, adopting a two-type categorisation. The results from both experiments show that children begin to understand conventional metonymy significantly later than they show a decent comprehension capacity with novel metonymy, and the results from explanation task and eye tracking task show that it is only after six years old that children show a good command of conventional metonymy comprehension.

In combination, the three tasks in the two experiments yield different results in terms of age trajectory, but similar and mutually explainable results in terms of metonymy type influence. For example, the developmental trajectory yielded from the three tasks are different in shape, as is shown in Figure 8 (e.g., chronological development suggested by explanation score; U-shaped (reversed) trend suggested by reaction time; rapid development in metonymy comprehension at age 4–5 suggested by eye-tracking task); children are able to understand and explain novel metonymy at a younger stage than conventional metonymy; conventional metonymy acquisition starts at around school age (6–7 years old).

In relation to some extant developmental studies on metonymy (Falkum et al., Reference Falkum, Recasens and Clark2017; Köder & Falkum, Reference Köder and Falkum2020), the present study only finds a U-shape in reaction time (by mean) – there is no U-shape detected in other measures in selection, explanation and eye-tracking tasks. In general, most of the measures in this study presents children’s chronological development with age in metonymy comprehension, but with different rate of development at different stages. In response to the on-going debate between the early onset hypothesis versus the literal stage hypothesis, the results of our study suggest an early onset of novel metonymy before the age of 4–5, and also show that there is no literal stage in children’s metonymy comprehension development. As for conventional metonymy, children’s onset of comprehension is delayed until 6 years old.

Although the present study does not find a U-shaped development with age in the picture selection score, our results can provide further explanation for the U-shape developmental trajectory which was detected in previous studies (Falkum et al., Reference Falkum, Recasens and Clark2017; Köder & Falkum, Reference Köder and Falkum2020).

4.2. High selection randomness of age-3 group

Combining the findings from selection task with explanation task, and also taking into account the three-year-olds’ performance in literal trials, the present study argues that three-year-old children’s performance can be highly random and thus be overestimated in picture selection tasks which were frequently conducted in previous studies.

As is shown in Figures 3–4, the youngest group still make mistakes in literal trials and choose unrelated choices significantly more often than the other two groups, suggesting that, although their selection accuracy is significantly above chance, their choices are largely random. In this way, the picture selection performance of the three-year-olds could be overestimated.

Figure 3. Literal trial performance between age groups.

Figure 4. Times of unrelated choice between age groups.

Figure 5. Metonymy explanation score in different age groups.

Figure 6. Explanation/choice score in different age groups.

Figure 7. Example stimulus.

Figure 8. Developmental trajectory found under different measurements.

Furthermore, in Figure 6 where the ratio of explanation score to selection score is calculated, the mean (0.8832) and the distribution (most being lower than 1) of the ratio of the youngest group mean that there must be cases where three-year-old children choose the correct picture but are unable to explain their choices. This can be caused by two reasons: (1) the three-year-old children get some of their choices right by chance; (2) the three-year-old children’s linguistic ability is still at a relatively low stage of development and thus they are unable to give decent verbal responses in the explanation task. The first reason, which aligns with the high selection randomness discussed above, would also provide further explanation for the U-shape detected in previous studies.

The U-shaped development in previous studies consists of three developmental periods: a good performance of the youngest group, a declined performance of the second youngest group, and a steady improvement of the oldest group. The present study suggests that the high selection randomness of the three-year-olds can, at least, partly explain why a U-shape could possibly occur in previous studies, especially why the youngest group could perform surprisingly well in the selection tasks.

However, as no U-shape is found in most of the measures in the present study, there is a need to explain the discrepancy in findings between the present study and previous studies. From the view of the authors, the discrepancy can be explained, from a methodological perspective, by the unique task design of picture selection experiment. The picture selection experiment in the present study consists of two sections (selection task and explanation task). The two tasks are not done totally separately but in an interweaving and interactive manner – participants, in each trial, needs to give a verbal explanation immediately after they make a choice. In this way, the participants will possibly be more cautious with their choice and their selection accuracy might incline towards the explanation score. Indeed, compared with the U-shape in previous studies, the trajectory found in selection accuracy in the present study is more similar to the explanation score developmental trajectory. In other words, although the present study finds and argues that three-year-olds can be highly random in choosing pictures, the participants of the youngest group in our study, influenced by the “tight” task design, are not so “random” as they could possibly be.

Also, as the present study takes different types of metonymy into consideration, instead of merely including novel metonymies as done in previous studies, it is possible that children’s comprehension of conventional metonymy and of novel metonymy follows different trajectories, and that the criss-cross of the two trajectories can conceal the U-shape in the general finding. For example, the three-year-olds show an almost chance-level selection performance in conventional metonymy trials (refer to Table 6) but they are almost unable to give explanations for their choices in these trials, which, to some extent, pulls down their overall performance in the picture selection and explanation tasks.

4.3. Performance of age 4–5 weakened in selection task

Comparing the findings between eye tracking experiment and picture selection experiment, and also between the selection task and the explanation task, the present study argues four-to-five-year-old children’s performance in the selection task can be underestimated and thus their real metonymy comprehension ability can be masked due to methodological reasons. Also, this argument can provide further explanations for the U-shaped curve (the decline of the four-to-five-year-olds) detected in previous studies, although we do not find a clear drop (only reaction time data show a slight drop by mean) of performance among them.

Compared with the findings from eye tracking experiment where the age 4–5 group shows a significant improvement from the youngest group in both types of metonymy, especially in novel metonymy, the findings from the selection task suggest that the performance of the age 4–5 group is weakened as there is no significant difference between the two younger groups and the age 4–5 group even react more slowly than the age-3 group in choosing pictures. Why would there be such discrepancy in the age 4–5 group’s performance between the two experiments?

The present study would argue that the difference of nature between online and offline tasks can cause the age 4–5 group to perform differently. In online tasks, children are not required to give any verbal response, and they are also under no influence from the examiner as they complete the task alone in the lab. Also, online eye tracking measures can collect data of children’s live eye movements which reflects their ongoing processing of the information on the screen.

However, offline tasks, where children are required to choose the correct picture, can be more pragmatically demanding and complex for children at this age and thus can cause a delay in showing their improved performance. Indeed, children at 4 and 5 are experiencing a sharp increase in vocabulary and thus in conventional lexical meanings (Falkum, Reference Falkum2022). In other words, the age 4–5 children’s growing sensitivity to conventions may not lead them to use and understand the words in a flexible and figurative manner, as figurative uses involve large departures from conventions (Falkum, Reference Falkum2022); and children at this age may also, being cautious, incline to the literal meanings when the context is uncertain, which was described as a literal preference in Köder and Falkum’s study (2020). Such delay in comprehension ability development and late literal preference also has their empirical equivalence in studies on other figurative devices. In tasks where children are required to verbally paraphrase the metaphorical meaning or tell a story containing a metaphor (Pearson, Reference Pearson1990), and tasks where children are forced to choose the correct metaphorical meaning from options (Di Paola et al., Reference Di Paola, Domaneschi and Pouscoulous2020; Waggoner & Palermo, Reference Waggoner and Palermo1989), participants tend to show a lack of comprehension ability until 4 to 5 years old, which may stem from the complexity and high linguistic demand of the tasks used (Köder & Falkum, Reference Köder and Falkum2020).

In the offline task findings of the present study, there is evidence that would suggest age 4–5 group’s increased linguistic ability and high cautiousness in selection. Age 4–5 group are found to be surprisingly slower in reaction than the youngest group, which would suggest that the age 4–5 children are more cautious and may experience more competition in the decision-making process. Being cautious and hesitating, the age 4–5 children are more likely to think too much of the choice itself and misinterpret the intention of the examiners, and thus to stick to literal choices which they think safer. Among the explanation responses, more than one child from the age 4–5 group, after choosing the literal picture in metonymy trials, gave similar responses to the following utterance “(I choose this picture) because you said it’s the big ear (itself), then I pick the big ear”; also, some of them added “(but) the boy there is also big ear”, which means that they may actually understand the expression but they, taking into account of their interpretation of the examiner’s intention and the task’s aim, refrain their real thoughts and do not choose the correct picture.

Furthermore, not only the comparison between the online and offline results, but also the comparison within the offline tasks between the selection and explanation score results can provide explanations for the age 4–5 children’s weakened performance in the selection task. The ratio of explanation score to selection score (shown in Figure 6) of some of the age 4–5 children is greater than 2, which means that there must be cases where the age 4–5 children choose the wrong picture in the selection task, but can still give correct explanations to the meaning of the target metonymy expression, which is in line with the point that age 4–5 children’s performance can be underestimated in the selection task.

Combining the two offline tasks and one online task, the present study would argue that, compared with the age-3 group, the age 4–5 children, in the selection task, show a higher vocabulary and language production ability, adopt a more cautious strategy in the decision-making process, and thus are more likely to stick to literal options in metonymy trials. However, online tasks suggest that the above-mentioned literal preference found in age 4–5 children is only a seeming inclination, which means that their declined performance in the selection task does not equal to a declined metonymy comprehension ability; instead, the apparent preference for literal meanings can be indicative of their growing ability of attending to conventions, an ability with great functional importance in children’s language and social learning (Falkum, Reference Falkum2022; Kalish & Sabbagh, Reference Kalish and Sabbagh2007), as age 4–5 children’s comprehension ability in both types of metonymy improved, compared with their age-3 counterparts in the results from the eye tracking task.

4.4. Socio-cultural experience as a source of difficulty until age 6

The hypothesised 3-type difficulty categorisation of novel metonymy has been found not to be supported as children’s performance in both selection and explanation tasks is not found to decline as the hypothesised level of difficulty increases. As the 3-type categorisation is based on children’s animacy preference (Piaget, Reference Piaget1964) and the metonymy expression’s degree of departure from the prototypical metonymy (Peirsman & Geeraerts, Reference Peirsman and Geeraerts2006), the present study would argue that the difficulty of metonymy comprehension for children does not stem from the inanimacy of content or the distance of internal-domain mapping.Footnote ³

The present study, as one of the first to take into account different types of metonymy, did not stop at the state where the 3-type difficulty for children has been proven wrong. We merged the three novel metonymy sub-types into one ‘novel metonymy type’ and compared it with the conventional metonymy type. The results show significant declines from novel to conventional metonymy and significant delays in learning of conventional metonymy. As is attested in both eye tracking and explanation tasks, children’s comprehension development in novel metonymy and conventional metonymy follows different trajectories: children’s comprehension development of novel metonymy shows a chronological trend, with steady improvements from age 3 to age 7; while in terms of conventional metonymy, significant improvements appears only after the age of 4–5 and it is only the age 6–7 group that start to show a good command of understanding and interpreting conventional metonymies. Considering this, the present study would argue that socio-cultural conventions and experience can be the source of difficulty for children’s metonymy comprehension development.

As is discussed in the literature review, the difficulty of a figurative expression can be estimated from the perspective of its familiarity and semantic transparency (Cacciari & Padovani, Reference Cacciari, Padovani, Spivey, McRae and Joanisse2012). For one thing, the familiarity of a figurative expression reflects the degree of exposure of a speaker to it (Cacciari & Padovani, Reference Cacciari, Padovani, Spivey, McRae and Joanisse2012). Highly socio-culturally related metonymy expressions can be more unfamiliar to children because such metonymy expressions are used on limited occasions and can not be comprehended until a child has experienced such occasions (e.g., the metonymy “Chairman Mao – bank notes/money” can only be understood by interlocutors who have communicated about bank notes in shopping-related contexts). For another, the semantic transparency of metonymy, as a measure of the relation between literal (default) and figurative meaning, could explain the difficulty of metonymy from conceptual aspects (e.g., strength of contiguity and distance of mapping) but would not capture the variability in usage aspects (e.g., socio-cultural experience). In other words, one metonymy expression, with fixed type of contiguity in formal analysis, can be of different difficulty for people with different social experience, as in the metonymy of “drink white”, which could provide mental access to “drink white wine” in a restaurant context only for people, unlikely as children, who have such conversational experience (e.g., being exposed to wine-related conversations at parties or formal dinners).

In what ways can children older than six be more socio-culturally experienced and thus perform far better than their younger counterparts in understanding conventional metonymies? From our perspective, children’s socio-cultural experience in figurative language acquisition can be influenced by family and education in terms of richness and diversity of language input.

For one thing, in mainland China, children at the age of six enter primary school and begin to receive standard education, which, from the very beginning, includes the subject of “yuwen” (literally translated as language and literature). From “yuwen” classes, primary school children receive focused input of literary reading materials which contain rich figurative uses of language and socio-cultural knowledge than daily colloquial language does, which familiarises primary school children with non-literal language uses.

Besides providing standard language education, primary school is also a place where children begin to build diverse interpersonal relationships (e.g., with schoolmates and teachers), which renders a more diverse language input and more chances for error corrections, especially teacher’s correction (Lasagabaster & Sierra, Reference Lasagabaster and Sierra2005), compared to kindergarten period when children’s language input almost comes from their family members, namely parents. As is discussed in §4.3, children at the age of four to five experience a sharp increase in vocabulary and an increasing sensitivity to conventions but still make mistakes in metonymy picture selection tasks, especially in conventional metonymy; it is just possible that age 6–7 children benefit from the correction process, especially peer correction and teacher correction, during which the learned vocabulary becomes more flexible in use.

Indeed, metonymy in real use covers a wider variety than is categorised in Peirsman and Geeraerts’ (Reference Peirsman and Geeraerts2006) model; the novelty and conventionality of metonymy are notions constructed by adults with richer language and social experiences, and thus would have opposing effects in predicting children’s metonymy comprehension (conventional being harder than novel). From the view of the present study, the comprehension process of metonymy not only relies on schematic ability of category levels rendering internal domain mapping but also largely depends on socio-cultural experiences, which could be the source of the children’s delayed acquisition of conventional metonymies.