The Influence of Exemplar Variability on Young Children’s Construal of Verb Meaning

ABSTRACT Verbs serve as the architectural centerpiece of sentences, making verb learning pivotal for language acquisition. Verb learning requires both the formation of a verb-action mapping and the abstraction of relations between an object and its action. Two competing positions have been proposed to explain the process of verb learning: (a) seeing a highly variable range of exemplars allows children to detect and abstract the commonalities across actions—the action invariants; and (b) seeing a less variable range of exemplars enables children to focus on and extract the action invariants. Using manner—a major component of verb meaning in English—as a test case, this study addressed this debate by examining the influence of manner variability on the ability to fast-map new verbs and extend them to novel exemplars in 2-, 3-, and 4-year-old English-speaking children. Results contribute to this debate by showing that high manner variability hindered fast-mapping but facilitated extension to manner variations in the 2- and 3-year-olds. Thus, high exemplar variability may affect verb fast-mapping and extension differently. Furthermore, manner variability did not affect 4-year-olds’ (or adults’) fast-mapping or extension, suggesting that the influence of exemplar variability on verb learning attenuates with age. Finally, manner variability did not affect agent or object extension, revealing a component-specific effect of exemplar variability on verb extension.

Life is fueled by action and activity. People generally refer to entities using nouns but convey an action, an occurrence, or state of change, using verbs. Verbs serve as the architectural centerpiece of sentences, making verb learning fundamental to language acquisition. Furthermore, the fact that words can be extended to refer to a category of exemplars across variations gives words their economy and power. For example, after a child can map the verb "hold" to the action of a woman carrying a baby with both arms, the child still needs to extend the label to actions with variations in the agent, the object (e.g., a man holding a bag), or the manner (i.e., how an action is carried out; e.g., holding the baby with one hand). Thus, mapping a verb to meaning includes two steps: fast-mapping and then extension-a lengthier meaning refinement process, the latter of which likely entails the learner witnessing the new verb used to describe new events and then extending the verb to new, previously unattested events (e.g., Casenhiser & Goldberg, 2005;. Thus, fast-mapping occurs when children form an initial hypothesis of the word meaning from just one or two exposures to it and its referent, while extension happens when children generalize the name of the action beyond the original exemplar over which it was learned (e.g., Carey, 1978;Naigles & Hoff, 2006;Saji et al., 2011).
Fleshing out the nuances of verb meaning is a slow process for children that likely requires exposure to a verb in different linguistic and situational contexts (e.g., Bowerman, 1980;Childers & Paik, 2009;Dromi, 1987;Forbes & Poulin-Dubois, 1997;Gentner, 1978;Huttenlocher et al., 1983;Theakston et al., 2002). One critical factor that might affect the construal of verb meaning is the variability of exemplars children are exposed to during verb learning. For example, an exposure to the label "moving" used to refer to actions that are visually more variable than the label "sprinting" allows learners to understand that "moving" has a wider semantic scope than "sprinting." But how does exemplar variability influence success in fast-mapping and what types of initial exposures likely lead to extensions? While there is no clear answer to this question in the literature, understanding the circumstances under which verb fast-mapping and extension best occur would be important for understanding the process of language acquisition and assisting children whose language is delayed, given that verb learning and extension is exceptionally difficult for children with specific language impairment (e.g., Conti-Ramsden & Jones, 1997;Rice & Bode, 1993;Thordardottir & Weismer, 2002).
This study examines the influence of high exemplar variability on children's fast-mapping and extension of motion verbs. Verbs express not an entire event, but only parts of an event as any event includes multiple actions (Gentner, 2003). For example, the event of a man walking into a house can be labeled as an instance of walking, stepping, swaggering, entering, coming, or going. Research has proposed that there is a core set of action components that are often encoded across languages, such as manner, path, cause, result, and object (Jackendoff, 1983;Langacker, 1991;Slobin, 2001;Talmy, 1985). For example, manner refers to the distinct way in which an agent (the action doer) moves, such as running, jumping, and waving, while path refers to the course followed by the agent related to an external reference point independent of the agent oneself, such as entering a house, leaving a theater, and crossing the border. Languages differ in terms of the relative frequencies of different types of verbs. In some languages (e.g., Spanish, Turkish, Greek), path is often encoded in the verb while manner is typically encoded in a satellite position as a propositional or adverbial phrase (e.g., she is existing the house running). By contrast, in other languages (e.g., English, German, Chinese), manner is often encoded in the verb, and path is typically encoded in a satellite position (e.g., she ran out of the house.). Using the component of manner as a test case, this study examines how exemplar variability affects motion verb learning in young English speakers. This study focuses on manner because it is a major component of verb meaning in English.

Fast-mapping verb meaning: what role does high manner variability play?
There are two views on this question. Some research on natural speech samples suggests that high exemplar variability can facilitate child verb learning (e.g., Gallivan, 1988;Rinaldi et al., 2004;Sandhofer et al., 2000). For example, Gallivan (1988) found that parents' production of particular motion verbs is related to their children's production of these verbs. In addition, parents tend to use very few verbs but repeat them often-a pattern of child-directed speech observed in parents speaking English and Mandarin (Sandhofer et al., 2000) and Italian (Rinaldi et al., 2004). It seems that hearing verbs in multiple situations over time is associated with children's ability to use them correctly. Arguably, verbs that are used in multiple situations over time are associated with exemplar variability, thus supporting the view that exemplar variability facilitates verb learning. However, verbs that are naturally used to describe a wider range of exemplars may also be used more frequently. Thus, this effect may merely be a product of input frequency since high-frequency words also tend to be learned early (e.g., Carroll & White, 1973). To disentangle these factors, the current study examined child learning of novel verbs whose frequency and variability was controlled.
A competing position suggests that high exemplar variability can hinder child initial verb learning. Child initial verb understanding tends to be specific to the objects, actors, manners seen in the event. For example, 2-to 3-year-old children have an overly specific understanding of the early-acquired light verbs that have broad meanings (e.g., make, go; Theakston et al., 2002). In addition, children may use sweep only when they sweep with a broom, and see only when they point and turn to the listener for eye contact (Bates et al., 1979;Harris et al., 1988;Tomasello, 1992). Thus, verbs with more specific meanings may have an advantage over verbs with broader meanings (e.g., Gentner, 2003;Smiley & Huttenlocher, 1995;Tomasello, 1992). Furthermore, words with high imageability ratings (e.g., the ease with which a word elicits a mental image; e.g., Bird et al., 2001;Gilhooly & Logie, 1980;Ma et al., 2009;McDonough et al., 2011) and highly specific meanings (Pulverman et al., 2008) are associated with an earlier age of acquisition than words with lower imageability ratings and less specific meanings. High imageability rating and meaning specificity have been thought to be associative with a less variable range of exemplars . However, the correlational nature of these studies (e.g., Gallivan, 1988;Ma et al., 2009;McDonough et al., 2011;Rinaldi et al., 2004;Sandhofer et al., 2000) precludes drawing causal inferences about the role of exemplar variability in initial verb learning. Furthermore, verbs that are naturally used in multiple situations and verbs that are less imageable and specific in meaning may have high variability in various semantic components (e.g., agent, object, results, instrument) besides manner. Thus, it is still unclear how high manner variability facilitates child initial verb learning.

Extending verb meaning: what role does high manner variability play?
Again, there are two competing views on this question. The view that high exemplar variability may facilitate verb extension is suggested by structural alignment theory, which proposes that children use a comparison process to extract regularities across multiple exemplars-a process that cannot occur in the presence of only one exemplar (Gentner & Namy, 1999Namy & Gentner, 2002). This comparison process allows learners to extend a word to visually different yet within-category exemplars (e.g., Gentner & Markman, 1994;Gentner & Namy, 1999;Waxman & Klibanoff, 2000). Golinkoff and colleagues also proposed that learning a verb such as "drink" requires the abstraction of the common relation from a range of drinking actions performed by different agents, with different drinks, and in different manners (Golinkoff et al., 2002;Maguire et al., 2006). These theories suggest that verb learning requires a process of extracting the commonality across action exemplars that share a common label, thus predicting that high exemplar variability may facilitate verb extension. However, it should be noted that this prediction is moderated by younger children's greater difficulty in seeing similarity across visually highly dissimilar displays.
This view is also supported by research on noun learning and categorization. An exposure to variable exemplars helps children learn and retain novel labels (Perry et al., 2010;Twomey, Ranson et al., 2014). In addition, seeing a broad range of exemplars facilitates children's formation and generalization of nonlinguistic and linguistic categories of geometric shapes (Bomba & Siqueland, 1983), categories of objects (Gentner & Namy, 1999;Graham et al., 2010;Oakes & Spalding, 1997), perceptual organization (Bhatt & Quinn, 2011), and non-adjacent dependencies of three-element strings (e.g., aXc or bXd; Gomez, 2002). Although these studies are not about verb extension, the effect may apply to word acquisition regardless of form classes: Viewing a variable range of exemplars may help children attend to the relational commonality among instances rather than the specifics of the elements instantiating the relation.
Verb extension has also been addressed in some studies. Behrend (1990) taught 3-year-olds novel verbs referring to novel actions that included three components (instrument, manner, result). Children were then asked to judge whether the verb still applied when one component was changed. Results showed that it was harder for children to extend verbs to new manners and results than to new instruments. However, Behrend did not examine the effect of exemplar variability on verb extension. In addition, when 24-month-old infants were habituated to a novel verb token (e.g., "dacking") accompanied by an animated star that was transforming either into the same shape (same exemplars) or different shapes (varied exemplars; Twomey, Lush et al., 2014), infants who viewed varied exemplars discriminated the out-of-category action in test. Thus, visual variability helped children map the novel verb to an action category. Since action category formation serves as the foundation of verb extension, this finding suggests that visual variability may facilitate verb extension. While this finding could just be about heightening children's attention to the display, it may also mean that different exemplars including variable objects are useful for cementing the meaning of a novel verb.
Research suggests that high exemplar variability may facilitate verb extension (Childers, 2011;Childers & Paik, 2009;Forbes & Farrar, 1995;Seston et al., 2009;Snape & Krott, 2018). For example, the 2.5-and 3-year-olds who learned a novel verb through exemplars with different instruments successfully enacted the verb using novel instruments, whereas those children who learned through similar exemplars did not (Childers, 2011;Childers & Paik, 2009). Snape and Krott (2018) found that 3-year-old children extended a novel verb to a new object when they saw an action performed on two different objects, but not when they saw the action performed on one object. These findings suggest that variability of the object facilitated verb extension to new objects. However, these studies did not manipulate the variability of manner-an essential component of verb meaning in English. Forbes and Farrar (1995) showed 3-year-old children a videotaped novel event-containing agent, instrument, manner, and outcome components-which was labeled with a novel verb. Children then saw either the same event three times (the same condition) or three actions with a change in a single component (the different condition). Children were then asked to judge whether the new verb applied to an event with a particular change (e.g., agent change, instrument change, manner change, outcome change). They found that the 3-year-olds were reluctant with manner extension in both the same and different conditions, but seeing multiple exemplars with different instruments and outcomes facilitated the 3-year-olds' verb extension to novel instruments and outcomes. However, this study did not single out the effect of manner variability on children's acceptance of verb extension.
A competing view suggests that high exemplar variability can hinder verb extension. This view is supported by research on infants' categorization of spatial relations. Quinn and colleagues found that 3-month-olds formed spatial categories of above vs. below when the spatial relations were depicted by the same object rather than multiple different objects (Quinn et al., 1996), and that the 6-to 7-montholds' categorical representation of between could be specific to the particular objects and locations used to depict the spatial relation (Quinn et al., 2003(Quinn et al., , 1999. Furthermore, categorical representation of the support relation of on was observed in the 14-month-olds when two (rather than six) exemplars were used to depict this relation (Casasola, 2005). Perhaps, showing fewer objects allows infants to attend to the relation between those objects rather than the specifics of those objects-a view that applies to the semantic acquisition of relational terms including verbs.
Research on verb extension also found that 2.5-and 3-year-olds extended a novel verb to a new agent more readily when shown action exemplars with the same agent rather than different agents (Maguire et al., 2008). In addition, 2.5-to 3-year-old children learning verbs linked to events with multiple objects were more successful if those events were enacted by one single agent rather than multiple agents-a pattern of results observed in children raised in the United States, China, Korea, and Singapore (Childers et al., 2017). Furthermore, Childers et al. (2016) taught 2.5-to 3.5-year-olds a new verb referring to (a) a pair of highly similar (though not identical) events and then varied events (the Progressive Alignment condition), (b) two pairs of more varied events (All Far condition), or (c) a single repeated event (Control). The children in the Progressive Alignment condition (rather than those in the All Far condition) were more likely to extend a verb than those in the control condition, suggesting that seeing a highly similar range of exemplars facilitated verb extension. However, these findings are not about manner variability per se. Furthermore, since an action can be carried out by a variety of agents and on many objects, while other components such as manner, outcome, or path are part of a verb's meaning, agent and object may not be informative to children about verb meaning (Talmy, 1985). Thus, agent and object variability might direct children's attention toward them rather than the action itself, leading to less successful verb learning (Kersten & Smith, 2002).
Thus, research is divided in their conclusion on the influence of exemplar variability on verb learning, with some studies suggesting that high exemplar variability facilitates verb learning (e.g., Childers & Paik, 2009;Forbes & Farrar, 1995), while other studies suggesting exactly the opposite (e.g., Maguire et al., 2008). Furthermore, it is unclear whether the unsuccessful verb extension observed in previous research (e.g., Childers et al., 2017;Forbes & Farrar, 1995;Maguire et al., 2008) was due to the children's reluctance to extend the verb or their failure to form an initial association (fast-mapping) between a verb and an action referent. This issue can only be addressed when both fast-mapping and extension are examined in the same study.
This study examined the influence of manner variability-an essential component of verb meaning (Slobin, 2001;Talmy, 1985)-on 2-, 3-, and 4-year-old children's verb learning. We asked whether exemplar variability affects verb fast-mapping and extension differently, and whether the effect of manner variability is specific to manner extension or is generalizable to agent and object extensions. A preferential pointing paradigm was then used at test, in which participants were shown a withincategory event and an out-of-category event side-by-side simultaneously and asked to choose which one could be named by the verb (Golinkoff et al., 2013;Imai et al., 2008;Ma, zhou et al., 2020).

Participants
The participants were 43 2-year-olds (M = 31.49 months, SD = 1.96; range = 28.03-34.67; 19 boys), 48 3-year-olds (M = 37.74 months, SD = 1.94; range: 35.06-41.97; 24 boys), and 48 4-year-olds (M = 50.21 months, SD = 2.09; range: 46.37-53.97; 24 boys) English-reared children. Eight additional children were excluded from the final sample due to failure to complete the experiment. Three age groups were recruited, permitting us to evaluate the possible differential effect of manner variability on verb fast-mapping and extension in 2-and 3-year-olds who are able to learn novel verbs under experimental conditions (e.g., Waxman et al., 2009), though this ability is still fragile at age 3 (Imai et al., 2008). The 4-year-olds were recruited to explore the development of the effect of manner variability on verb learning. This study was approved by the research ethics committee of the University of Delaware. The parental consent obtained for all child participants was both informed and written.

Procedure and design
Children were tested individually by an experimenter, who is a native speaker of English, in a laboratory at the University of Delaware. Following the procedure established in previous research (e.g., Imai et al., 2008;Ma, zhou et al., 2020), all visual stimuli were presented on a computer screen, while all speech stimuli were presented in child-directed speech by the experimenter who stood behind the child (Ma et al., 2011, Ma, Fiveash et al.,2020. An experiment started with a motivation phase, during which children were shown an animation of an astronaut, Jimmy, landing on a new planet and wanting to learn a new language (Figure 1). The experimenter asked the child, "Would you like to help Jimmy learn some new words? Do you want to help Jimmy to learn some new words?" The two-trial task familiarization phase then introduced children to pointing at one of two actions in a Preferential Pointing Paradigm (Imai et al., 2008;Ma, zhou et al., 2020). Children were shown two videotaped actions of familiar verbs side-by-side on a computer screen (i.e., clapping vs. jumping; turning vs. waving). The experimenter asked children to point to one of the actions on each trial, selected at random (e.g., "Where is she jumping?"). After children pointed, the experimenter said, "Thanks," and proceeded immediately to the next trial. This study used pointing behavior as the dependent variable, as children start to reliably use pointing during novel word learning at two years of age (Merriman & Bowman, 1989), thus allowing us to use the same dependent variable on all participants in this study: 2-, 3-, and 4-year-old children and adults.
Then, the word learning task started, where children learned two novel verbs (twilling /twiling/, clooming /kluming/), one in each block. Each block contained a training phase and a test phase. In the training phase, each novel verb was instantiated by two 8-sec video-taped action tokens, which were presented in either a narrow version or a broad version. The terms narrow and broad are used to distinguish between how much the manner changes in the two actions: the narrow version was instantiated by two actions with slightly different manners, while the broad version was instantiated by two actions with more variable manners. For example, the narrow version of twilling was instantiated by an actress moving a ball up and down repetitively (a) with one arm-referred to as Trained Action 1 hereafter-and (b) with two arms-referred to as Trained Action 2 hereafter. The broad version of twilling was instantiated by the actor moving the ball repetitively (a) up and down with one arm (Trained Action 1) and (b) in a circle with two arms-referred to as Training Action 3 hereafter (Table 1). On each training trial, children saw one action and heard the novel verb presented twice in complete sentences ("Look! She is twilling the ball. Look at her twilling it.") spoken by the experimenter. A parallel design was used for clooming (Table 2). This study used a within-subject design, where each child learned both twilling and clooming, one in the narrow version (referred to as the narrow verb hereafter) and the other in the broad version (referred to as the broad verb hereafter). The presentation order of the two blocks and the verb type assignment (narrow, broad) of the two verbs were counterbalanced across participants. This set was shown 3 times in total Figure 1. Examples of the experimental sequence (two of the conditions). Each child learned two novel verbs (twilling, clooming): one narrow and one broad. All participants saw the same test phase regardless of whether they learned a verb as a narrow or broad version. Participants were randomly assigned to 1 of 12 between-subjects conditions, created by counterbalancing 4 factors: (1) the order of presentation of the two verbs; (2) the assignment of version type (narrow, broad); (3) the order of presentation of the tests; and (4) the left-right assignment of the target.
Four measures were taken to enhance child verb learning. First, a live experimenter described all the events occurring on the screen, as children learn verbs from video with a live person better than from video alone (Roseberry et al., 2009). Second, the target verbs were presented in sentences, as the use of sentence frames facilitates child word recognition (Fernald & Hurtado, 2006;Imai et al., 2008). Third, familiar objects (a basketball, a stuffed Pooh Bear) were used as prop, since familiar objects help children attend to the action (rather than the object) in verb learning (Kersten & Smith, 2002). Finally, each set of two training trials was followed by an 8 sec contrast trial, in which a new action was performed accompanied by the sentence ("She is not twilling the ball. She is not twilling it"). Contrastive stimuli aid learners to narrow down the range of possible meanings attested (e.g., Au & Markman, 1987;Diesendruck, 2005;Heibeck & Markman, 1987;Piccin & Waxman, 2007;Waxman & Klibanoff, 2000). Participants saw each set of the two training trials and a subsequent contrast trial * Moving the ball from shoulder to shoulder under the neck while the one hand is holding the ball and keeping the ball close to the body the entire time. ** Moving the ball from side to side using both arms, while both hands are holding the ball the entire time. *** Tossing the ball into the air using one hand and then catching it with the other hand, then continuing this juggling action. Both hands are at the same height related to the floor.
three times during the training phase before proceeding to the test phase. In total, in the training phase, children heard the label "twilling" (or "clooming") 12 times in the training trials (2 times/trial × 2 trials × 3 sets) and 6 times (2 times/trial × 3 trials) in the contrast trials. The test phase started with two retention tests assessing child fast mapping performance, followed by three extension tests assessing child acceptance of extensions to a new manner, a new agent, and a new object, respectively. On each test trial, children were shown two dynamic actions side-by-side and were asked to point at the target action (e.g., "Where is she twilling the ball?"). The same tests were used for all participants regardless of whether they learned twilling (or clooming) as a narrow or broad verb, while the left-right assignment of the target of the five tests and the presentation order within the two retention tests and that within the three extension tests were counterbalanced across participants. Since Trained Action 1 was used in the training for both the narrow and broad versions of a verb, Trained Action 1 served as the target in the test phase except the manner extension trial. Take twilling for example. Retention Test 1 examined whether children could map twilling to Trained Action 1 when paired with the contrast action seen in the training. Retention Test 2 examined whether children could map twilling to Trained Action 1 when paired with a novel out-of-category action (i.e., the actress moving the ball from one shoulder to the other). Retention Test 2 was included as a strong test of child fast-mapping performance because Trained Action 1 was paired with a novel action not seen in the training. Compared with Retention Test 1, Retention Test 2 was arguably more difficult and thus more likely to reveal children's potentially different fast-mapping performance across verb types. Then, the extension trials began. In each extension trial, only one action component (manner, agent, or object) was changed. In the object extension test, the actress seen during training performed the pair of actions used in the Retention Test 2 on a new object: a red box. In the agent extension test, a male actor performed the pair of actions used in Retention Test 2. The object and agent extension tests were included to determine whether exemplar variability influences verb extension in a component-specific or component-general way. Does the change of a component (manner) influence the extension of other components (object, agent)? In the manner extension test, children were shown two novel actions side-by-side. In one action, the actress was moving the ball with two arms from left to right-a manner that is different from but still similar to the action seen in the training. In the other action, the actress was tossing the ball between her hands-a manner that is highly dissimilar to the action seen in the training.

Verification of verb type (narrow, broad) and target assignment in the tests
Sixteen adult native speakers of English (M = 23.13 years, range = 19-31 years; 9 females)-all undergraduate or graduate students at the University of Delaware-were administered a perceptual similarity rating task. Participants first read the following instruction: You will watch pairs of events side-by-side. You will decide how similar the pairs of actions are on a scale from 1-9. A rating of 1 means "not similar at all", and a rating of 9 means "extremely similar." Feel free to use the whole range of numbers on the scale. The adult similarity judgment task served two purposes. First, adult ratings could justify the narrow and broad verb assignment. That is, the two actions used to instantiate the narrow version of a verb should be more perceptually similar to each other than the two actions used to instantiate the broad version of the verb. Second, adult ratings could justify the assignment of targets in the extension tests. By pairing the target and non-target in the three extension tests with the three actions used in teaching the narrow and broad version of a verb, we could determine whether the trained actions were perceptually more similar to targets than non-targets in extension tests.
The perceptual similarity task started with two practice trials, where a pair of highly similar objects (two non-identical baby bottles) and a pair of less similar objects (a baby bottle vs. a car) were shown, respectively. The practice trials aimed to introduce the participants to a task where the perceptual similarity of the two visual stimuli paired together may differ across trials. The practice trials were followed by 40 test trials-20 for each verb (twilling, clooming). For each verb, there were: (a) two trials, where the Trained Actions 1-2 (i.e., the two actions used for training the narrow version of the verb) and the Trained Actions 1-3 (i.e., the two actions used for training the broad version of the verb) were shown, respectively; (b) nine trials, on each of which the target of the manner, agent, or object extension trial was paired with Trained Actions 1, 2, and 3, respectively; and (c) nine trials, on each of which the non-target of the three extension trials was paired with Trained actions 1, 2, and 3, respectively. The presentation order of the 40 tests was semi-randomized within a participant-with the constraints that no more than two twilling or clooming trials were presented consecutively and the same action was not shown on two consecutive trials-and was counterbalanced across participants. * To make this motion possible, the Pooh Bear was attached to the right lower leg (or both lower legs) using transparent tape. ** Pushing the upper part of the Pooh Bear up and down with her right foot while stabilizing the Pooh Bear by pressing the lower part of its body against the floor using her left foot. *** Moving the Pooh Bear from side to side using both legs, above and parallel to the ground. **** Wiggling the Pooh Bear's head back and forth with her left leg, while stabilizing the Pooh Bear by pressing its left leg against the floor using her left leg.
For the two practice trials, a paired sample t test showed that the similarity rating of the two highly similar objects (i.e., two baby bottles) (M = 5.88, SD = 1.67) was significantly higher than that of the two less similar objects (i.e., a baby bottle vs. a car) (M = 1.44, SD = .73, t (15) = 11.80, p < .001), demonstrating that this is a valid method of assessing perceptual similarity.
For the 36 trials (18 for each verb) where the target (and the non-target) of each extension trial was paired with Trained Action 1, 2, and 3, respectively, we calculated (a) the average ratings across the two trials where the target (and the non-target) of each extension test was paired with Trained Actions 1 and 2, respectively-the two actions used for training the narrow version of a verb, and (b) the average ratings across the two trials where the target (and the non-target) of each extension test was paired with Trained Actions 1 and 3, respectively-the two actions used for training the broad version of a verb. Separate paired sample t tests showed that the trained actions were more similar to the target than the non-target for each extension test, each verb, and each version (narrow, broad; p's < .001), thus verifying the target assignment of the extension tests (Table 3).
In addition, another group of 16 adult native speakers of English (M = 22.45 years, range = 19-33 years; 8 females)-all were undergraduate or graduate students at the University of Delawarewere administered the word learning task (without the motivation and task familiarization phases) that the children completed. Among the total 160 test trials (16 participants × 2 verbs × 5 test trials), there was only one trial (a manner extension trial), where a single participant did not select the target, showing that adults performed at ceiling, and that verb type (narrow vs. broad) did not affect adults' fast-mapping or extension. This finding further verified the target assignment of the extension trials.

Results
The dependent variable used in data analyses is child response on each test (1 = target; 0 = non-target).
Child performance on the two task familiarization trials approached the ceiling level in all age groups: 2-year-olds (90%; 77 target responses out of 86 trials [43 children × 2 tests]); 3-year-olds (97%; 93 target responses out of 96 trials [48 children × 2 tests]); and 4-year-olds (100%; 96 target responses out of 96 trials [48 children × 2 tests]). Thus, children had no difficulty with the task. We also conducted data analyses including only the children who gave target response on both task familiarization trials, and found the same pattern of results as reported herein.
To determine whether child performance differed by verb type, age, and test type, mixed effect model analyses were conducted to account for variability by item and variability by participant. Because the dependent variable was dichotomous (target, non-target), a mixed effect direct-entry logistic regression analysis was conducted with 1390 trials (139 children × 10 tests [(2 retention tests + 3 extension tests) × 2 verbs]) and 6 predictors: verb type (narrow, broad), test type (retention, extension), age (2-, 3-, 4-year-olds), and the 3 interactions among them (verb type × age, verb type × test type, test type × age). The main effects of age (p < .001), test type (p < .001), and verb type (p = .03), and the interactions of verb type × age (p = .01) and of verb type × test type (p < .001) were significant (Table 4)  * The average of the two trials whether the target action was paired with Trained actions 1 and 2.
** The average of the two trials whether the non-target action was paired with Trained actions 1 and 2. *** The average of the two trials whether the target action was paired with Trained actions 1 and 3. **** The average of the two trials whether the non-target action was paired with Trained actions 1 and 3. To determine whether the verb type × test type interaction was still evident when a wider age difference existed across the 2-, 3-, 4-year-olds, we also conducted data analyses within the younger 2-, 3-, 4-year-olds and within the older 2-, 3-, 4-year-olds, respectively, based on the median split of age in each age group. Separate logistic regression analyses showed a significant verb type × test type interaction in the younger and older groups, respectively (see Supplemental Online Information), suggesting that the finding that verb type (narrow, broad) affected fast-mapping and extension differently applied to both the younger and older 2-, 3-, 4-year-olds. Thus, child performance was examined in three age groups (2-, 3-, and 4-year-olds), respectively. Six analyses were conducted within each age group. First, to examine whether verb type (narrow, broad) affected verb fast-mapping and extension differently within each age group, a mixed effect logistic regression analysis was conducted with three predictors (verb type, test type, the verb type × test type interaction). The differential effect of verb type on fast-mapping and extension was indicated by a significant verb type × test type interaction.
Second, to determine whether children reliably fast-mapped and extended the verbs, ten Sign tests examined whether they were highly likely to give a target response on each of the ten tests ([2 retention tests + 3 extension tests] × 2 verbs). An adjusted significance cutoff level of .005 (.05/10) was used. This analysis also allowed us to determine whether certain tests were exceptionally difficult.
Third, to explore how verb type affected fast-mapping and extension performance, five nonparametric Wilcoxon matched pairs tests compared child response in each test between verb types (narrow, broad). An adjusted significance cutoff level of .01 (.05/5) was used.
Fourth, two Wilcoxon matched pairs tests examined whether Retention Test 2 (Trained Action 1 vs. a novel out-of-category action)-a strong test of fast-mapping-was more difficult than Retention Test 1 (Trained Action 1 vs. contrast) for the narrow and broad verbs, respectively. If fast-mapping is exceptionally difficult with a certain verb type (narrow, broad), child performance may differ across the two retention tests for that particular type of verb. An adjusted significance cutoff level of .025 (.05/ 2) was used.
Fifth, to investigate whether the manner, agent, and object extension tests were similarly difficult, six Wilcoxon matched pairs tests compared child performance among the three extension tests for the narrow and broad verbs, respectively. An adjusted significance cutoff level of .008 (.05/6) was used. This analysis included all the participants (n = 139) in this study * < .05 ** < .01 *** < .001 Sixth, we examined whether children's fast-mapping performance can predict their responses (target, non-target) in each extension test. For the narrow verb, we first calculated the fast-mapping target response rate across the two retention tests (either 0, .5, or 1) within each child. Then, three point-biserial correlation analyses examined whether children's fast-mapping target response rate was associated with their responses on each of the three extension tests. Then, the same analyses were conducted for the broad verb. An adjusted significance cutoff level of .008 (.05/6) was used.

The 2-year-olds' performance
First, a mixed effect logistic regression analysis was conducted with 430 trials (43 children × 10 tests) and three predictors (verb type, test type, the verb type × test type interaction). Data reporting is focused on the main effect and interaction depending on verb type. The verb type × test type interaction was marginally significant (p =. 055; Table 5) , suggesting that verb type affected fastmapping and extension differently in the 2-year-olds. This marginally significant interaction was explored in the following analyses.
Then, ten Sign tests examined whether the children were highly likely to give a target response in each test. The 2-year-olds were highly likely to give a target response in Retention Test 1 for both the narrow (p = .005; 31 target responses out of 43 children) and broad (p = .002; 32 target responses) verbs, suggesting that they reliably fast-mapped the verbs. In addition, they were marginally highly likely to give a target response in Retention Test 2 for the narrow verb (p = .01; 30 target responses out of 43 children; a significance cutoff level of .005 was used) but not for the broad verb (p = .76; 23 target responses out of 43 children), suggesting that Retention Test 2 was exceptionally difficult with the broad verb. Furthermore, the 2-yearolds were not highly likely to give a target response on any of the six extension trials. Given the children's reliable verb fast-mapping, their unsuccessful verb extension cannot be due to a failure to form an initial association between a verb and an action.
Next, five nonparametric Wilcoxon matched pairs tests compared child response in each test (target, non-target) between the narrow and broad verbs. In the manner extension test, the 2-year-olds were marginally more likely to give a target response with a broad verb (28 target responses out of 43 children) than with a narrow verb (18 target responses; z = 2.24, p = .03; a significance cutoff level of .01 was used), suggesting that high manner variability facilitated manner extension. However, performance did not differ by verb type in the two retention tests or in the agent and object extension tests (Table 6; Figure 2).  Then, two nonparametric 1 Wilcoxon matched pairs tests compared child performance between the two retention tests for the narrow and broad verbs, respectively. For the narrow verb, performance did not differ between Retention Tests 1 and 2 (z = .28, p = .78). However, for the broad verb, performance was marginally better in Retention Test 1 than in Retention Test 2 (z = 1.96, p = .05), further verifying that Retention Test 2 was exceptionally difficult with the broad verb.
Next, six Wilcoxon matched pairs tests compared the 2-year-olds' performance among the three extension tests for the narrow and broad verbs, respectively. Child performance did not differ among the three extension tests for either the narrow or broad verb, suggesting that the manner, agent, and object extension tests were similarly difficult for the 2-year-olds.
Finally, we examined whether the 2-year-olds' fast-mapping performance predicted their responses (target, non-target) in each extension test. Among the five tests for the narrow verb, children's target response rate across the two retention tests was not associated with their responses in the manner (r(43) = .10, p = .54), agent (r(43) = .23, p = .12), or object (r (43) = .04, p = .82) extension test. Similarly, among the five tests for the broad verb, children's target response rate across the two retention tests was not associated with their responses in the manner (r(43) = .10, p = .54), agent (r(43) = .23, p = .12), or object (r(43) = .04, p = .82) extension test.

The 3-year-olds' performance
First, a mixed effect logistic regression analysis was conducted with 480 trials (48 children × 10 tests) and three predictors (verb type, test type, the verb type × test type interaction). Data reporting was again focused on the main effect and interaction depending on verb type. The main effect of verb type (p = .004) and the verb type × test type interaction were significant (p = .002;  In the 2-year-olds, the main effect of Test Type was also significant (p = .003; Then, ten Sign tests examined whether the 3-year-olds were highly likely to give a target response in each test. In Retention Test 1, the 3-year-olds were highly likely to give a target response for both the narrow (48 target responses out of 48 children) and broad (p < .001; 46 target responses) verbs. In Retention Test 2, they were highly likely to give a target response for the narrow verb (p < .001), and marginally highly likely to do so for the broad verb (p = .01, 33 target responses; a significance cutoff level of .005 was used). These findings suggested that the 3-year-olds reliably fast-mapped the verbs. In addition, for the narrow verb, the 3-year-olds were highly likely to give a target response in the agent (p < .001; 38 target responses) and object extension (p < .001; 37 target responses) tests, but not in the manner extension test (p = .88). These findings suggested that manner extension was exceptionally difficult with the narrow verb. For the broad verb, the 3-year-olds were highly likely to give a target response in the agent (p = .005; 34 target responses) and manner extension (p < .001) tests, and marginally highly likely to do so in the object extension test (p = .06; 31 target responses).
Next, five nonparametric Wilcoxon matched pairs tests compared child response in each test between the narrow and broad verbs. In Retention Test 2, the 3-year-olds were more likely to give a target response with a narrow verb (47 target responses out of 48 children) than a broad verb (33 target responses; z = 3.50, p < .001). By contrast, in the manner extension test, they were more likely to give a target response with a broad verb (40 target responses) than with a narrow verb (23 target responses; z = 3.55, p < .001). However, performance did not differ by verb type in Retention Test 1 or in the agent or object extension test (Table 6; Figure 2).
Then, two non-parametric Wilcoxon matched pairs tests compared child performance between the two retention tests for the narrow and broad verbs respectively. For the narrow verb, child performance did not differ between Retention Tests 1 and 2 (z = 1.00, p = .32). However, for the broad verb, child performance was significantly better with Retention Test 1 than with Retention Test 2 (z = 3.61, p < .001), further confirming that Retention Test 2 was exceptionally difficult with the broad verb.
Next, six Wilcoxon matched pairs tests compared child performance among the three extension tests for the narrow and broad verbs respectively. For the narrow verb, the 3-year-olds were more likely to give target responses in the agent (z = 3.44, p = .001) and object (z = 2.99, p = .003) extension tests than in the manner extension test, confirming that manner extension could be more difficult than agent and object extension (Behrend, 1990). However, the analyses on the broad verb revealed no significant results, suggesting that the manner, agent, and object extension tests were similarly difficult for the 3-year-olds for the broad verb.
Finally, we examined whether the 3-year-olds' fast-mapping performance predicted their responses (target, non-target) in each extension test. Among the five tests for the narrow verb, children's target response rate across the two retention tests was marginally significantly associated with their responses in the agent extension test (t(48) = .28, p = .05; a significance cutoff level of .008 was used), but not in the object (t(48) = −.08, p = .59) or manner (t(48) = .14, p = .34) extension test for the narrow verb, suggesting that when learning the narrow verb, the high performers in the retention tests were more likely to give a target response in the agent extension test. Among the five tests for the broad verb, children's target response rate across the two retention tests was significantly associated with their responses in the manner extension test (t(48) = .42, p = .003), but not in the agent (t(48) = −.08, p = .59) or object (t (48) = .08, p = .61) extension test, suggesting that when learning the broad verb, the high performers in the retention tests were more likely to give a target response in the manner extension test.

The 4-year-olds' performance
First, a mixed effect logistic regression analysis was conducted with 480 trials (48 children × 10 tests) and 3 predictors (verb type, test type, the verb type × test type interaction). The analysis revealed no significant results, suggesting that performance did not differ by either verb type or test type (Table 5).
Then, ten Sign tests showed that the 4-year-olds were highly likely to give a target response in each test (p's < .001).
Next, five nonparametric Wilcoxon matched pairs tests compared child response in each test (target, non-target) between the narrow and broad verbs. In Retention Test 2, the 4-year-olds were marginally more likely to give a target response with a narrow verb (46 target responses out of 48 children) than with a broad verb (41 target responses; z = 2.36, p = .025; a significance cutoff level of .01 was used). However, child performance did not differ by verb type on either Retention Test 1 or the three extension tests (Table 6).
Then, two nonparametric Wilcoxon matched pairs tests showed that child performance did not differ between Retention Tests 1 and 2 for either the narrow (z = .58, p = .56) or broad (z = 1.67, p = .10) verbs.
Next, six Wilcoxon matched pairs tests compared performance among the three extension tests for the narrow and broad verbs, respectively. For the narrow verb, the 4-year-olds were marginally more likely to give a target response in the agent extension test (47 target responses out of 48 children) than in the manner extension test (41 target responses; z = 2.12, p = .034). However, performance did not differ between the agent and object extension tests (z = 1.41, p = 16) or between the object and manner extension tests (z = 1.63, p = .10; Table 6). For the broad verb, the analyses revealed no significant differences among the three extension tests (Table 6).
Finally, we examined whether the 4-year-olds' fast-mapping performance predicted their responses (target, non-target) in each extension test. Among the five tests for the narrow verb, children's target response rate across the two retention tests was significantly associated with their responses in the agent (t (48) = .56, p < .001), object (t(48) = .64, p < .001) and manner (t(48) = .38, p = .008) extension tests, suggesting that when learning the narrow verb, the high performers in the retention tests were also more likely to give a target response in the three extension tests. Among the five tests for the broad verb, children's target response rate across the two retention tests was significantly associated with their responses in the object extension test (t(48) = .41, p = .004) but not in the agent (t(48) = −24, p = .10) or manner (t(48) = .05, p = .74) extension test, suggesting that when learning the broad verb, the high performers in the retention tests were also more likely to give a target response in the object extension test.

Discussion
This study examined the effect of exemplar variability on 2-, 3-, and 4-year-old English speakers' verb fast-mapping and extension. Each child learned two novel verbs (twilling, clooming)-one narrow and one broad-while the narrow/broad version assignment of the two verbs was counterbalanced across participants. We assessed verb fast-mapping and extension to new actions with a manner, or agent, or object change. Using the verb component of manner as a test case, this study aimed to differentiate the influence of exemplar variability on fast-mapping and extension. This study found that high manner variability could hinder fast-mapping, as (a) performance on Retention Test 2 was better with the narrow verb than with the broad verb in the 3-and 4-year-olds, (b) the 2-year-olds reliably found the target on both retention tests for the narrow verb, but they did not reliably find the target on Retention Test 2-a strong test of fast-mapping-for the broad verb, and (c) for both the 2-and 3-year-olds, Retention Tests 1 and 2 were similarly difficult for the narrow verb, but Retention Test 2 was more difficult than Retention Test 1 for the broad verb. By contrast, this study found that high manner variability could facilitate manner extension, as (a) the 2-and 3-year-olds' performance on the manner extension test was better with the broad verb than with the narrow verbs, and (b) the 3-year-old reliably found the target on the manner extension test for the broad verb but not for the narrow verb. Thus, this study found that exemplar variability may affect the verb fast-mapping and extension differently. Notably, children-for all three age groups-reliably found the target in Retention Test 1, suggesting that the 2-and 3-year-olds' unsuccessful verb extension cannot be due to the failure of forming an initial association between a label and an action.
This study also found that manner variability did not affect agent or object extension. In addition, the association between fast-mapping and extension performance is mostly evident in the 3-and 4-year-olds rather than the 2-year-olds. Furthermore, consistent with past research (e.g., Behrend, 1990;Forbes & Farrar, 1995;Imai et al., 2008), we found that: (a) verb learning performance increased with age, and the 4-year-olds performed almost at ceiling, suggesting that the influence of manner variability on verb learning attenuated between ages 3 and 4; (b) extension is more difficult than fastmapping; and (c) manner extension could be more difficult than agent and object extension, as indicated by the 3-and 4-year-olds' performance on the narrow verb.

How did high manner variability affect verb fast-mapping and extension?
We found that seeing a visually more similar range of exemplars facilitated fast-mapping, thus supporting the conclusion that words with high imageability ratings or highly specific meanings tend to be learned earlier than words with lower imageability ratings or less specific meanings (e.g., Bird et al., 2001;Ma et al., 2009;McDonough et al., 2011;Pulverman et al., 2008). High imageability and meaning specificity have been thought to be associative with a narrow set of exemplars , which may facilitate the abstraction of commonalities across action exemplars. Furthermore, this study supports the conclusion that high exemplar variability could facilitate verb extension (Childers, 2011;Childers & Paik, 2009;Forbes & Farrar, 1995), thus arguing against the findings that high exemplar variability hindered verb extension (Maguire et al., 2008). This may be related to a major methodological difference across studies: Maguire manipulated a verb component (agent) that is apparently not as essential to verb meaning as manner.
The differential effect of manner variability on verb fast-mapping and manner extension is consistent with the structural alignment theory, which suggests that verb extension requires a comparison process to extract regularities across multiple exemplars (Gentner & Namy, 1999Namy & Gentner, 2002). Thus, exemplar variability should facilitate verb extension. However, children may have difficulty in seeing similarity across highly different displays. This study suggests that while it may be easy to form a category from a narrow range of exemplars, the category's tight scope may inhibit the inclusion of novel exemplars. Although it is more difficult to form a category from a broad range of exemplars, the wider category scope may allow novel exemplars to be more readily included.
The current findings are also consistent with the Bayesian probability theory on the effect of exemplar variability on word meaning acquisition. In Xu and Tenenbaum (2007), after adults and 3-and 4-year-old children were shown either three identical red apples, or three different colored apples, or three vegetables of different kinds, which were labeled by a novel noun (e.g., fep), they were asked to find another fep from a set of objects. Participants who saw three identical red apples did not extend the label at the basic (e.g., yellow apple) and superordinate levels (e.g., an eggplant). Participants who saw three identical red apples extended the label at the basic level, but not at the superordinate level. Only the participants who saw three vegetables of different kinds extended the label at both the basic and superordinate levels. The findings suggest that in order to extend a label at a certain level, language learners may require prior exposure to the label that has been applied at that corresponding level. The current study revealed that in order to extend the verb to a manner change, the 3-year-olds need to hear the verb label a wide range of manners. To appropriately form a hypothesis about a verb's meaning, children must hear the verb applied to a wide set of exemplars, exemplifying the action category to which a verb refers. Otherwise, children tend to form an overly narrow construal of the verb's meaning and perhaps encoding too much irrelevant detail into the meaning of the verb.

Why did the 2-and 3-year-olds underextend the narrow verbs?
It is important to note that action invariants can be abstracted at different levels. For example, the action invariants across "moving a ball up and down with one or two arms"-the two actions used for training the narrow version of twilling-can be abstracted narrowly as "moving a ball up and down," or more broadly as "moving a ball," or highly broadly as "moving." Thus, it would be possible to form a broad action category even from a narrow range of exemplars. However, when twilling was trained as a narrow verb, the 2-and 3-year-olds formed a narrow action category, precluding them from extending twilling to a new manner of moving the ball from side to side. The 4-year-olds (and adults) extended the narrow version of twilling to the new manner, revealing a flexible encoding of manner in their verb meaning construal. A question of great theoretical import then arises: Why did younger children form a narrow action category from a narrow range of exemplars.
We propose three explanations. First, this may be related to young children's tendency to store overly specific information about a word's meaning (Huttenlocher et al., 1983;Tomasello, 1992). To learn the meaning of a verb, children first form an initial hypothesis. Then, they decide where the boundary of this lexical category falls, with further exposure in which the verb is used to name previously unseen and yet within-category action exemplars. Children initially tend to attach too much information to a verb's meaning (Bates et al., 1979;Harris et al., 1988;Tomasello, 1992). As this tendency decreases with age, the 4-year-olds might form more abstract action invariants of twilling even when it was instantiated as a narrow verb. This is analogous to how infants initially encode the phonological information in their word representations. Infants begin by encoding acoustic features such as speaker gender, emotional tone, and prosody-specifics that are irrelevant to word identityinto the phonemic representation of a word; these acoustic features are also processed by older children and adults but not at the phonemic level (Houston & Jusczyk, 2000;Ma et al., 2017;Newman, 2008).
Second, the 2-and 3-year-olds may already understand that verbs tend to name the manner of an action in English (Slobin, 2001), thus leading them toward a bias for naming the manner of an action in a word learning task. This bias has been observed in English-reared children's learning of novel verbs as early as age 3 under experimental conditions (Maguire et al., 2010) and in their verb vocabularies that contain more manner verbs than other verbs by 36 months of age (Fenson et al., 1994). This bias may prevent the 2-and 3-year-olds from accepting a manner change as the exemplar of a newly learned verb. Perhaps, the 4-year-olds may already understand that verbs also referred to action components other than manner in English-such as leaving (path), hammering (instrument), and breaking (result)-which may reduce their manner sensitivity between ages 3 and 4. Research suggests that even before 24 months of age children can use familiar verbs to refer to a variety of appropriate actions enacted by various actors and with various affected objects (Naigles et al., 2009). Perhaps, in parents' naming practice, the early-acquired verbs are used to refer to a high variable range of exemplars with variations in agent, object, and manner, thus facilitating children's verb extension. Nevertheless, the 2-and 3-year-olds' failure of extending the verb to a manner change suggests that they are sensitive to manner variations in verb learning, and that they understand the core semantic elements of the tested verbs.
Third, the ability to extend newly learned verbs may improve with one's language development and/or general cognitive development. Compared with the 2-and 3-year-olds, the 4-year-olds and adults likely had more experience extending newly learned verbs to previously unseen exemplars, better working memory, and memory retrieval strategies. Thus, they could quickly adjust their initial interpretation of verb meaning to identify the action that was more likely to be the target in a twooption forced-choice task. Supporting this explanation is the current design where the same actions were used in multiple trials. Trained Action 1 was used as the target in Retention Tests 1 and 2 and the agent and object extension tests, and the same non-target action was used in Retention Test 2 and the agent and object extension tests. It is possible that the older children and adults were superior in encoding and remembering the actions seen on earlier training and test trials, which then informed their responses on later trials The third explanation is also supported by the finding that the association between child fastmapping and extension performance is mostly evident in the 3-and 4-year-olds rather than the 2-year-olds. Thus, the relationship between child fast-mapping and extension performance may change with age. The 3-and 4-year-olds who were better at retention tests were also better at extension tests-a finding that does not apply to the 2-year-olds. It is important to note that, unlike the current study, past research found that English-reared 2-year-olds extended newly learned verbs to new objects under experimental conditions (Arunachalam & Waxman, 2015). This may be related to several methodological differences across studies. Arunachalam and Waxman (2015) trained children on only one novel verb and tested on only object extension, while the current study trained children on two novel verbs and tested on three types of extension. Thus, the verb learning task used in this study may be more difficult than that used in Arunachalam and Waxman (2015). In addition, Arunachalam and Waxman (2015) showed children the trained action performed on a new object versus a new action performed on the old object in the test. The current study showed children the trained action performed on a new object versus a novel action performed on the same new object in the object extension test. Perhaps, showing a new object and an old object on the same trial helped children realize that object is not an essential component of the meaning of the newly learned verb, thus helping them to direct more attention to the manner of the action (Arunachalam & Waxman, 2015). Nonetheless, the current findings are consistent with the past findings that (a) the 2-to 3-year-olds have difficulty extending a newly learned verb to the same action performed on a new object-a finding observed cross-culturally in English, Chinese, and Japanese learning children (Imai et al., 2008;Snape & Krott, 2018). The current study also found that the 3-year-olds' fast-mapping performance predicted their manner extension performance for the broad verb but not for the narrow verb. This finding should be due to the exceptional difficulty of extending a narrow verb to a manner change, suggesting that the association between fast-mapping and extension may also differ according to exemplar variability.
The three explanations may not be mutually exclusive. Furthermore, the initial under-extension of manner verbs may be a mechanism that young children use to minimize promiscuous extension that is likely to be wrong. There are two possible developmental patterns in semantic acquisition: (a) language learners initially underextend and then broaden a verb's meaning; and (b) they initially overextend and then progressively narrow down a verb's meaning. The second approach is referred to as the hypothesis elimination approach (Xu & Tenenbaum, 2007), which, as applied to acquisition of verb meaning, proposes that the act of learning consists of eliminating incorrect hypotheses. Logically, learners who have initially overextend need negative evidence to revise their initial hypothesis. However, negative evidence may not necessarily be available in the input. This study found a pattern of underextension in verb meaning acquisition-a pattern that should apply to verbs in general (e.g., path verbs, instrument verbs) and even words of other form classes (e.g., nouns, adjectives). Notably, children may also overextend word meaning (e.g., Clark & Clark, 1977;Rescorla, 1980). However, overextensions may decrease with child age for four reasons. First, overextensions are related to younger children's imitations in vocabulary, which inevitably increase the chance of using a familiar word to label novel objects (or actions, properties) that cannot be labeled by that word. Second, overextensions may decrease with children's growing semantic knowledge of their native language. For example, understanding that manner is an essential component of verb meaning in English may help English-learning children realize that even slight differences in manner can require another verb. Third, overextensions may decrease with the knowledge that a language only has a small set of general-purpose words (e.g., thing, stuff, do, get) that can be used to label a wide range of exemplars. Thus, the probability of a new word being a general-purpose word is low. Fourth, overextensions may decrease as children receive corrective feedback from their caregivers and teachers (Chapman & Thomson, 1980). However, it should be noted that parents often accept children's overextensions, which may unintentionally prolong child use of over-extension (Gruendel, 1977).
We found that high exemplar variability affected verb fast-mapping and extension differently, thus reconciling the debate on the influence of exemplar variability on verb learning. Verbs that are easy to fast-map may be difficult to extend. For example, Theakston et al. (2002) examined the 2-to 3-yearolds' acquisition of the different forms of the verb "go" (e.g., go, going, gonna, gone). Children preferred to restrict these meanings to different situations as if they were unrelated. They used "go" to mean movement, as in "I am going in the car"; "goes" to mean belonging, as in "This book goes on the shelf"; and "gone" to mean disappear, as in "Where's that book gone?" Thus, to adequately gauge children's verb knowledge, both fast-mapping and extensions must be probed.
The current finding that 2-and 4-year-olds tend to underextend verbs with highly specific meanings also helps to explain the conflicting findings on Chinese children's verb extension. Chinese children tend to have a verb learning advantage in their early vocabularies (Tardif et al., 2008). For example, there are about 20 Chinese verbs that distinguish between various manners of carrying or holding, such as carrying on the arm and carrying on the back . These carrying verbs are among the earliest words in Chinese-learning children's receptive and productive vocabularies (Hao et al., 2008;Tardif et al., 2008). However, the relative verb learning advantage does not mean a good understanding of verb meaning. For example, Chinese-speaking 3-year-olds tend to underextend highly specific, familiar Chinese carry verbs (Ma et al., 2021;Saji et al., 2011). Perhaps, the meaning specificity of the early acquired Chinese verbs facilitates fast-mapping but hinders verb extension in the young Chinese learners.

Why did manner variability not affect agent and object extension?
We found that child performance in the agent and object extension tests did not differ between verb types, suggesting that manner variability did not affect agent or object extension. Here, we propose three explanations to this finding. First, the effect of exemplar variability on verb extension may be component-specific. Thus, manner variability may not affect verb extensions to an agent or object change. This explanation is supported by the finding that the variability of an action component (agent, Maguire et al., 2008;object, Snape & Krott, 2018) seen in the training affected 2-to 3-year-olds' extension of a newly learned verb to a change of that particular action component. However, Maguire et al. (2008) and Snape and Krott (2018) only tested children on agent and object extension, respectively, leaving the component-specificity of the effect of exemplar variability of verb extension unclear.
Second, perhaps this finding was driven by the English-reared 2-and 3-year-olds' sensitivity to manner change in language learning, since verbs tend to name the manner of an action in English (Slobin, 2001). Thus, when the 2-and 3-year-olds were exposed to manner changes in a language learning task, their attention was biased toward the manner changes rather than the agent and object. This explanation is supported by the past finding that English-speaking 3-year-olds reliably assumed that a new verb labeled the manner of a novel action, while their Spanish-speaking counterparts more often mapped the new verb to the path of the action (Maguire et al., 2010).
Third, the underlying mechanism of agent and object extension may differ from that of manner extension. Supporting this explanation is the fact that manner changes and agent/object changes are usually labeled by words from different parts of speech. For example, in English, a manner change may warrant a new verb label, but an agent or object change is typically indicated by a new noun or pronoun. This explanation is also supported by the exceptional difficulty associated with manner extension related to agent and object extensions found in experimental research (Behrend, 1990;Forbes & Farrar, 1995) including the current study, where only one action component was changed in an extension test. For example, in this study, the target of the agent and object tests are the same action performed by a new agent or on a new object, respectively.
Notably, the current data do not suggest that visual variability is the only or the strongest factor affecting verb meaning construal. Verb meaning interpretation is also affected by other factors, such as the syntactic structure in which a verb appears (Gertner et al., 2006;Lidz et al., 2003;Ma, Fiveash et al., 2020;Ma et al., 2019;Naigles, 1990;Yuan et al., 2012), the input frequency of verb use in child-directed speech , and the extra-linguistic cues in word learning (Imai et al., 2008). Mandarinspeaking children may turn to extra-linguistic cues-such as social cues-to identify the form class of novel words as Imai et al. (2008) found. Removing a 1-sec segment of object-holding from the beginning of a videotaped scene, allowed children to interpret the novel word as a verb, suggesting that Mandarinlearning children are highly sensitive to extra-linguistic cues in distinguishing between nouns and verbs.

Conclusion
This study examined the effect of exemplar variability on verb fast-mapping and extension. High variability could hinder verb fast-mapping but facilitate verb extension in 2-and 3-year-olds. This study is the first to show that variability in the portrayal of action verbs may affect initial verb learning and extension differently under the same experimental conditions. Furthermore, the effect of exemplar variability attenuated between 3 and 4 years of age.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by Institute of Education Sciences (IES) funding (R324A160241).].