When more is less: the impact of multimorphemic words on learning word meaning

Abstract

Monomorphemic words have been found to influence category formation, as they encode one general category and thus activate it more than other related categories in the same lexical network. On the other hand, multimorphemic words can encode multiple categories from the same network by the multiple forms they combine. Superordinate categories are encoded by sub-lexical forms (e.g., affixes), while the entire words encode lower categories in the hierarchical structure. In the present study, we asked whether sub-lexical forms influence the learning of the meaning encoded by the entire word they underlie. We used Semitic-like words where sub-lexical forms (syllabic patterns) encode superordinate categories of manner-of-motion, and the entire words encode lower-level categories (moving characters). In our main experiment, a word-learning test showed that a shared syllabic pattern had a negative effect on the learning of the moving characters encoded by the entire word. This effect was revealed mainly in dimensions related to the superordinate category encoded by the pattern. The effect and its direction are beyond the expectations of enhanced category representations suggested in previous literature. We conclude that the effect of word-form is beyond the specific category they encode and can have different directions at different hierarchical levels.

1. Introduction

Word learning is a process that involves the mental representation of multidimensional concepts. Some dimensions are related to the word-form, and others shape the word meaning. For example, the sequence of phonemes /k/, /æ/ and /t/ constitutes one aspect of the word-form that may refer to a pet cat, while a feature such as ‘whiskers’ may be a conceptual dimension related to the meaning of the word. The presence of each dimension can influence the representation of the other. This was explored empirically in studies that investigated the effect of labels (using novel words such as ‘lorp’ and ‘pim’) on the learning of word meaning (Fulkerson & Waxman, 2007; Johanson & Papafragou, 2016; Lupyan et al., 2007; Lupyan & Casasanto, 2015; Nazzi & Gopnik, 2001; Waxman & Markow, 1995; Welder & Graham, 2006). In these studies, children and adults were required to fit labeled and unlabeled stimuli into given categories.

To date, such studies have focused mainly on one level of categories that are encoded by entire word labels, namely, common nouns. In the real world, words that encode one category can be associated with items belonging to other conceptually related categories. For example, the word ‘rose’ activates the members of the specific category rose and may also activate exemplars of other types of flowers that share similar features and belong to the same superordinate category. This activation raises the question of whether and how a word-form that encodes one category influences the representations of conceptually related categories. Addressing such questions is important for understanding the role of word-forms in the overall conceptual organization process.

The relationship of one word-form with more than one category is significant in the case of hierarchical categories. For example, the label animal that encodes a superordinate category is also related and can be used to refer to categories such as mammal and dog. These categories are distinct enough to be encoded by independent labels but still share conceptual dimensions, especially since the same exemplars can participate in their learning and representation. Nevertheless, the effect of word-form on categorization has been investigated only on one level of representation, either basic or superordinate (Blewitt & Krackow, 1992; Gervits et al., 2023; Johanson & Papafragou, 2016; Lupyan et al., 2007; Waxman, 1990), expanding our understanding of how the word-form influences category learning. Given the complexity of mental representations and the relationship of one word-form with multiple categorical levels, here we aim to shed light on the effect of word-forms that encode one categorical level on learning and representing other levels in the same hierarchical structure.

To investigate representation at multiple levels, we rely on forms that simultaneously encode different levels of meaning. For example, in the noun library the suffix -ary encodes the meaning of place, while the entire word encodes a category member that is a specific place. Similarly, the Hebrew word mazrék ‘syringe’ encodes the category of instrument via the abstract syllabic pattern maCCéC (C = consonant), while the entire word denotes a specific instrument. That is, in such cases, the category meaning encoded by the pattern that underlies this and other sister words (e.g., mazlég ‘fork’ and masrék ‘comb’) is also encoded by the words as a whole.

In the present study, we explore the effect of the word-form, specifically the sub-lexical form, across representational levels, by inducing learning of novel Hebrew-like multimorphemic words. We rely on the formal and conceptual structure created by the mechanism of word derivation unique to Semitic languages. Thus, we used a shared, novel syllabic pattern to encode a superordinate category across different words while the whole word encoded a lower-level category. Previous studies have shown that speakers of Semitic languages rely on sub-lexical morphemes in segmenting and retrieving words (Bentin & Frost, 2013; Feldman et al., 1995; Frost et al., 1997; Kolan et al., 2011; Omar et al., 2021). However, studies have yet to investigate the role of such forms in the conceptual organization of word meaning. Here, we ask whether syllabic patterns that encode a superordinate category affect the learning of lower-level categories encoded by the entire word.

1.1. The effect of word-forms on category learning

Previous studies have shown that a label shared across entities facilitates category learning (Althaus & Mareschal, 2014; Balaban & Waxman, 1997; Gervits et al., 2023; Johanson & Papafragou, 2016; Lupyan et al., 2007; Nazzi & Gopnik, 2001; Waxman & Markow, 1995). Category learning in such studies was typically evaluated by the responses of participants to new stimuli, which either belong to the same category of the familiarization items or not. Studies on adults as well as on children usually measure category formation by inductive generalization or by explicit judgment about category membership. For example, Lupyan et al. (2007) examined whether giving novel names to aliens that differ in their perceptual features affects the categorization of these aliens according to a conceptual feature (e.g., an alien to be avoided or an alien to be approached). The results of the study demonstrated that adults learned categories encoded by word-forms better than linguistically unmarked categories.

It is well known that categorization is a dynamic process that can yield multiple levels of abstraction. Ambridge (2020) recently suggested that all these levels are exemplar-based. Being exemplar-based means that each category is represented by memories of instances that the individuals encounter in different contexts (Medin & Schaffer, 1978; Medin & Smith, 1981). These memories represent every single exposure to the category members and retain the context and the details of the specific exposure event (Barsalou et al., 1998; Medin & Schaffer, 1978; Medin & Smith, 1981; Nosofsky, 1986, 1988). Along this line, the same exemplars can participate in the learning and representation of more than one level of abstraction. For example, a blue balloon can participate in forming the categories blue and balloon. Establishing multiple levels of categories based on the same exemplars while learning monomorphemic words is possible but not necessary. For example, when the words encode very specific categories that are low in the hierarchal structure, the possibility that the category members are shared with another category is low.

However, the involvement of the same exemplar in establishing multiple categories is inevitable in multimorphemic words. The exemplars involved in learning the word factory in English or its Hebrew counterpart mifʕal (based on the pattern miCCaC, indicating place) also participate in learning the suffix ‘er’ or the pattern miCCaC, respectively. Therefore, investigating this process and the different levels it could involve should take into consideration the effect of word-form on categorization.

Lupyan (2008) was the first to investigate the effect of word-forms beyond the specific category they encode, focusing on exemplar representations. He investigated this using a recognition memory task of labeled items belonging to categories such as chairs and tables. His study examined whether the production of labels affected the perceptual representation of exemplars in cases where a picture of an exemplar was presented. The results showed that overtly classified items with a category name were less recognized than those with unlabeled items. This indicates that the positive effect of the label on category learning might be at the cost of the accuracy of exemplar representation. Based on these results and the idea of multiple-level categories of exemplar-based representation suggested by Ambridge (2020), we expect the word-form effect to be reflected in learning different hierarchical levels of representations.

The theoretical framework that underlies Lupyan’s studies particularly what he terms the Label Feedback Hypothesis can explain how the word-form can influence lower levels of representations. The Label Feedback Hypothesis argues that labels produce an online modulation of exemplar representations. The word-forms activate the category they encode. The activated category often includes dimensions beyond the specific exemplar. This activation leads to a flow of information from the category level to the exemplars, which augment their perceptual representations (Lupyan, 2008; 2012). For example, naming an item as belonging to the basic-level category ‘chair’ results in the activation of category dimensions. These properties flow from the activated category and influence the representation of the new exemplar. This top-down effect was also found in neurophysiological studies supporting the Label Feedback Hypothesis. They found that labels decrease the perceptual distinctiveness between objects that belong to different conceptual categories as a result of enhanced activation of category dimensions (Jouravlev et al., 2018) and they impact visual processing and visuo-cortical responses within a 6-minute learning period (Kutlu et al., 2023).

Based on the assumption that the categories themselves are exemplar-based, we expect that an enhancement of category dimension at the exemplar representation influences not only the exemplar itself but also the other categorical level (e.g., furniture) to which this exemplar is linked. This effect can be reflected in how well the category dimensions are learned as part of the category member (e.g., a table has legs). It can also be found in tasks that require discrimination between members of the same category (a blue balloon and a red balloon), as highlighting category dimensions increases similarities between the category members and makes them less distinct (Althaues & Mareschal, 2014; LaTourrette & Waxman, 2020).

In the current study, we are interested in exploring this effect as reflected when learning multimorphemic words. Sub-lexical forms (here, Semitic syllabic patterns) that highlight dimensions of the superordinate level in the exemplar representation may influence the representation of the category encoded by the entire word, especially since these categories can be represented by the same exemplars. Thus, we will examine both whether and how sub-lexical forms influence the representation of the category dimension in particular, and whether and how this effect is reflected in discriminating members of the same superordinate category.

1.2. The structure of Semitic words

Most Semitic words are assumed to combine two interleaved morphemes: a consonantal root and a syllabic pattern (Aronoff, 2007; Berman, 2003; Ravid, 2003, 2006; Shimron, 2003). These morphemes do not appear individually as surface phonetic forms but only in combination with each other. Each morpheme carries semantic information that can be associated with different conceptual categories (Bentin & Frost, 2013; Berman, 2013; Ravid, 2003). Consonantal roots are typically viewed as carrying the semantic core of the Hebrew word (see Ravid, 2003), whereas patterns serve various functions. Several nominal syllabic patterns encode superordinate categories (Shatil, 2006), while the words themselves encode lower categories in the hierarchical structure. For example, in addition to the patterns used for deriving instrument names and place names (introduced above), the Hebrew pattern CaCéCet encodes a category of illnesses, to which belong words such as cahévet ‘hepatitis’ and xazéret ‘measles’. In these words, the superordinate category is marked by a specific element of the word-form, the shared pattern. At the same time, the entire word encodes another category in the same hierarchical structure: a specific illness. In learning words with this structure, the learning of the superordinate category encoded by the syllabic pattern is based on exemplars that are also involved in forming the basic-level category encoded by the entire word. This creates two levels of representations within the single word that overlap phonologically by the identical vocalic pattern as well as conceptually by the shared exemplars. Here, multimorphemic words with a similar structure will be compared with monomorphemic words that encode only one categorical level, where the superordinate category is not formally marked.

Back to the Label Feedback Hypothesis, we expect activating the superordinate dimensions in multimorphemic words to influence the representation of the same dimensions at the exemplar level. This activation may lead to a better learning of superordinate dimensions when they are encoded by the syllabic pattern in the multimorphemic word compared with the same dimensions when they are encoded by the monomorphemic words. This effect can also be reflected in decreased discrimination between members of the same category encoded by words that share the same syllabic pattern.

1.3. The current study

The current study aims to explore the effect of the syllabic pattern underlying words on learning the categories that these words encode. This is investigated in the first experiment presented below. This experiment included two different conditions – what we term the multimorphemic word condition and the monomorphemic word condition. These conditions include words that encode specific entities that share specific categories of manner-of-motion (skipping vs. flipping). We chose this category based on linguistic and conceptual considerations. The categories of manner-of-motion are not encoded morphologically in Semitic languages, so that previous knowledge is less likely to influence learning. Moreover, these categories are related to visual perceptual features that can be perceived bottom-up from the learning environment. Learning members of the manner-of-motion categories involves learning the relevant features of a complex motion event (Slobin, 2005, 2006; Talmy, 1991). Each category is a combination of several features, such as motor pattern (skip versus flip) and rate (single versus multiple occurrences).

In the multimorphemic word condition, the category manner-of-motion is encoded by the syllabic pattern. That is, the word-forms simultaneously encode the superordinate category and a specific moving entity. In the monomorphemic word condition, the shared manner-of-motion is not encoded via the word-forms used. Any effect of the word-form on the exemplar representation is expected to be found in enhanced learning of the manner-of-motion in the multimorphemic word condition, and a high degree of confusion between members that belong to the same category of manner-of-motion. The categories of manner-of-motion used in this study, the patterns they encode and the word they underlie were found to be learnable in a previous study that used the same paradigm (Omar et al., 2021).

Encoding the category manner-of-motion by the syllabic pattern creates phonological similarity between the words. To rule out the effect of this phonological similarity on word learning, we designed the second experiment presented below. This experiment also included two learning conditions – the phonological similarity condition and the no-similarity condition. These conditions include the same word-forms used in Experiment 1. In the phonological similarity condition, the words share the same syllabic pattern and vowels. However, in this experiment, the patterns do not encode any meaning. That is, the words partially overlap phonologically and do not overlap conceptually. In the no-similarity condition, the words are phonologically and conceptually distinct, with no overlap between either form or meaning.

2. Experiment 1

In this experiment, we explored word learning in the multimorphemic and monomorphemic word conditions, followed by generalization tests of the superordinate categories. Word learning test examines whether encoding a superordinate category by the syllabic pattern affects the learning of lower categories encoded by the entire words. This test examines both learning the category dimensions as part of the concept and the discrimination between members of the same category, using different types of trials. Generalization was assessed by asking whether learning the manner-of-motion of the entities in the exposure phase affords the learning of new members of the category (Conceptual generalization). Finally, following the multimorphemic word condition, the relationship between these categories and the syllabic patterns is tested (Morphological generalization).

2.1. Participants

Forty-eight native Hebrew-speaking adults participated in this experiment (28 females, mean age: 26, SD = 4). All the participants were students at the University of Haifa and were paid for their participation. By self-report, all participants had normal hearing and no history of neurological disorders or learning disabilities. All aspects of this study were approved by the Ethics Committee of the University of Haifa.

2.2. Stimuli

The stimuli were inspired by the word learning paradigm used in Banai et al. (2020). Two sets of six novel multimorphemic words were devised (see also Omar et al., 2021) and introduced to the participants in the learning phase. The words in the two sets consist of two pseudo-morphemes: a root and a pattern that do not exist in Hebrew but are based on the consonants, vowels and phonotactic properties of Hebrew (e.g., xutirúk: consists of the root x.t.r.k and the pattern CuCiCúC). Each word had a unique pseudo-root. The frequency of each four-root consonant was balanced across the two sets so that the two sets had the same consonants in similar proportions. The pseudo-syllabic patterns used in the two sets shared the same prosodic pattern Cv.Cv.CvC (C = consonant, V = vowel) and an ultimate stress pattern. The two sets included the same number of roots and different numbers of syllabic patterns (two patterns versus six patterns), creating two learning conditions of different type-token frequencies.

In one set, the words were based on six pseudo-patterns CiCoCúC, CuCiCúC, CoCaCéC, CeCoCáC, CuCiCeC and CiCuCáC. That is, the words were unrelated phonologically. These word-forms were used in the monomorphemic word condition and provided unique names for six cartoon characters (see Appendix A). Although the words in this set were composed using the nonlinear mechanism of Semitic word formation, we consider them monomorphemic since they did not share any sub-lexical unit. The named characters were divided into two categories of manner-of-motion (skipping or flipping) of three members each. In this condition, the categories of manner-of-motion are not encoded morphologically.

In the second set, the words were based on only two different pseudo-patterns, CiCuCáC and CuCiCúC, so every three words shared the same syllabic pattern. The pseudo-roots were not shared between the sets. This set of word-forms was used in the multimorphemic word condition to name two groups of characters. In this condition, the characters belong also to two categories of manner-of-motion (skipping or flipping). However, here the concept of manner-of-motion was encoded morphologically. Characters that moved in the same manner-of-motion also had names that shared the same syllabic pattern (e.g., xutirúk, kumilúf and nurivús shared the syllabic pattern CuCiCúC and were names of flipping characters).

The two sets of words created two learning conditions with different type-token frequencies. While the multimorphemic word condition included two syllabic patterns, each shared by three names (1:3 type-token frequency), the monomorphemic word condition included six different patterns, each unique to a specific name (1:1 type-token frequency).

Note that proper names in Hebrew and other Semitic languages such as Arabic are often decomposable multimorphemic words, constructed nonlinearly by combining roots and patterns. In fact, many proper names are derived from common nouns in the language. Examples for this can be found in names such as Simxa, Bracha and Tikva (lit. Joy, Blessing and Hope). Proper names also allow for categorization based on the characteristics of the person, for example, their gender, which is morphologically marked (e.g., via the morpheme ‘-a’ in the feminine names Sara and Rivka in Hebrew, or Samira, Samiḥa and Saʕida, in Arabic, as well as ‘-e’ in the Arabic names Malake, Fatme and Amne). Furthermore, color names in Hebrew such as Ɂadóm (red), yarók (green), cahóv (yellow) and kaxól (blue) share a pattern (CaCóC) and so do the equivalent color words in Arabic, Ɂáḥmar, Ɂáxdar, Ɂásfar and Ɂázrak, that share their own unique pattern (ɁáCCal). Other salient examples of multimorphemic names in Semitic languages are place names, family names and even names for illnesses. Cahevét (jaundice), šaxefét (tuberculosis) and Ɂademét (rubella) are distinct illnesses that are derived from color names and share the pattern (CaCeCét). That is, proper names are highly informative in these languages.

In addition to the two sets of words used in the exposure session, six pseudo-words were used in the Morphological generalization test to examine the generalization of the two patterns used in the multimorphemic word condition ‘CiCuCáC’ and ‘CuCiCúC’ (see Appendix A).

2.3. Design

2.3.1. Exposure phase

During the exposure session, participants were exposed to the novel words either under the multimorphemic word or the monomorphemic word condition using an interactive narrated video story (a Birthday Party story or a Picnic story). The two stories were counterbalanced across conditions: half of the participants of each condition were exposed to the Birthday Party story, and the other half were exposed to the Picnic story. The video scripts of the stories were designed with MIT’s Scratch software, and the audio stories narrated by an adult female native speaker were recorded using ‘Audacity’. The duration of each movie was 4–5 minutes. The characters and their names were repeated ten times in the different episodes of the story to induce the learning of both the superordinate categories and the specific moving characters based on the same exemplars (see examples of the exemplars in Appendix B). To maintain the participants’ attention during the story, they were required to click on different objects or characters.

2.3.2. Testing phase

Following the exposure session of each learning condition, word learning and category generalization were tested.

2.3.2.1. Word learning test

Learning the moving characters encoded by the entire words was examined with a three-alternatives-forced-choice (3AFC) test. In this test, the participants heard the question ‘Who is ____?’ and were asked to select the correct answer out of three moving characters that appeared simultaneously on the screen (see Figure 1). The three characters remained visible until a choice was made. The following question was heard immediately after the participant answered by clicking on one option. No feedback was provided during the test.

Figure 1. Example of an Appearance, Motion and Combination trial in the Word learning test. Participants hear the question ‘Who is siguváŝ?’ and three different moving characters appear on the screen.

The test included three types of trials that require the identification of different dimensions of the category members encoded by the entire words: (a) Appearance trials test the learning of the visual properties of the moving characters; (b) Motion trials test the learning of the manner-of-motion of each moving characters; and (c) Combination trials simultaneously test the learning of the appearance and manner-of-motion of each character. The distractors in Appearance trials shared the same manner-of-motion with the target character. This allows testing the discrimination between members of the same category of manner-of-motion based on the character’s appearance. The distractors in Motion trials shared the same appearance as the target. This allows testing the learning of each manner-of-motion as part of the specific moving character. In Combination trials, one of the distractors shared the same appearance with the target and moved in a different manner-of-motion, while the other distractor moved in the same manner-of-motion and had a different appearance. The six characters’ names were presented nine times in the test, equally divided across the three types of trials. Each test included 18 trials of each type of trial, for a total of 54 trials in the entire test.

2.3.2.2. Conceptual generalization test

This test examined the generalization of manner-of-motion categories (skipping and flipping) to new items. The test included six newborn moving characters, each appearing in six different trials for a total of 36 trials. During the test, each character appeared on the left side of the screen, while at the same time two stationary characters from the exposure session appeared on the right (and were larger). The participants were required to select the ‘big brother’ of the new moving character out of two stationary pictures (see Figure 2). The stimuli of each trial remained visible on the screen until a choice was made. The next trial appeared immediately following the participant’s click on a character. No feedback was provided during the test.

Figure 2. Example of a trial from the Conceptual generalization test. On the left – the newborn moving character. On the right – two stationary characters from the exposure phase.

2.3.2.3. Morphological generalization test

This test examined the generalization of the association between the syllabic patterns and the meanings of skipping and flipping (the two novel morphemes designed for the study) to words that were not part of the exposure phase. The participants heard new multimorphemic names and were required to select a matching character out of three moving characters that were not encountered before and that appeared simultaneously on the screen (see Figure 3). The six new names consist of the syllabic patterns introduced in the exposure phase, CiCuCáC, and CuCiCúC, encoding skipping and flipping, respectively. Six novel consonantal roots were used (e.g., CiCuCáC and x.g.v.l yield the word xiguvál). Each name was presented six times for a total of 36 trials. The transition between the trials was immediately after the participant’s response, and no feedback was provided during the test.

Figure 3. Example of a trial in Morphological generalization test. Participants hear ‘Who is xiguvál?’ and three different moving characters appear on the screen.

2.3.3. Experimental procedures

Participants were assigned randomly to one of the two conditions (multimorphemic word and monomorphemic word). Each participant completed an exposure session followed by a test session, in which all three tests were presented in a fixed order presented above. The test was conducted with headphones in a quiet room in the Auditory Cognition Lab at the University of Haifa.

2.3.4. Data analysis

The response to each trial in the three tests was coded as correct (1) or incorrect (0). For the full dataset, please refer to https://osf.io/uk58a/?view_only=b110951dff7946f0869a75f144daf1cd. First, we assessed whether the performance of the participants was above the chance level in word learning and generalization tests. Second, we modeled the participants’ performance in the Word learning tests to assess whether the learning conditions influence word learning. We used mixed-effects logistic regressions (Jaeger, 2008) using the lme4 package in the R environment (Bates et al., 2014). We considered the models with random intercepts for participants and trials, as well as by-participants and by-trial random slopes of the main effects. To measure the main effects and interactions, we used sum-contrast. That is, the effect of all factors and interactions is not compared to a baseline level of each variable as in treatment contrast commonly used in logistic regression, but rather to help understand the interactions, they are compared to an overall mean of the dependent variable. In sum-contrast, positive estimates in the model indicate that the corresponding factors have a positive effect on the performance and vice versa. Both marginal and conditional R-squared are reported for each model. Marginal R-squared represents the proportion of variance explained by the fixed factors in a model, while conditional R-squared provides an estimate of the proportion of variance explained by both fixed and random factors combined (Nakagawa & Schielzeth, 2013).

2.3.5. Results

2.3.5.1. Word learning test

The performances of the participants in the Motion and Combination trials were similar in the two learning conditions. In the monomorphemic word condition, the proportion of correct answers was 0.77 ± 0.02 in Motion trials and 0.75 ± 0.02 in Combination trials (χ2(1) =0.4, p = 0.52). In the multimorphemic word condition, the proportion of correct answers was 0.62 ± 0.02 in Motion trials and 0.66 ± 0.02 in Combination trials (χ2(1) = 0.25, p = 0.62). The participants’ high performance in Appearance trials (0.92 ± 0.013) and their similar performance in Motion and Combination trials indicate that the choice of the correct answer in Combination trials was based on the identification of motion. Furthermore, analyzing the type of errors in Combination trials shows that 92% of the incorrect choices in the multimorphemic word condition and 94% in the monomorphemic word condition were for distractors that share the same appearance with the target. This supports the idea that Combination trials examined motion learning. Thus, in the analysis of the results, we combined Motion and Combination as they reflect identification by motion, taking into consideration the different chance levels in the two types of trials (0.33 for Motion trials and 0.5 for Combination trials).

As shown in Figure 4, participants recognized both the appearance and the manners of motion of the characters well above chance (monomorphemic word: Appearance: $ \hat{\beta} $ = 6.18, Z = 6.62, p < 0.001; Motion: $ \hat{\beta} $ = 2.37, Z = 6.96, p < 0.001; multimorphemic word: Appearance: $ \hat{\beta} $ = 6.9, Z = 4.64, p < 0.001; Motion: $ \hat{\beta} $ = 1.67, Z = 4.71, p < 0.001), suggesting that these two dimensions were reliably learned following exposure. Appearance learning was similar in the two conditions (monomorphemic word: 0.92 ± 0.012; multimorphemic word: 0.92 ± 0.013). Participants’ performance in learning the motion of the characters was lower than in learning their appearances, with an advantage in the monomorphemic word condition (monomorphemic word: 0.75 ± 0.015; multimorphemic word: 0.64 ± 0.016).

Figure 4. Word learning test. Accurate identification of the appearance and the manner-of-motion of the moving characters across the two conditions – monomorphemic word and multimorphemic word. Box edges mark the interquartile range, the line within each box marks the median and the whiskers mark 1.5 times the IQR. The horizontal line marks the chance level (0.33).

To assess the effect of the shared syllabic pattern on learning the members of the two categories (skipping and flipping), we modeled the learning of the appearance of the characters and their manner-of-motion in the multimorphemic word vs. the monomorphemic word condition. The full model included the fixed effects of Condition (multimorphemic word and monomorphemic word), Type-of-Trial (Appearance and Motion), Conceptual Category (skipping and flipping) and all their pairwise interactions. In the model, two random slopes were included: Conceptual Category by-participant and Condition by-trial. The random slopes that account for the by-participant effect of Type-of-Trial were not included in this model. In our study, we tested word learning after exposure to two different conditions. We expect that the different conditions will create internal differences within each participant that could be reflected in their performance in different types of trials. Thus, including these random slopes could mask the effect created deliberately in the exposure phase, where participants were exposed to different conditions, and may interfere with answering the research question.

The full model displayed in Table 1 was compared with a model that does not include the independent variable Condition and its interaction with other fixed variables. The full model showed significant improvement over the model without Condition ( $ \Delta AIC=2.6 $ , χ2(3) = 8.6, p < 0.01), indicating that the condition of learning influences the learning of the words. The model shows a significant negative effect of Type-of-Trial ( $ \hat{\beta} $ = −1.15, Z = −10.23, p < 0.001), indicating that the appearances of the characters were learned significantly better than their manners of motion. This effect was influenced by the manner-of-motion these characters have, as shown by the significant interaction term between Conceptual Category and Type-of-Trial ( $ \hat{\beta} $ = −0.35, Z = −3.25, p < 0.01). This indicates that the advantage of learning the characters’ appearance over their manner-of-motion tended to be more pronounced in the flipping category. Furthermore, our results showed no significant main effect of Condition on word learning ( $ \hat{\beta} $ = −0.18, Z = −0.86, p = 0.39). However, word learning was influenced by the interaction term between Condition and Type-of-Trial ( $ \hat{\beta} $ = −0.19, Z = −2.22, p < 0.05). The negative sign of this interaction indicates that in the multimorphemic word condition, the learning of the manner-of-motion of a given character was significantly worse than other types of trials across the conditions.

Table 1. Fixed effects of the logistic regression model predicting learning of the Appearance and Motion of exposed exemplars from two conceptual categories skipping and flipping at the two conditions – multimorphemic word and monomorphemic word. Model formula: Response ~ Type_of_Trial*Conceptual Category + Condition*Type_of_Trial + Condition*Conceptual Category + (1 + Conceptual Category | Participant) + (1 + Condition | Trial). Marginal R² = 14.45% and Conditional R² = 41.06%

A further model that includes only Motion trials was fitted to explore whether the difference we found in learning the manners of motion in the two conditions is not related to the participant performance in the Appearance trials. The model included Conceptual Category and Condition as fixed effects. Moreover, we included the random slopes: Condition by-trial and Conceptual Category by-participant. The full model displayed in Table 2 was compared with a model that does not include the variable Condition. The full model showed significant improvement over the model without Condition ( $ \Delta AIC=1.5 $ , χ2(4) = 9.5, p < 0.5), indicating that the condition of learning influenced the learning of the manner-of-motion dimension as part of each character. The model shows a significant effect of Condition ( $ \hat{\beta} $ = 0.41, Z = 2.07, p < 0.05), indicating that learning the manner-of-motion of each character was significantly worse in the multimorphemic word condition when they were encoded morphologically.

Table 2. Fixed effects of the logistic regression model predicting learning of the Motion of characters from two conceptual categories skipping and flipping at the two conditions – multimorphemic word and monomorphemic word. Model formula: Response ~ Condition*Conceptual Category + (1 + Conceptual Category | Participant) + (1 + Condition | Trial). Marginal R² = 2.82% and Conditional R² = 48.42%

2.3.5.2. Conceptual generalization test

In both conditions, participants generalized the two conceptual categories (skipping and flipping, Figure 5) evident by their above-chance performance (Monomorphemic Word Condition: Skipping: $ \hat{\beta} $ = 2.71, Z = 6.32, p < 0.001; Flipping: $ \hat{\beta} $ = 2.38, Z = 6.58, p < 0.001; Multimorphemic Word Condition: Skipping: $ \hat{\beta} $ = 1.77, Z = 4.8, p < 0.001; Flipping: $ \hat{\beta} $ = 2.02, Z = 6.04, p < 0.001). These findings indicate that the manners of motion (skipping and flipping) were learned as distinct superordinate categories of the characters. The proportion of the correct responses in the monomorphemic word condition was better than multimorphemic word condition in the two categories (Monomorphemic Word Condition: Skipping: 0.78 ± 0.02; Flipping: 0.77 ± 0.02; Multimorphemic Word Condition: Skipping: 0.67 ± 0.02; Flipping: $ \hat{\beta} $ =0.72 ± 0.02).

Figure 5. Conceptual generalization test. Accurate generalization of the two conceptual categories skipping and flipping, across the two conditions – the multimorphemic word and monomorphemic word condition. Box edges mark the interquartile range, the line within each box marks the median and the whiskers mark 1.5 times the IQR. The horizontal line marks the chance level (0.5).

To explore the effect of the shared syllabic pattern on learning the conceptual categories encoded by the morphological patterns, we modeled participants’ performance in generalizing the two categories in the multimorphemic word condition versus the monomorphemic word condition. The model included Condition and Conceptual Category as fixed effects. This model showed no significant effect of Condition (β = 0.25, Z = 1.22, p = 0.22), Conceptual Category (β = −0.06, Z = −0.67, p = 0.5), or the interaction between them (β = 0.11, Z = 0.74, p = 0.08).

2.3.5.3. Morphological generalization test

The generalization of each morphological pattern to three different items was tested using a 3AFC test. The mean proportion of correct responses in the skipping category was 0.49 ± 0.02 and in the flipping category 0.46 ± 0.02. The results of the test showed that the participant’s performance in the two morphological patterns exceeded what was expected by chance (Skipping: $ \hat{\beta} $ = 1.05, Z = 4.63, p < 0.001; Flipping: $ \hat{\beta} $ = 0.91, Z = 4.11, p < 0.001). These findings indicate that manners of motion were not only learned as conceptual categories but also as meanings that are encoded by the syllabic patterns.

2.3.5.4. The relationship between pattern generalization and manner-of-motion learning

A further analysis was conducted to explore whether the generalization of the morphological pattern encoding the concept of manner-of-motion, as reflected in the morphological generalization test, predicts the representation of these concepts at the level of the specific character, as reflected in the Word learning test. For this purpose, we modeled the learning of manner-of-motion in the Motion trials of the Word learning test. The proportion of the generalization for each morphological pattern (Skipping generalization and Flipping generalization) in the morphological generalization test, along with the Conceptual Category, were included as fixed effects. The model also incorporated Trials as a random effect and the Conceptual Category by-participant as a random slope.

The model in Table 3 showed a significant effect of the generalization of the skipping manner-of-motion on learning the manner of motion of the characters ( $ \hat{\beta} $ = 4.71, Z = 3.75, p < 0.001), indicating that participants who generalized the skipping manner-of-motion well tended to better learn the manner-of-motion of the specific characters. Additionally, the model revealed a significant negative effect of the interaction between the generalization of the skipping manner-of-motion and the conceptual category ( $ \hat{\beta} $ = −1.15, Z = −10.23, p < 0.001), suggesting that participants who generalized the skipping manner-of-motion tended to perform worse in identifying the manner of motion of skipping characters in the Word learning test.

Table 3. Fixed effects of the logistic regression model predicting learning of the Motion of characters from two conceptual categories skipping and flipping in the multimorphemic word condition. Model formula: Response ~ Conceptual Category*Skipping Generalization + Conceptual Category*Flipping Generalization + (1 + Conceptual Category | Participant) + (1 + Trial). Marginal R² = 19.34% and Conditional R² = 44.65%

2.3.6. Discussion

The major outcome of Experiment 1 is that the shared syllabic patterns across words impeded the learning of the manner-of-motion of specific characters (Figure 4, Table 1). These manners-of-motion were learned as dimensions that allow categorization (Figure 5), and their association with the syllabic patterns was learned and generalized following the multimorphemic word condition. The lower performance in morpheme generalization compared to word learning aligns with the idea that these two processes might engage different memory systems and cortical regions. While the consolidation of a general schema involves identifying commonalities between specific exemplars, this type of learning typically requires more time and repeated exposures (Davis & Gaskell, 2009). The effect of the shared syllabic pattern on the representations of specific characters is consistent with studies that investigated the effect of noun labels on exemplar representation (Katz, 1963; Lupyan, 2008). Because the novel names we used are based on Hebrew-like word structure and the shared syllabic pattern encoded a general category but influenced the learning of the meaning encoded by the entire word, the current finding suggests that word-form can influence not only the categories they encode but also lower categories in the hierarchical structure.

The negative effect of a shared syllabic pattern on word meaning is also consistent with the finding of the previous study conducted on Hebrew-speaking preschool children (Banai et al., 2020). The shared patterns in that study created a phonological similarity between the words without encoding a particular meaning. The results of the study were interpreted to mean that the phonological overlap across different words decreases the distinctiveness between entities and impedes their learning. To address the involvement of phonological similarity in the effect of the syllabic pattern on word learning in the current study, we conducted an additional experiment in which we used the same words without encoding the meaning of manner-of-motion in the syllabic patterns.

3. Experiment 2

This experiment aims to explore the involvement of phonological similarity in the effect of the shared syllabic pattern on word learning that we observed in Experiment 1.

3.1. Participants

Participants were forty-eight adult native Hebrew speakers who did not participate in Experiment 1 (30 females, mean age: 27, SD = 3). By self-report, all participants had normal hearing and no history of neurological disorders or learning disabilities.

3.2. Stimuli

The two sets of words used in Experiment 1 were again used as names of cartoon characters. The same characters were used in the current experiment. However, each character had its distinct manner-of-motion. In one condition (no-similarity), the six stimuli were six words; each comprised of a distinct root and a distinct syllabic pattern and each identifying a specific character. The stimuli in the phonological similarity condition were identical to the no-similarity condition, except that each subgroup of three words shared the same syllabic pattern. Thus, unlike Experiment 1, the concept manner-of-motion was not encoded morphologically in any of the sets.

3.3. Design

3.3.1. Exposure session

Similar to Experiment 1, participants were exposed to the novel words under one of the two conditions (no-similarity and phonological similarity), using the same interactive narrated video stories (a Birthday Party story or a Picnic story) counterbalanced across conditions, with the appropriate modifications to the stimuli. The names of the characters were repeated ten times, and each moving character appeared seven times in the different episodes of the story.

3.3.2. Testing session

Following the exposure session of each learning condition, word learning was tested using a 3AFC test similar to the Word learning test we used in Experiment 1.

3.3.3. Procedure

Participants were randomly assigned to one of the two conditions. All participants were tested with headphones in the Auditory Cognition Lab at the University of Haifa.

3.3.4. Results

Mixed-effects logistic regressions were used to examine whether words were learned and to assess the effect of phonological similarities on this learning (Jaeger, 2008). As shown in Figure 6, participants learned the appearances of the moving characters in the two learning conditions (the proportion of correct answers in the no-similarity condition 0.97 ± 0.008 and in the phonological similarity condition 0.97 ± 0.008). This performance is more than expected by chance (no-similarity: $ \hat{\beta} $ = 5.46, Z = 6.83, p < 0.001; phonological similarity: $ \hat{\beta} $ = 35.2, Z = 6.72, p < 0.001). The learning of the specific motion of the characters (no-similarity: 0.74 ± 0.015; phonological similarity: 0.76 ± 0.015) was worse than learning their names but above the expected by chance (no-similarity: $ \hat{\beta} $ = 2.85, Z = 8.16, p < 0.001; phonological similarity: $ \hat{\beta} $ = 2.75, Z = 7.98, p < 0.001).

Figure 6. Word learning test. Accurate identification of the appearance and the manner-of-motion of the moving characters across the two conditions – no-similarity and phonological similarity. Box edges mark the interquartile range, the line within each box marks the median and the whiskers mark 1.5 times the IQR. The horizontal line marks the chance level (0.33).

To assess the effect of phonological similarities on word learning, we modeled the learning of the characters’ appearances and their specific manner-of-motion in the phonological similarity and the no-similarity conditions. The full model included Condition and Type-of-Trial as fixed effects. Trials and the by-participant random slops of the learning condition were included as random effects. This model was compared with another model that does not include Condition and its interaction with other fixed variables. No difference was found ( $ \Delta AIC=-3.9 $ , χ2(2) = 0.04, p = 0.98). An estimate of the full model is displayed in Table 4. The model demonstrates that the learning of participants is negatively affected by Type-of-Trials ( $ \hat{\beta} $ = −1.43, Z = −10.3, p < 0.001). This finding indicates that learning the characters’ appearances was significantly better than learning their motions. The model also showed that the learning condition does not affect learning ( $ \hat{\beta} $ = 0.05, Z = 0.18, p = 0.85). Moreover, the interaction between Condition and Type-of-Trial was insignificant, and thus no evidence of condition effect on learning manners-of-motion as features of the words was found ( $ \hat{\beta} $ = 0.002, Z = −0.02, p = 0.99). That is, phonological similarity across the Hebrew-like words used in the present study did not affect their learning.

Table 4. Fixed effects of the logistic regression model predicting the learning of the appearances and the manners of motion of the characters at the two conditions – no-similarity and phonological similarity. Model formula: Response ~ Condition*Type-of-Trial + (1 | Participants) + (1 + Condition | Trial). Marginal R² = 15.77% and Conditional R² = 41%

Comparing the learning of the manner-of-motion across the conditions of the two experiments (see Figure 7) shows that the performance of the participants in the multimorphemic word condition was significantly worse than the other conditions ( $ \hat{\beta} $ = −0.55, Z = −2.09, p < 0.05, see Table 5). This finding indicates that the negative effect of the syllabic patterns of motion learning we found in the multimorphemic word condition cannot be attributed to the phonological similarity they create between words. The model included only the Motion trials, Condition as an independent variable, Participant as a random variable and Trial by Condition as a random slope.

Figure 7. Motion trials in Word learning test. Accurate identification of the manner-of-motion of the moving characters across the four conditions – monomorphemic, multimorphemic, no-similarity and phonological similarity. Box edges mark the interquartile range, the line within each box marks the median and the whiskers mark 1.5 times the IQR. The red point marks the mean.

Table 5. Fixed effects of the logistic regression model predicting the learning of the manners-of-motion of the characters at the four conditions of the two experiments – monomorphemic, multimorphemic, phonological and no-similarity. Model formula: Response ~ Condition + (1 | Participants) + (1 + Condition | Trial). Marginal R² = 1.85% and Conditional R² = 43.18%

3.3.5. Discussion

The results of the Word learning test in Experiment 2 showed that a shared syllabic pattern that does not encode any meaning did not affect the learning of the categories encoded by the word. Similar results were found when all the conditions were compared, as only in the multimorphemic word condition motion was learned worse than in other conditions. These findings are not consistent with Banai et al. (2020), perhaps due to developmental factors. Studies have shown that auditory information, including speech sounds, overshadows visual information when both are presented simultaneously to children. This auditory overshadowing impedes the discrimination of visual stimuli, so entities that share similar labels become less distinct. This overshadowing effect was not found in adults (Napolitano & Sloutsky, 2004; Sloutsky & Napolitano, 2003). Thus, the results of Experiment 2 indicate that phonological similarities by themselves do not influence the learning of word meaning when these similarities are not related to any semantic information.

4. General discussion

The present study aimed to explore whether a word-form that simultaneously encodes general and more specific categories at the same hierarchical structure affects the learning of word meaning. We used multimorphemic Semitic-like words to investigate the effect of complex word structure on categorization and word learning. Three findings are of note: first, consistent with a previous study (Omar et al., 2021), conceptual regularities across exemplars in Experiment 1 allowed for learning the manners-of-motion of the individual characters and the generalization of this concept to new entities. That is, the current study successfully replicated the results of our original experiment, confirming the validity of the methodology. This indicates that despite the novelty of the learning paradigm we employed (with the aim to minimize the influence of prior knowledge), the concept of manner-of-motion, typically syntactically expressed through adverbial forms, was learned and generalized as a category of entities when morphologically encoded by novel nouns. Second, the novel comparison between the monomorphemic and multimorphemic word conditions allowed us to explore the effect of sub-lexical forms on categorization. A shared syllabic pattern that encoded the superordinate category manner-of-motion impeded the learning of the category-related dimension (skipping or flipping) at the level of the specific moving characters. Third, although the shared syllabic pattern negatively affected word learning in Experiment 1, Experiment 2 showed no evidence of this interference when the same patterns created only phonological similarities between conceptually distinct words.

The effect of word-form we found in Experiment 1 is consistent with the findings of previous studies on monomorphemic noun labels that encode whole categories (Lupyan, 2008; 2012). However, the way multimorphemic words influenced learning countered our predictions, given previous findings. Section 4.1 will consider the learning of the multimorphemic words in our experiments and the word-form effect we found in light of the exemplar-based approach, to illustrate the process of constructing the representation of words with a complex internal structure. Moreover, we discuss our findings in relation to the expectations of the word-form effect suggested by the Label Feedback Hypothesis, the only theory that takes into consideration the effect of labels on lower-level representations (Lupyan, 2012). Section 4.2 will discuss how hierarchical categories encoded by multimorphemic words are represented based on our results. In Section 4.3, we will suggest an explanation of the effect of word-form we found from a perspective that considers the interconnection between hierarchically linked categories encoded by multimorphemic words.

4.1. On categorization – the view from multimorphemic word effects

Let us consider a possible theorization of the process of learning multimorphemic words of the type used in the present study, following principles of the exemplar-based approach (Barsalou, et al., 1998; Medin & Schaffer, 1978; Medin & Smith, 1981; Nosofsky, 1986; 1988, among others). The exemplars involved in learning these words combine more than one dimension (in our case, we focus on the dimensions of visual appearance, manner-of-motion, quadri-consonantal root and syllabic pattern). During the learning process, these dimensions allow complex categories to form, as each dimension contributes to the process of analogizing each stimulus with similar representations. The exemplars of the present study were categorized simultaneously into (at least) two categorical levels. A stimulus such as Xutirúk was compared and categorized (following initial exposure) with (1) exemplars that are labeled by the same combination of root and pattern (x.t.r.k + CuCiCúC) and that share the same visual appearance and manner-of-motion (constituted from the feature rate and motor pattern, see Slobin, 2005, 2006; Talmy, 1991) and (2) exemplars that are labeled by the same pattern but with a different root (e.g., k.m.l.f + CuCiCúC) and that share only the same manner-of-motion (see Appendix B for examples of the exemplars in each category). This creates a network of connected exemplars where each exemplar is activated in categorization events at different levels: a specific level represented by exemplars described in (1) and the superordinate levels represented by the exemplars described in (2).

As explained at the outset of this paper, our unique data allow us to explore the interconnection between the different levels of representation. Specifically, we consider our results in light of Lupyan’s (2008, 2012) suggestions regarding the possible effect of the category activation on the exemplar representation, which is a lower level of abstraction, and whether these suggestions apply in our case. Given that in our data the same exemplars were involved in the representation of the two hierarchically linked categories of the learned words, we expected that any effect on the exemplar representation should be reflected at all categorical levels (as in the case of the word-form effect, as suggested by Lupyan). The expected effect is an enhancement of dimensions related to the general category (e.g., manner-of-motion) in the exemplar representation when this category is encoded by the syllabic patterns. Our findings indeed showed that the representation of the category-related dimensions was affected at the level of the moving character (see Word learning task), confirming an interconnection between the levels. However, the direction of the effect we found contradicted the expectations derived from Lupyan’s suggestion. Syllabic patterns impede the representation of these dimensions.

One could argue that these findings result from using novel rather than existing categories. The enhancement of the superordinate dimension suggested by the Label Feedback Hypothesis is assumed to occur after a relationship has been established between the word-form and the category meaning. This effect has been reported in the context of novel word learning, particularly when the training session allowed for learning form-meaning association (Lupyan, 2007). However, in our study, such a relationship was indeed established, as indicated by the Morphological generalization test. In fact, a previous investigation using the same experimental paradigm revealed that already following only five exposures to these multimorphemic words (half the number of exposures in the present study) Hebrew-speaking participants were able to learn the relationship between the syllabic patterns and their meaning (Omar et al., 2021).

The discrepancy between the word-from effect expected by the Label feedback Hypothesis and the way the moving characters were represented in the present study indicates that the effect of the word-forms is not unidirectional across categories. Rather, the direction of the effect seems to depend on the level of representation that is being tested. The fact that the manner-of-motion was not highlighted in the representation of the moving characters at the multimorphemic word condition, as would have been expected by the Label Feedback Hypothesis, revealed that the way a general category is activated by the entire label (Lupyan, 2008; 2012) is different from the way exemplars of lower categories were activated by the syllabic patterns.

With this suggestion, we now go back to the assumption of the exemplar-based model, namely, that the exemplar is activated at the superordinate as well as lower levels of the same hierarchical structure. The discrepancies between our findings regarding the effect of superordinate-level word-forms (the syllabic pattern) on the moving characters representation and the findings reported by Lupyan, which examined the interconnection of exemplars within the same category, require that we reexamine this assumption, specifically with respect to the representation of hierarchically linked categories.

4.2. On the representation of hierarchically linked categories

Exemplar-based models have yet to theoretically and empirically tackle the representations of hierarchically linked categories and how the same exemplars are represented at different levels (Murphy, 2016). These models suggest analogy processes between a new instance and existing representations that vary in similarity and thus create multiple categories that differ in their structure (e.g., Medin & Smith, 1981; Nosofsky, 1986; 1988). Murphy (2016) calls for studies that would pay more attention to issues such as the representation of hierarchical categories in exemplar-based models. For example, he suggests that representing each squirrel we encounter along with all its categories (e.g., mammal and animal) would be inefficient (especially since he himself seldom thinks about a squirrel as a mammal or animal when he encounters an exemplar of this category). Murphy also argues that exemplar-based models cannot explain the basic-level advantage, as the identification of entities that belong to the basic level category (e.g., dog) is faster than superordinate (e.g., mammal) and subordinate (e.g., bulldog) levels (Murphy & Brownell, 1985; Murphy & Lassaline, 1997). Since exemplar models assume an analogy between a new instance and existing exemplars and not the whole category, when the same exemplars exist at the different hierarchical levels of categories, there is no reason to access one level faster than the others.

Recently, Ambridge (2020) suggested that exemplar representations are re-represented at different levels of abstraction. This theoretical view may explain the faster access to one categorical level relative to the other. Assuming that the representation of the exemplar changes with abstraction, the more the categorical level is similar to the stimulus, the faster it is expected to be accessed. An additional aspect that should be considered is the hierarchically linked categories discussed by Murphy, which can be learned based on different types of information, depending on the level of the category. While the basic-level category ‘squirrel’ can be learned bottom-up based on the processing of perceptual exemplars that share similar attributes, members of superordinate categories share a general framework that is correlated with fewer perceptual similarities (see Rosch et al., 1976; Murphy, 2004). Thus, some superordinate categories such as animals or mammals are often learned based on explicit verbalizable knowledge (e.g., mammals give birth) that may but does not have to be accompanied by an encounter with perceptual exemplars (see Brooks & Kempe, 2020 on exemplar-based models and explicit learning of categories).

These categories differ from the hierarchical categories encoded by the multimorphemic words we investigated, which were learned naturally and implicitly from the linguistic input. In our study, the participants learned two hierarchical categories bottom-up based on the same exemplars as shown by both the Word learning and the Morphological generalization tests. Although the exemplars of the manner-of-motion categories (skipping and flipping) also share a general framework of the motion event, this framework is correlated with the attributes of rate, and motion patterns that are perceived from the exposure to the moving characters. Nevertheless, our findings validate Murphy’s questions regarding the exemplar representations across hierarchical categories. They indicate that even when the same exemplars were involved in learning two hierarchical categories, these exemplars may have been represented differently at the superordinate- and lower-level categories.

Differences in the way exemplars are represented across the categorical levels were reflected in the present study in the learning of the manner-of-motion dimension. We suggest that these differences are related to the function of the word-form at the different levels. At the more specific level (the moving characters) investigated in our study, the syllabic pattern appeared consistently with both dimensions: manner-of-motion and visual appearance in all the exemplars (e.g., bizudax always had the same visual form and moved in a skipping manner-of-motion). Thus, there is no reason that at this level of categorization, the syllabic patterns would activate the manner-of-motion more than the visual appearance and highlight it in the exemplar representation. On the other hand, at the superordinate level, the manner-of-motion was consistently connected with the syllabic pattern for all the exemplars. The consistent association between the two dimensions can yield an activation of the manners-of-motion with each exposure to the syllabic patterns, enhancing their learning.

In the Word learning test, we were able to examine the learning of the specific moving characters encoded by the entire words, controlling categorization at the superordinate level. In this test, we examined the learning of both the manners of motion of the characters and their visual forms. We used the question ‘Who is ____?’ that directs the participant’s attention to the identity of the specific character in contrast to other characters. In the Appearance trials, the distractors shared the same manner-of-motion, which induce categorization based only on the visual form of the moving characters. In the Motion trials, the distractors shared the same appearance, directing the participant’s attention to the specific character to retrieve the missing information. The results revealed that indeed there is no benefit in learning the manners-of-motion of the characters when they were encoded morphologically. Contrarily, the shared syllabic patterns impeded the learning of the manner-of-motion at these levels of categories. These results indicate that even when hierarchically linked categories share the same network, and several exemplars are activated at the different levels, the function of the word-form and the way they activate other dimensions vary across the levels.

4.3. On the effect of the sub-lexical form on word representation

The main finding of the current study is that the syllabic pattern that encodes a superordinate category impedes the learning of lower-level categories encoded by the entire word, namely, dimensions that are related to the two hierarchical levels such as manners-of-motion (an effect that cannot be attributed to the phonological similarities between words that shared a syllabic pattern, as shown in Experiment 2). This effect may be related to the multiple form-meaning associations learned in the multimorphemic word condition, when the manner-of-motion is encoded by the syllabic pattern, compared with the single association learned in the monomorphemic word condition. The relationship between the superordinate category manner-of-motion and the syllabic pattern in the multimorphemic words may have influenced how this concept was represented at lower levels.

The influence of higher-level representations on lower levels was found previously in studies on learning conceptual categories (Clapper & Bower, 1991; 1994). These studies showed that learning a general schema influences the value of some dimensions of the individual instance that belong to the same category. They found that dimensions related to the general schema are less involved in identifying these instances compared with idiosyncratic dimensions that are specific to each of them. Thus, Clapper and Bower conclude that such category-related dimensions are less valuable in the representation of individual instances. In our view, this top-down effect can result from the structure of hierarchical categories, where one category is nested in the other. In these categories, a diagnostic dimension that is highly activated at a superordinate level can be confusing and less informative at lower levels. For example, ‘moving the body to music’ is a valuable dimension of exemplars under the category ‘dancer’ as they are common across the category members and play an essential role in defining the boundaries of the category relative to contrasting categories. The exact dimension, however, might not have a similar value or degree of informativeness in lower-level categories. This dimension is shared across category members such as ‘ballet dancer’ and ‘butoh dancer’ and thus contributes neither to the categorization of exemplars at these lower levels nor to the discrimination between them.

In the present study, the participants learned dimensions diagnostic of the superordinate category, such as skipping and flipping manners-of-motion, at both the super- and lower-level categories in the two learning conditions. Although our experimental design does not allow direct comparison between the hierarchical levels, we expect that the value of the manner-of-motion dimensions would vary across the representational levels. While manners-of-motion contribute to the categorization at the superordinate level as they allow distinguishing between skipping vs. flipping characters, their value must be reduced at lower levels, where participants should distinguish between moving characters under the same category of manner-of-motion (e.g., finupál and bizudáx that are skipping characters). Our findings showed that the value of manners-of-motion in the lower-level category (specific character) was even more reduced when they were encoded by syllabic patterns. This was reflected in the lower performance of the participants in learning the concepts ’skipping’ and ‘flipping’ in the multimorphemic word condition. This can be a result of an enhancement of manner-of-motion activation at the superordinate level when it was encoded by the syllabic patterns, as suggested by the Label Feedback Hypothesis. This activation may have simultaneously weakened the value of the manner-of-motion dimension and its degree of informativeness in the process of the moving character’s identification.

The ability to predict the learning of the manner of motion, as measured in the Word learning test, based on participants’ performance in the morphological generalization task, supports our suggestion that enhancing the manner-of-motion dimension at the category level decreases its values at lower levels. This effect was observed mainly in the skipping manner of motion, where generalization of this manner predicted lower performance in learning it at the level of the moving character (see Table 3).

The fact that this correlation was found only in the skipping and not in the flipping manner-of-motion might be related to differences in saliency and novelty between the two. Skipping, while rhythmic, involves less visual complexity compared to flipping. The change in orientation and position involved in flipping creates a more complex visual event. Moreover, skipping is closer to human-like motion, as it involves rhythmic movements similar to walking and running, making it a more typical example of human locomotion compared to the specialized and less frequent motion of turning and spinning around an axis while moving forward. Due to these differences, the representation of the flipping manner-of-motion at the level of the moving character might be more stable and protected from the top-down effect we observed in other manner-of-motion. Further studies on different types of categories would help us understand whether and how the top-down effect we found is related to the type of category.

5. Limitations and future research

In the present study, we investigated the effect of higher-level categories on word learning using multimorphemic proper nouns. While it is more common in Semitic languages to form proper nouns using more than one bound morpheme (see our methods section for detailed examples), multimorphemic proper nouns are less common in other languages, such as English. However, there are still cases where suffixes form proper nouns (e.g., city names: in English -cester in Leicester and Gloucester, Dutch -drecht in Dordrecht and Papendrecht, German -ow in Teltow and Gatow, and Danish -lev in Tinglev and Herlev) (Schlücker & Ackermann, 2017). Names of medications provide another example of multimorphemic proper nouns, as suffixes indicate the belonging to the same family and sharing the same structure, mechanism and function (e.g., amoxicillin, ampicillin and dicloxacillin belong to the penicillin class, while atorvastatin, fluvastatin and lovastatin belong to the statin class). Given the variations in morphological productivity across languages, future studies on speakers of non-Semitic languages, where the morphological system is less productive and proper nouns are primarily monomorphemic, would be important. This would help us gain a more complete picture of whether similar results are found when participants are less familiar with decomposable proper nouns.

Moreover, we suggest that future studies should examine the learning of superordinate categories (e.g., skipping and flipping in our study) independently of specific word learning. This would provide a more precise picture of conceptual representations at the superordinate level. The form-meaning associations we examined in the morphological generalization test can only be assessed following the multimorphemic word condition. Thus, we need more information to determine how the superordinate category is represented in the monomorphemic condition. The conceptual generalization test used in our study can indicate that skipping and flipping are concepts that can be generalized to new items. However, it does not reveal the value of these dimensions at the superordinate category level. In this test, the characters from the exposure phase are stationary, requiring participants to retrieve the manner of motion for each character (a process that has been examined independently in Motion trials in the Word learning test) before generalizing it to a new character. This process involves activating two representational levels, where category learning cannot be teased apart from learning the specific characters. A categorization test that does not involve characters from the exposure phase would provide more insight and allow comparing superordinate category learning in the two conditions.

6. Conclusions

The current study showed that the effect of the word-form on categorization is not limited to the category it encodes but involves lower categories in the same hierarchical structure. However, the effect of the same word-form can be reflected differently at the different hierarchical levels. Based on our findings, we suggest that differences in the internal structures of two levels of hierarchically linked categories influence the function of the word-form across the levels. Although the two categories share the same network and some of the exemplars are activated at the two levels, the function of the word-form and the way they activate other dimensions vary across the levels. Our suggestion aligns with Ambridge’s claim that exemplars are re-represented at each level of abstraction (Ambridge, 2020). Here we add to this suggestion a ‘how’ aspect, specific to the initial stages of learning hierarchical categories in Semitic-like multimorphemic words. We further suggest that dimensions that are highlighted in the superordinate category encoded by a sub-lexical form are less activated at lower levels at the same hierarchical structure encoded by the entire word. That is, the more a dimension is valuable and activated in the superordinate level, the less informative it is in lower levels.

In conclusion, future studies on the effect of sub-lexical forms on categorization using the same stimuli and research paradigm would further support our discussion of exemplar representation across the categorical levels. Moreover, we would like to highlight the importance of investigating words of morphologically distinct structure that would challenge the underlying assumptions of common theoretical approaches and expand our understanding of how hierarchically linked categories are learned and represented.