Learning Concrete and Abstract Novel Words in Emotional Contexts: Evidence from Incidental Vocabulary Learning

ABSTRACT This study investigates the role of emotional linguistic input in learning novel words with abstract and concrete denotations. It is widely accepted that concrete words are processed more easily than abstract ones. Several theories of vocabulary acquisition additionally propose a critical role of sensorimotor and emotional information during novel word learning. In this study, proficient adult speakers of English read novel words denoting concrete and abstract words (e.g. boat vs religion) embedded in informative passages with different emotional valence (positive, neutral, and negative). After five exposures to each novel word in an emotionally consistent context, participants were tested on orthographic and semantic vocabulary learning, and provided valence judgments of these novel words. A concreteness advantage was seen in both tasks measuring semantic learning. Critically, valence of linguistic contexts was more influential for novel words with concrete denotations. In line with previous reports, the transfer of context emotionality to novel words (i.e. semantic prosody) took place in concrete stimuli but it was not found in abstract stimuli, even though both were embedded in emotional contexts. An equal advantage was seen for semantic learning of novel words with both concrete and abstract denotations seen in positive contexts. These findings provide support for weak embodied theories of cognition, which propose experiential and linguistic information as critical for concrete and abstract novel word learning.

elicitation of a sensorimotor experience related to the referent of a given word, arguing that processing semantic information activates the same neural systems that are activated during perception and action (Barsalou, 1999, Gibbs, 2006;Decety & Grèzes, 2006).For instance, similar brain activation patterns were observed for an individual's actual body movements and merely reading words that describe body movements (Hauk, Johnsrude, & Pulvermüller, 2004;Pulvermuller et al., 2005).
Words that are abstract (i.e., low in concreteness) present a challenge to the above accounts: How can a word with no tangible, real-world referent be represented and processed?What information is activated when learning an abstract word?The strong embodiment account (e.g., Gallese & Lakoff, 2005, Glenberg & Kaschak, 2003;Lakoff & Johnson, 1999;Zwaan, 2004) suggest that abstract concepts are represented through "conceptual metaphor:" Concrete concepts provide grounding for abstract concepts through metaphorical extension.For example, the abstract concept of "understanding" can be learned through an expression such as "to grasp an idea."(see Boulenger et al. (2009) and Aziz-Zadeh et al. (2006) for conflicting findings).An alternative proposal arouse under a family of theories falling under weak embodiment.While varying in detail, several such accounts suggest that, in the absence of tangible representation in the world, the learning of abstract words is particularly attuned to additional dimensions of semantics, including linguistic and sensorimotor aspects of the novel word meaning (e.g., Barsalou, 2009;Vigliocco et al., 2009;Wilson-Mendenhall et al., 2011).Thus, Vigliocco et al. (2009) suggest that two sources of information are utilized, regardless of concreteness of the word: 1) experiential, which includes sensorimotor and emotional information, and 2) linguistic, which is the associations learned through co-occurrences of words in texts.The Affective Embodiment Account (Kousta et al., 2009), for instance, provides evidence in a study showing that abstract words are learned using emotional information, and furthermore, abstract words with increased emotional associations are learned earlier.Similarly, Barsalou (2009) proposes the Perceptual Symbol Systems framework, which highlights the role simulation of situated knowledge, including emotional experience, as being crucial to the processing and representation of abstract words.Wilson-Mendenhall et al. (2011) describe abstract and emotional words as being acquired and processed in the context of certain activities, including sensations in the body.In a similar vein, Kousta et al. (2011) propose an Embodied Theoretical View of Abstract Representation, proposing that abstract and concrete words bind different types of information, with concrete ones putting greater weight on sensorimotor information in the word-learning environment while abstract words put a greater weight on emotional information.This perspective is further developed by Ponari et al. (2018) who advance an Abstract-via-Emotion (AvE) hypothesis.It views emotion as a bootstrapping mechanism for learning abstract words, such that learners make use of information related to emotion when sensorimotor information is not available.This emotional information can be re-used through bootstrapping when learning subsequent abstract words based on those already acquired (see Moffat et al., 2015).In sum, there is a family of accounts that specifically ties the success in learning relatively abstract words to the emotionality and other semantic characteristics of those words.
This theoretical stance finds support in mounting empirical evidence that abstract word learning, recognition and categorization in children is particularly affected by emotional experience (e.g., Newcombe et al., 2012;Siakaluk et al., 2016Siakaluk et al., , 2014; see also Gleitman et al., 2005).For instance, a regression study by Kousta et al. (2011) found that abstract words that tend to have emotional connotations are acquired earlier.Lund et al. (2019) also found that children aged 6 had quicker responses to positive abstract words in a lexical decision task, and children aged 7 had quicker responses to positive words regardless of concreteness.Ponari et al. (2017) found that valence and age affected lexical decision times: children aged 8-9 performed better with valenced abstract words compared to children aged 10-11, suggesting that emotional valence facilitates learning of abstract words in children.Specifically, Ponari et al. (2020) provided emotional vs. neutral vocabulary learning contexts to children aged 7-10 years and tested orthographic and semantic learning of novel words.In the semantic task, only children aged 7-9 years performed better in defining valenced abstract words than neutral ones.However, since there was no effect of teaching strategy, Ponari et al. suggested that the emotional information in the learning context does not facilitate learning.Likewise, Vigliocco et al. (2018) found that after the age of 9, children start to make use of linguistic information in addition to previously used emotional information, as seen by less sensitivity to emotional valence.These findings suggest the differential engagement of sensorimotor and emotional information begins early in the time-course of language development.
Almost invariably, the theoretical discussion of abstract and concrete word learning has so far focused on the semantic properties of the words to be learned.Yet there is evidence for an independent role of such connotations in the surrounding linguistic context.The Lexical Quality Hypothesis (Perfetti, 2007;Perfetti & Hart, 2002) proposes that the emotional and sensorimotor information found in linguistic contexts influences semantic representation.Using large corpora of texts, Snefjella and Kuperman (2016) investigated the role of concreteness and valence of the linguistic context, finding that words that tend to occur in positive contexts come with reduced cognitive effort in lexical tasks -shorter recognition times, better memory recall and earlier age of acquisition -compared to words occurring in negative contexts.To our knowledge, only one study so far has used the novel word learning paradigm to investigate the role of context affect.Snefjella et al. (2020) embedded novel words in contexts that were kept consistently positive, neutral or negative for a given novel word and participant, finding that novel words learned through contexts with positive affect show advantages in posttests measuring semantic and orthographic learning.Thus, an item plurk showed a reliably higher valence if encountered by a participant in positive contexts, lower if in neutral contexts, and even lower in negative contexts.Moreover, Snefjella et al. (2020) observed higher/medium/lower scores in posttests of semantic knowledge for novel words learned in positive/neutral/negative contexts.This positivity advantage for novel words with concrete denotations is in line with the processing advantage established for existing positive words over negative ones (see review in Kuperman et al., 2014) and the advantage reported for existing words occurring in more positive contexts (Snefjella & Kuperman, 2016).Importantly for the present purposes, all novel words in Snefjella et al. (2020) study had concrete denotations (e.g., a tool, a piece of clothing) and thus could not shed light on whether emotional information of linguistic context is recruited particularly strongly in the learning of abstract rather than concrete words.This word concreteness x context affect interaction is implemented in the present study.

The present study
The main interest of this study is to investigate the link between the dimension of concreteness with emotional dimensions to explain learning mechanisms for concrete and abstract word learning in adult speakers of a language, and more specifically, during fast-paced incidental learning of novel words in a short experiment: see motivation below.Most of the empirical basis for related theoretical accounts has come from observational studies that document vocabulary acquisition and language development in children as well as empirical studies employing lexical decision tasks to study concreteness processing in children (see the Introduction).We test the effects of emotional and sensorimotor information in a novel word learning paradigm, where factorial manipulations allow for a high degree of experimental control over both the input, the duration of learning, and the measurement of learning outcomes.Our target population is proficient adult native speakers of English and the learning opportunities are confined to a relatively short experimental session.Observing a greater impact of emotional information on abstract rather than concrete word learning in this setup, as predicted by a family of theoretical accounts reviewed above, would contribute to the current understanding of vocabulary acquisition.Specifically, it would confirm that the proposed learning mechanisms are in place long after childhood and that they can be active instantaneously (i.e., after five exposures to a novel word in a short time) rather than over years of language use.
We build on previous word learning studies to factorially manipulate concreteness of word meaning and emotionality of that word's linguistic context and to assess theoretically predicted main effects of emotion and concreteness and their interaction.To this end, we modify Snefjella et al. (2020) stimuli and place words with abstract (e.g., temptation) and concrete (e.g., device) intended meanings into informative contexts that are either positive, neutral, or negative.By keeping context emotionality consistent for each novel word within a given participant, we associate a specific emotional connotation with each learning occasion.
The predictions of the study are as follows.In line with the reported robust concreteness advantage, we expect novel words with concrete intended meanings to be learned better than those with abstract ones.We also expect context emotionality to be absorbed to some degree by novel words (that are originally devoid of either denotations or connotations) in a process typically labeled semantic prosody (see details in Snefjella & Kuperman, 2016;Winter, 2019).Following the Embodied Theoretical View of Abstract Representation (Kousta et al., 2011), we predict emotionally valenced linguistic contexts to influence learning of novel words with abstract denotations more so than learning of novel words with concrete denotations.Based on results of Snefjella et al. (2020), we refine this prediction.Namely, we expect novel words occurring in positive contexts to show better semantic learning and come with higher scores in posttests of novel word form or meaning as compared to words learned in neutral or negative contexts.Furthermore, this positivity advantage is anticipated to be particularly salient in novel words with abstract connotations, because of a hypothesized greater reliance of abstract words on affective information (see above).In sum, embodied theories predict the main effect of concreteness (higher-quality learning for words with concrete denotations), the main effect of affect (higherquality learning for words occurring in more positive contexts) and an interaction between the two (more positivity advantage in abstract rather than concrete words).We test these predictions in a large-sample online study set up in a novel word learning paradigm.

Participants
Eighty-eight participants were recruited from Amazon Mechanical Turk, an online crowdsourcing marketplace (mturk.com),and 198 through McMaster's undergraduate participant pool.This study was approved by McMaster's Research Ethics Board (2011-165).All participants were compensated with $4 USD or with course credit.Participants were removed if they indicated that their first language is not English, if they had a vision problem, hearing impairment, language disability, or learning disability, or if they had completed the study twice, resulting in 154 remaining participants (Mage = 28.17,53 male, 100 female, 1 other).The two participant pools did not differ in performance in any test, and so they are considered jointly below.

Materials
The setup of this study is a modification of Snefjella et al. (2020).It has a 3 × 2 design, with three levels of emotion of the linguistic context (positive, neutral, and negative) and two levels of concreteness of the novel word (abstract and concrete).Nine novel words were generated using the Wuggy software (Keuleers & Brysbaert, 2010) along with two homophones for each (e.g.plurk, plirk, plerk), and rated in a separate norming study, which demonstrated a valence range from 2.44 (rotch) to 4.22 (ceammy) to ensure that the novel words do not differ significantly.The full list of novel words is available in Supplementary Materials S3.These novel words were used to replace abstract and concrete nouns in the target passages.
The passages were created in triplets, such that each of the three passages contained two highly positive, highly negative or neutral words, while most of the carrier passage remained the same, see Table 1 for example stimuli.Moreover, all passages in a triplet had an identical final sentence containing one and the only occurrence of the novel word in a passage.Five of the triplets describe abstract words (e.g., temptation, psyche, religion, economics, chance) and four describe concrete words (e.g., boat, plant, kitchen appliance, clothing).For each novel word, five sets of triplets were created.Snefjella et al.'s stimuli for concrete words were used here, while abstract ones were developed anew.These triplets were rated in a separate norming study on Amazon Mechanical Turk, asking participants to rate the passages on a scale of 1-9 for valence, see Table 2.This norming study enabled us to ensure that the passages are reliably different in emotionality and that there are no significant differences between concrete and abstract conditions (t(2126) = −4.42,p < .001)see Table 11 in Supplementary Materials S1.Flesch Reading Ease and Flesch-Kincaid Grade-Level estimates were collected from the computational tool Coh-Metrix to ensure that contexts (abstract vs. concrete) are reasonably matched on readability (t(2074.8)= −3.69,p = <.001)see Tables 12  and 13 in Supplementary Materials S2 (Graesser et al., 2004).The full list of passages can be found in Supplementary Materials S3.
For each participant, each novel word appeared in five passages.In total, each participant read 45 passages.The presentation of the passages was counterbalanced so that each participant saw each novel word in only one emotional type of context and always with the same intended (concrete or abstract) denotation.Thus, for one participant, a novel word plurk would also occur five times in negative contexts and with a denotation "boat," while for another plurk would occur in neutral contexts and with a denotation "psyche."

Procedure
The experiment was administered as an online study.The experimental paradigm was that of incidental novel word learning, that is, participants were unaware that vocabulary posttests were to be administered after the learning phase.Prior to beginning the experiment, participants provided informed consent to participate in the study.When beginning the experiment, participants received the instructions: "This is a study of different reading styles and the ability to understand texts.You will be shown a few short texts.Please read them carefully for comprehension."In this part of the experiment, participants read 45 short passages, constituting the "learning" phase.In the second part, they were tested on their vocabulary learning through four tasks: orthographic choice, valence rating, definition prompting, and definition matching.Instructions appeared on the screen prior to each task.Details about each task are given below.

Post-tests
Below, the posttests are presented in the order that they appear to the participant.In the orthographic choice task, word form knowledge was tested, while in the definition prompting and definition matching tasks semantic recall and recognition were tested, respectively (Laufer et al., 2004).Additionally, valence ratings to novel (initially meaningless) words enabled testing semantic prosody, i.e., the transfer of emotional information from linguistic context to the novel word.
Orthographic choice.The orthographic choice task measures learning of the forms of the novel words.Participants saw the novel words (e.g., plurk) along with homophones (e.g., plirk, plerk) and were instructed to indicate on the screen which novel words they recall reading.The baseline performance in this task is 50%.See Supplementary Materials S2 for instructions.
Valence rating.The valence rating task measures the learned valence of the novel words.Participants were instructed to use an on-screen scale ranging from 1 to 9 to indicate the emotional connotation of the novel word or by pressing "0" to indicate that they had not seen this novel word.This task included filler novel words that were not included in the passage readings.See Supplementary Materials S2 for instructions.
Definition prompting.The definition prompting task measures semantic recall of the novel words.Participants were presented, one at a time, with novel words and fillers.For each word, they were asked to give a definition for words which they recall reading, and to click the appropriate box indicating this word was not seen for words they did not recall reading.For seen novel words, hints were provided after each response from the participant (after participant either submits a typed a definition, or indicated that they did not see the word), totaling two hints.The hints consisted of fragments of passages that this novel word was seen in.Responses were scored on a scale of 0-3 by two research assistants.Three points were given for correct responses that were elicited before any hints were received.A response was considered "correct" when similar in semantic meaning to the target definition (e.g., for target "kitchen appliance," responses of "appliance," "stove," "microwave" would all be considered correct).Correct response given after one hint was given two points, and after two hints, one point.If the correct response was not provided after 2 hints, 0 points were earned.Only the first correct response was used to calculate the score for each question.See Supplementary Materials S2 for instructions.The baseline performance in this task, if responses are given randomly, is 0 points.
Definition matching.The definition matching task measures semantic recognition of the novel words.It is administered last so as not to confound the other tasks.The participants were provided with a list of nine novel words and nine foils and asked to match the words they read to the list containing nine definitions and nine foil definitions.See Supplementary Materials S2 for instructions.The baseline at-random performance in this task is 5.6%.

Individual difference measures
Participants filled out an online survey measuring individual differences.Participants were asked their age, gender, country of birth, country of residence, and highest level of education.They were also asked detailed questions about their language background: which language is most dominant (able to list up to five languages), and the order in which they learned their languages.If they were not born in an English-speaking country, they were asked about the age that they moved to one and the age that they began to learn English.They were also asked to rate, on a scale of 1-10, their proficiency in each of the following: speaking, understanding spoken language, and reading.Finally, participants reporting any vision problems, hearing impairments, language disabilities or learning disabilities were removed from analysis.The summary of individual difference measure responses can be found in Supplementary Materials S2.

Variables
Dependent variables.Vocabulary learning was measured through three posttests, described in detail above: orthographic choice, definition prompting, and definition matching.The outcome of orthographic choice was response accuracy in percent correct, that of definition prompting was the average score in points (for scoring scheme see above), and that of definition matching was the percent of accurate matches between seen novel words and intended denotations.An additional test, the valence task, measured the potential learning of emotional connotation from context on a 1-9 scale.
Independent variables.One critical variable was whether the word's intended denotation was concrete or abstract.Another was emotional valence of linguistic contexts of novel words.A possible operationalization of context valence would be a tri-partite division of contexts into positive, neutral, and negative.However, a different operationalization is both theoretically advantageous and has greater explanatory power (Snefjella et al., 2020).Namely, we use the rating of valence for a specific novel word given by a specific participant as a predictor of their performance in vocabulary posttests.This way, we tap into an individual experience of how valenced a given word is and associate it with the quality of learning those novel words.In all analyses below, this individual measure of valence explains more variance than the categorical three-level characterization of passages in terms of emotionality.
Control variables.Reading time for each passage is both a dependent and an independent variable in this study.We test the influence of context valence and concreteness on how long participants read those passages.We also include passage reading times as a co-variate in regression models predicting vocabulary posttests.This step enables us to account for Schmidt's Noticing Hypothesis (1990;Schmidt, 2001), which proposed that longer reading times may indicate more attention paid to the stimuli and should therefore lead to better scores in posttests.Reading times were measured for each participant and passage by collecting the time intervals between the appearance of the text on screen and the moment the participant clicked "Next" to proceed to the next passage.

Statistical considerations
Generalized linear mixed effect regression models (lme4 package, Bates et al., 2015) were used in this analysis.These models allow for accounting for between-items and between-participants variability when estimating effects of covariates.All models were initially fitted with the full random-effects structure and downgraded if convergence errors were received.The lmerTest package (Kuznetsova et al., 2017) in the statistical platform R (R Core Team, 2018) was used to estimate p-values for fixed-effects with Satterthwaite's approximation for degrees of freedom.Library effects (Fox, 2003;Fox & Weisberg, 2019) was used to visualize critical effects and interactions.

Results and discussion
This section presents results for each posttest and then discusses them jointly.

Valence rating
A total of 6350 responses were recorded to the seen novel words.Participants accurately discriminated between seen and unseen novel words (89% accuracy) while giving valence ratings, in line with Snefjella et al. (2020) accuracy rates of 88% and 90% (experiments 1 L, 1O).Table 3 summarizes mean valence ratings given to learned novel words, broken down by their emotionality and concreteness, see also Figure 1.As Table 3 demonstrates, a numerical effect of semantic prosody emerges in both abstract and concrete conditions: novel words appearing in positive, neutral or negative contexts acquired values of valence polarized in respective directions.The range and dispersion of valence ratings to four concrete denotations was similar to that observed in Snefjella et al. (2020) two experiments (1 L, 1O) that used identical contexts (ranges: 4.95-5.87,5.04-5.77).Thus, the present experiment replicates prior results for the overlapping stimuli.
An important new finding is that for novel words with concrete denotations valence ratings had much more of a range between negative and positive contexts (5.09-5.83)compared to words with abstract denotations (5.3-5.5).This concreteness x context valence interaction was significant (F (6331, 2) = 6.06, p = .002),see Table 7.Moreover, post-hoc analyses (see Tables 8 and 9) revealed that the increase in novel word valence in abstract denotations determined by context valence was not significant (p > .1 in all contrasts).

Orthographic choice
The quality of learning the orthographic forms of novel words was measured in the orthographic choice posttest.Participants showed high accuracy (around 90%) in discriminating between seen novel words and their homophones (see Table 4), which is comparable to Snefjella et al. (2020) accuracy scores (81%, 81%, 86%, and 94% for experiments 1 L, 1O, 2O, and 3O).A regression model fitted to orthographic choice responses (Table 10) did not indicate an effect of either individual  valence ratings or concreteness, nor was there a reliable interaction of the two (all p > .05).That is, affective and sensorimotor characteristics of the learned novel words and their context did not affect orthographic form learning.This agrees with the findings of Snefjella et al. (2020) who did not observe effects of context emotionality on orthographic choice scores in words with concrete denotations.Notably, passage reading time showed a strong positive effect on the test scores, with participants who read materials in the learning phase longer having higher scores in this task (β = 0.934, SE = 0.144, z = 6.496, p < .001), in line with the Noticing Hypothesis (Schmidt, 2001).

Definition prompting
Semantic recall of the novel words was tested in the definition prompting posttest.Responses were scored on a scale of 0-3, with 3 being the highest possible score.In all conditions, the scores were much higher than the baseline of the test (0 points) suggesting that semantic learning did take place, see Table 5 with mean scores for each experimental cell.The regression model fitted to definition prompting scores (Table 11) revealed a concreteness advantage.Words with concrete denotations were learned better overall than words with abstract ones in terms of semantic recall (β = 0.234, SE = 0.234, z = 0.104, p = .025).Importantly, the model did not indicate any effect of valence ratings on test scores, nor was there an interaction between denotation concreteness and context valence (all p > .05).Emotionality of context did not affect this facet of semantic learning for either abstract or concrete words.Passage reading times showed an effect expected under the Noticing Hypothesis (Schmidt, 1990;Schmidt, 2001): a longer inspection of passages in the learning phase led to higher scores on this semantic knowledge test (β = 0.119, SE = 0.119, z = 0.048, p = .013).

Definition matching
Semantic recognition of the novel words was tested in the definition matching posttest.Overall, participants had 43% accuracy in recognizing seen novel words and matching them to correct denotations.This performance exceeded by far the baseline score of 5.6% that random responses to the test would have led to.A regression model confirmed the anticipated concreteness advantage (Table 12; β = 1.785,SE = 0.471, z = 0.471, p < .001).Novel words were matched to concrete denotations more accurately than to abstract ones (Table 6).Another important finding was that novel words learned in more positive contexts and associated with higher valence ratings showed higher definition matching scores (β = 0.139, SE = 0.055, z = 2.502, p = .012).This positivity advantage was demonstrated in Snefjella et al. (2020) for concrete denotations, while here we observed it in both concrete and abstract conditions.The definition matching score for a novel word with the lowest valence rating was estimated to be about 0.2 points lower than for a novel word with the highest valence rating.Crucially, this valence effect did not interact reliably with whether the novel word's intended denotation was concrete or abstract, see Figure 2 (F(1, 2 = 0.55, p ≥ .1).That is, (a) positive contexts led to better semantic learning and (b) the strength of affective support for learning was equal for words with denotations that are (e.g., tool) or are not (e.g., psyche) directly linked with sensorimotor experiences.Standard deviation of the by-participant intercepts is 1.38, and standard deviation of the word intercepts is 0.43.N = 152, number of observations = 6350.The reference level for condition is −1 (negative), while 0 stands for neutral and 1 for positive.The reference level for denotation is abstract.Standard deviation of the by-participant intercepts is 1.42, and standard deviation of the word intercepts is 0.41.N = 152, number of observations = 3585.The reference level for condition is −1 (negative), while 0 stands for neutral and 1 for positive.The reference level for type is MTurk.Standard deviation of the subject intercepts is 1.48, and standard deviation of the word intercepts is 0.40.N = 152, number of observations = 2870.The reference level for condition is −1 (negative), while 0 stands for neutral and 1 for positive.The reference level for type is MTurk.

Reading times
Reading times were collected from each passage and used as a dependent variable.A regression model (Table 13) showed no effect of valence (F(1172.43, 1) = 4.47, p = .03)or concreteness (F(78.35, 1) = 2.93, p = .09)on passage reading times, nor was there an interaction of the two factors (F(1135.60, 1) = 0.51, p = .48).We conclude that the amount of noticing during the learning phase is unaffected by affective or sensorimotor dimensions of the novel words and their contexts.

General discussion
The goal of this study was to investigate an argument made by the family of weak embodiment theoretical proposals that emotional and linguistic information in the connotations of the word to be learned and its context can influence the learning of the word, and especially so if that word is abstract rather than concrete (Vigliocco et al., 2009, Kousta et al., 2009;Kousta et al., 2011).In the absence of the tangible references, abstract word meanings arguably rely more on other available information sources that help encode the learned linguistic material in the mental lexicon: Emotional information is identified as one such important source (Kousta et al., 2011).This mechanism is argued to account for acquisition of abstract words in childhood and has found most support in studies of children as well as corpus studies.This paper tested this claim in proficient adult native speakers of English in  a brief experimental session.Specifically, we manipulated the valence of contextual information surrounding novel words which were further manipulated to denote either concrete or abstract words.A large-scale experiment using an incidental vocabulary learning paradigm had five abstract and four concrete nouns that were replaced with novel words in five contexts for each word.Each participant saw each novel word in five consistently positive, neutral, or negative contexts, which can be used to infer a consistently concrete or abstract denotation for that word.Following the reading phase, participants were tested on orthographic and semantic learning of the novel words.
The motivation for selecting an adult population rather than children and examining fast-paced learning with a few learning opportunities rather than a result of natural language learning was to study whether proposed learning mechanisms are in place throughout the lifespan and whether they are accessible and active "on a moment's notice," i.e., over a few exposures to a novel word.Critically, several theories of embodied word learning predict that emotionality of linguistic context is more influential for learning novel words with abstract rather than concrete denotations (see the Introduction).The present predictions consisted of an expected learning advantage to novel words with concrete denotations (supported by much prior research, see the Introduction) and to novel words occurring in more positive contexts (see Snefjella & Kuperman, 2016;Snefjella et al., 2020).We discuss these predictions one by one.The well-established concreteness advantage in the quality of learning was evident in both semantic knowledge tasks.While both concrete and abstract denotations (e.g., boat vs psyche) resulted in scores well above chance for both recall (definition prompting) and recognition (definition matching), scores were consistently higher for words with concrete denotations.
The tendency of novel words learned in more positive contexts to come with a better quality of semantic learning (Snefjella et al., 2020) was replicated here as well, in the definition matching task.The effect was equally strong for abstract and concrete words.No effect of context valence was observed in the definition prompting task.We also note that neither the concreteness nor the positivity advantage surfaced in the orthographic learning quality or in the inspection time of passages during the learning phase.Apparently, emotional and sensorimotor dimensions of the learning events have a focused influence on semantic rather than form learning, and more in word recognition than recall.
The present study answers in the negative its central question, whether affective information conveyed by linguistic context influences learning concrete and abstract words differently.While either concreteness of denotation or context valence or both showed an impact on semantic knowledge, neither semantic task revealed an interaction between these two factors.Context valence was equally influential for concrete and abstract denotations in definition matching (Figure 2) and equally ineffective for the two types of denotations in definition prompting.
The piece of evidence that is perhaps most difficult to reconcile with the hypothesized role of emotion in abstract word learning is the valence rating task.Novel words with concrete denotations were the ones that absorbed affect of linguistic contexts the most.After the learning phase, they were judged as more positive, neutral or negative in accordance with the contexts in which they occurred.Thus, semantic prosody, or the transfer of connotations from context to a word, demonstrably took place in these stimuli (in line with Snefjella & Kuperman, 2016;Snefjella et al., 2020;Winter, 2019).Conversely, novel words with abstract denotations were judged to have the same neutral valence, even though -like their concrete counterparts -they were embedded in contexts with different affective polarity.
This goes against several theoretical accounts (see the Introduction) which argue that emotional information is particularly required and recruited when sensorimotor information is not available (i.e., in words denotating abstract concepts).We outline some of the possible reasons for this discrepancy.An important deliberate deviation of our study from the empirical evidence that supports those accounts is its population (adults vs children) and the time scale of learning (a brief experimental session, compared to years of accumulated linguistic knowledge) give less opportunity to use context in an effective way and thus may affect learning.In a sense, our experimental setup presents conditions under which the critical interaction (context valence x word concreteness) is the least likely to emerge.This is because with age, the use of affective information in learning is either supplemented or replaced by the use of linguistic information: see Ponari et al. (2018) contrast between 9-and 10-year-old children.Additionally, the tightly controlled experimental stimuli with novel words come in quick succession, possibly impeding memory consolidation processes.Moreover, presented for reading on the screen, outside of a natural communicative environment, these stimuli are more difficult to link to either sensorimotor or affective information for the learner.
The present data suggest that at some point in the developmental trajectory -contingent perhaps on age, language experience, emotion regulation, and world knowledge -vocabulary acquisition of abstract words transitions from a heavy reliance on available emotional information to the disuse of this information, and instead equal reliance on linguistic information for learning of both abstract and concrete novel words.Charting when and explaining how this transition happens is a theoretical and empirical goal for future research.
An additional observation that theories of word learning need to account for is that acquisition of concrete words does rely on emotional information in adult proficient learners for recognition, but not recall, of novel words, with an effect noticeable over a few encounters with the novel word.Whether this reliance is absent in early vocabulary acquisition and rather develops with age or it is present at all stages of vocabulary development is also a topic for further examination.

Conclusion
The results of the current study support a partial dependence of novel word learning on sensorimotor systems, in line with weak embodiment accounts (Kousta et al., 2011;Vigliocco et al., 2009) which propose dependence on both experiential and linguistic information underlying abstract word meaning.The current study is in line with these accounts, as novel words denoting both concrete and abstract concepts were successfully learned using the surrounding linguistic context, regardless of context emotionality and concreteness of denotations.
This study reveals that novel word learning is a complex process demonstrably influenced by semantic dimensions of both the novel word itself and those of its linguistic contexts.We show that experiential information and linguistic information contribute to encoding of novel words and their surrounding linguistic contexts, even in proficient adult speakers of a language and in as little as five exposures to the word.

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

Figure 1 .
Figure 1.Partial effects of context valence on valence ratings to novel words, presented for abstract (left) and concrete (right) contexts.−1 stands for negative, 0 for neutral, and 1 for positive.Error bars represent standard errors.

Figure 2 .
Figure 2. Abstract and concrete definition matching accuracy by individual valence ratings.

Table 1 .
Example of stimuli for both concrete and abstract concepts.

Table 2 .
Norming study ratings of valence, means and standard deviations in parentheses.

Table 3 .
Means of valence responses per condition.Standard deviations are reported in parentheses.

Table 4 .
Orthographic choice scores by levels of context valence and novel word concreteness.

Table 5 .
Concrete and abstract definition prompting scores per condition.

Table 6 .
Concrete and abstract definition matching scores per condition.

Table 7 .
Model of valence ratings for words with concrete and abstract denotations.

Table 8 .
Model of valence ratings for words with abstract denotations.

Table 9 .
Model of valence ratings for words with concrete denotations.

Table 10 .
Model of orthographic choice scores by individual valence ratings.Standard deviation of the by-participant intercepts is 1.11.N = 152, number of.observations = 1283.The reference level for denotation is abstract.The reference level for denotation is abstract.

Table 11 .
Model of definition prompting scores by individual valence ratings.

Table 12 .
Model of definition matching scores by individual valence ratings.

Table 13 .
Model of reading times.