Introduction

Compounding is a word formation process by which existing words are combined to express a new meaning. Psycholinguistic research on compound word processing indicates that people arrive at the meaning of a compound by using a conceputal relation to mentally link together its constituent parts(see review in Gagné & Spalding, 2013). For instance, wristband is a ‘band LOCATED ON the wrist’ and bandstand is a ‘stand FOR bands’. In addition, recent word processing studies provide evidence that people attempt to construct a compound’s meaning by generating multiple relational meanings at once (e.g., Gagné & Spalding, 2013; Schmidtke, Kuperman, Gagné, & Spalding, 2016; Schmidtke, Gagné, Kuperman, Spalding, & Tucker, 2018a). Specifically, these studies demonstrate that multiple conceptual relations compete with one another during processing and that greater competition between the possible relational meanings inhibits word recognition speed. The present study addresses two outstanding gaps in the literature on the competitive nature of relational processing during compound word recognition.

The first issue is that it is unclear whether the mechanism of competition that is found during the processing of existing compound words, e.g., tinfoil, also generalizes to the processing of novel compound words, e.g., beefolk. Although the results of previous research imply that relational processing is competitive for novel compounds, in the interest of parsimony, we wish to establish whether relational competition of existing and novel compounds can be captured with a single measure of competition: namely, ‘entropy of conceptual relations’.

The second issue is that most prior studies of conceptual combination have adopted lexical decision techniques, which measure word processing effort of compounds that are presented in isolation. It therefore remains unclear whether the competitive aspect of conceptual combination plays a role in the processing of contextually embedded compounds. Since compounds are typically encountered within sentences, it is crucial to establish whether relational competition is shaped by contextual factors. We address both outstanding issues in a series of self-paced reading experiments. In the following sections, we briefly review the theoretical and empirical work on relational processing during compound word recognition, and then outline the research gaps and methodological approach of the study.

Relational structure

Early work on the linguistic expression of combined concepts contended that relational information is integral to the semantic representations of complex nominal phrases (Downing, 1977; Gleitman & Gleitman, 1970; Kay & Zimmer, 1976; Lees, 1962; Levi, 1978). Relational structure provides information about how the concepts denoted by each constituent of a compound combine with one another to construct the meaning of the whole word. The compound meatball might be understood as a ‘ball made of meat’, where ‘made of’ is a conceptual relation that posits a connection between the head ball and the modifier meat. Critically, conceptual relations are flexible in that the same conceptual relation does not always apply to the same modifier or head. For instance, the ‘made of’ conceptual relation does not apply to the compound mothball, which might be better paraphrased as ‘a ball for moths’. Understanding how humans might rely on relational structures during word recognition is important to establish because relational processing is not always considered in mainstream models of compound word recognition. More generally, the processing of conceptual relations also provides a window into the nature of conceptual combination, the cognitive process by which existing concepts are synthesized to generate new meanings (e.g., Estes, 2003, Lucas, Hubbard, & Federmeier, 2017, Parrish & Pylkkänen, 2021; Wisniewski, 1996)

Importantly, relational structure is separate from morphological structure. Relevant to the present study, endocentric compounds are composed of a head noun which denotes the general class to which the compound word belongs, and a modifier, which serves to attribute a more restricted meaning to the head (Lieber, 2005; Bauer, 2009). For example, the English compound sandpaper consists of the modifier sand and head paper, whereby the meaning of ‘sandpaper’ is a hyponym of ‘paper’. In English, almost all endocentric compounds are right-headed, meaning that the head of the compound appears as the right constituent position of the compound. Thus, while morphological information such as headedness is encoded in the surface form of a compound word in English, relational structures are ‘invisible’ in that they are not attested in a word’s orthographic (or acoustic) form.

Relational competition

A large body of experimental research on compound words presented in spaced (e.g., school yard) and unspaced (e.g., schoolyard) spelling formats demonstrates that conceptual relations are registered by readers during compound word processing (e.g., Gagné & Shoben, 1997; Gagné & Spalding, 2004a; 2009; Gagné, Spalding, Figueredo, & Mullaly, 2009; Spalding & Gagné, 2011). More recent word recognition research indicates that relational processing is competitive (Spalding et al., 2010; Schmidtke et al., 2016; 2018a). That is, when a compound word is encountered, multiple possible relational interpretations are generated and then compete with one another for selection as the most likely relational meaning (Gagné & Spalding, 2013). Findings from these studies show that compounds that generate many equally plausible and highly competing relational meanings take longer to process than compounds that have a more dominant and readily available relational meaning. The process of relational competition forms the basis of the relational-interpretation-competitive-evaluation hypothesis (RICE; Spalding, Gagné, Mullaly, & Ji, 2010).

Extant research on the RICE hypothesis quantified competition among conceptual relations using data from the ‘possible relations task’ (Spalding & Gagné, 2014; Schmidtke, Gagné, Kuperman, & Spalding, 2018b). In this task, participants are presented with a list of compound words, and are asked to choose the most plausible conceptual relation out of a set of relations for each compound (see Table 1 for a list of 16 canonical conceptual relations). Entropy of conceptual relations for a compound is defined as Shannon entropy (Shannon, 1948) computed over the resulting probability distribution of relational choices. A lower entropy of conceptual relations score indicates that there are fewer stronger competitor relational meanings for a compound, i.e., most participants tended to converge upon one conceptual relation for a compound in the possible relations task. A higher entropy score indicates that there are more equally plausible conceptual relations for a compound, i.e., participants tended to be less systematic in choosing a conceptual relation that fits the compound. Across multiple visual and auditory lexical decision (unprimed) mega-studies, Schmidtke et al., (2016; 2018a) found that entropy of conceptual relations had a clear and consistent inhibitory effect on the processing speed of existing compounds. In all studies, higher entropy of conceptual relations – stronger competition among relational meanings – was associated with longer lexical decision response times. This pattern of results supports the proposal of the RICE hypothesis, demonstrating that competition among conceptual relations plays an influential role in compound word processing.

Table 1 Semantic relations adapted from Gagné and Shoben (1997). h: head; m: modifier
Full size table

Research gap 1: Novel compound words and relational competition

The first contribution of the present study is to address the nature of conceptual combination in novel compound word processing. The role of relational information on the processing of novel compounds has been examined in prior research (e.g., Coolen, Van Jaarsveld, & Schreuder, 1991; 1993; Gagné, 2002; Gagné & Spalding, 2004b; Gagné, Spalding, & Gorrie, 2005; Štekauer, 2005). Pertinent to the present research, Gagné et al., (2005) investigated the role of competition among conceptual relations for novel compounds in two experiments. In an initial experiment, participants were presented with a list of spaced novel compounds, and for each compound were required to select one out of two plausible meanings as their preferred meaning. For example, for the compound bark cream, participants were required to choose ‘cream for bark’ or ‘cream made from bark’ as their preferred interpretation. In a follow-up experiment, a separate group of participants were asked to judge the plausibility of the same set of novel compounds which were embedded in a sentence context. Participants took longer to judge the plausibility of the compounds for which the two possible meanings were judged in the first experiment as equally likely relative to those for which there was a clear bias toward a preferred meaning. This study points to a competition effect, indicating that novel compounds are processed more slowly when one meaning does not clearly dominate over another.

While Gagné et al.’s (2005) study was important in that it pointed to a competition effect for novel compound word recognition, competition was examined as the relative dominance (measured as a probability) of one relational interpretation out of only two competing conceptual relations. If the RICE hypothesis claims that relational competition is a general property of compound word processing, then it ought to be possible to demonstrate that novel compound processing is also sensitive to competition between all 16 possible relations (see Table 1), as measured by entropy of conceptual relations (Schmidtke et al., 2016; 2018a).

Research gap 2: Context and conceptual combination

The second contribution of the present research concerns the role of sentence context on the processing of conceptual relations. Several studies have confronted the question of whether the context in which a compound is encountered during reading influences the way in which conceptual relations are accessed during reading. Gagné and Spalding (2004b) showed that relation-based interpretations for novel compounds could be verified more quickly if they read the compound placed in a supportive context versus when they were placed in a neutral context. Further, they found that novel compounds were read faster when they were associated with more frequently chosen conceptual relations compared to novel compounds that were associated with less frequent conceptual relations. Crucially, context type did not moderate the effect of the overall frequency of the modifier relation, suggesting that relational information is evaluated for a compound regardless of whether the context is supportive of a particular relational interpretation or not.

Gagné et al.’s (2005) study also manipulated discourse context. Participants were asked to provide acceptability judgments for compounds embedded within sentences that supported either a dominant or sub-dominant relational meaning. The result indicated that relational meanings were attuned to context even for established compounds. They found that contexts that supported the plausibility of sub-dominant meanings of compounds elicited shorter response times to acceptability judgments about their relational meanings. Based on these results, the processing of relational meanings appears to be a facet of novel and existing compound word processing irrespective of whether they are encountered in contexts that are both supportive and non-supportive of a particular relational meaning.

Eye-tracking during reading has also been used to investigate the issue of relational competition during sentence comprehension. In one experiment, Brusnighan and Folk (2012) found that, overall, novel compounds were read more slowly compared to existing compounds. However, novel compounds that were presented in meaning-supportive contexts elicited shorter re-reading times compared to neutral contexts. Interestingly, Brusnighan and Folk also reported that readers spent longer reading sentences containing novel compound words that they accurately remembered in a subsequent vocabulary test. The authors reasoned that increased time spent reading novel words led to a stronger semantic representation of these words in memory, resulting in better memory retention in the vocabulary test. This latter result may be explained by the Noticing Hypothesis (Schmidt, 1990), which claims that the learning of novel linguistic features requires the input to be consciously attended to and registered.

In a corpus of eye movement studies in which existing English compound words were read silently in sentence contexts, Schmidtke, Van Dyke, and Kuperman (2021) presented evidence that relational competition may not even be a feature of compound word processing during sentence reading. They reported one significant main effect of entropy of conceptual relations in gaze durations (the sum of all fixations on the word in the first pass of reading), the one significant effect of this measure out of seven sub-studies. Moreover, the entropy effect was not reliably present in an analysis of the entire corpus, showing that insufficient power may have been the reason for the failure to detect the entropy effect in the eye-movement record.

While there may have been insufficient power to find a significant effect of entropy of conceptual relations in Schmidtke et al.’s (2021), there are two additional reasons why the effect may not have been found. The first is that competition between possible relational candidates is not a feature of compound word processing during sentence reading. That is, contextual cues provide sufficient information to disambiguate the many relational interpretations of the upcoming compound word, nullifying the potential for competition between conceptual relations. The second reason for an absence of an effect null effect on eye movement during reading is that the ratings that were used to compute the measure of entropy of conceptual relations were obtained from compounds presented in isolation. These ratings were therefore not attuned to the sentences that the compounds were embedded in during the eye-movement study. Thus, in Schmidtke et al., (2021) the entropy of conceptual relations measure was an imprecise measure of competition for compounds presented in sentences and did not predict reading behavior.

Hypotheses

The present study addresses two hypotheses based on the research gaps identified in the literature. The first hypothesis tests whether relational competition effects in existing and novel compound processing can be captured by a single information-theoretic metric: entropy of conceptual relations. The RICE hypothesis predicts that competition effects during conceptual combination should arise irrespective of whether the compound is novel or exists in the language, i.e., is familiar to the language user. If the mechanism of competition generalizes across novel and existing compounds, then the same measure of entropy computed over all 16 possible relations (Schmidtke et al., 2016; 2018a) should predict the processing times of both novel and existing compounds.

The second hypothesis tests Gagné and Spalding’s (2004b; 2013) claim that although contextual cues might be used to help reach a relational interpretation of a compound, relational processing is competitive irrespective of whether compounds are encountered in context or in isolation. We test this hypothesis by examining the predictive role of entropy of conceptual relations on processing speed when compounds are read in contexts that provide support for a more restricted meaning versus those that do not. To address the issue of Schmidtke et al.’s (2021) failure to detect a competition effect for compounds read in sentence contexts, sentence-specific measures of entropy were obtained for compounds, and their ability to predict reading times were compared against that of an entropy measure obtained for compounds that were presented isolation. On the one hand, if sentence-specific entropy measures predict reading times of compounds in sentences, in both supportive and non-supportive contexts, then this would show an extension of the entropy of conceptual relations to sentence reading tasks, confirming that possible relations for novel and existing compounds are not “insulated from the sentential context” (Gagné et al., 2005, p. 215). On the other hand, if neither measure of entropy of conceptual relations reliably predicts the processing of compounds read in sentences, then this would lend support to the hypothesis that context cancels-out the potential for competition between conceptual relations.

The present study

To address the research goals, data was collected from two separate possible relations tasks wherein compound words were presented in isolation; one task included existing compounds as critical stimuli and the other task included novel compounds. In addition, data was collected from two self-paced reading studies in which compounds were read in sentences. One self-paced reading study included existing compounds as stimuli and the other included novel compounds. In both self-paced reading studies, the same compounds were presented in one of two discourse contexts. In the non-supportive context condition, compounds were embedded in sentences that did not provide any semantic information that might guide the possible relational interpretation of the compound word, e.g., There is no airforce anywhere around here. In the supportive context condition, compounds were embedded in sentences that were designed to provide a narrower semantic interpretation leading up to the compound word, e.g., Pilots in the airforce need a lot of experience flying. After the presentation of each sentence, participants were asked to select a relational interpretation that best fit the meaning of the compound.

We present our results as two separate analyses. In Analysis 1, we assess distributional and qualitative differences in the relational meanings generated by participants for existing and novel compounds across all three presentation types: isolation, supportive context and non-supportive context. The motivation for this analysis is to provide an initial step in understanding the ways in which the meaning generation of novel and existing compounds might be differentially affected by contextual information and its absence. In Analysis 2, we turn to word processing during self-paced reading, examining the role of competition between conceptual relations on processing of novel compounds, and whether the effects of entropy for existing and novel compounds are moderated by context type.

Method

Materials

A list of 100 existing compounds and a list of 100 novel compounds were created. The novel compounds and existing compounds always belonged to separate experimental lists and were shown to separate groups of participants. For the existing compounds, 100 compounds were randomly sampled from the possible relations data set made publicly available in Schmidtke et al., (2018a) while ensuring that the entropy values were normally distributed. The list of novel compounds was created using the LADEC database (Gagné, Spalding, & Schmidtke, 2019). Novel compounds were created by crossing all the left and right constituents from existing compounds in the database. The final list of 100 novel compounds were cross-checked for frequency of occurrence across all sections of the Corpus of Contemporary American English (COCA; Davies, 2008) and the Wikipedia corpus (Davies, 2015) to ensure that none were already lexicalized.

Sentences

Both lists of compounds were embedded in sentences for the self-paced reading task. Two sentence frames were created for each compound, one with a supportive context, and one with a non-supportive context. Supportive contexts were provided for novel compounds based on the compounds’ constituents. For example, in the sentence They always forget to lock the lifegate after they close it, lock is semantically related to the right constituent gate in lifegate. We ensured that the critical compound never occupied the first or last position in the sentence. After constructing the stimuli, we used a corpus-derived computational measure of semantic similarity to verify that there was a stronger semantic relation between the constituents of the compounds (including the whole words themselves for existing compounds) and all content words in the preceding sentence frame. The results verified that the semantic association between words in the preceding sentence frame and the constituents of the critical compounds (and also the whole compound words) was stronger in the supportive vs. non-supportive context conditions (see online Supplementary Materials S1 for details).

After creating the sentence frames, there were 200 sentences containing existing compounds and 200 sentences with novel compounds. Each of the 200 sentences were divided into two counterbalanced lists of 100 sentences (List A and List B) ensuring that novel and existing compounds were never inter-mixed. Each list contained one sentence for each compound, 50 of which provided supportive contexts and 50 of which provided non-supportive contexts. The materials are provided in Tables 9 and 10 in the Appendix. An example sentence frame for each experimental condition is provided in Table 2.

Table 2 Examples of a sentence in each of the experimental conditions in the self-paced reading tasks
Full size table

Procedure

There were two experimental tasks: a possible relations task where compounds were presented in isolation and a possible relations task combined with self-paced reading. Each task is described in turn below.

Possible relations task in isolation

Data for the possible relations task for existing compounds was obtained from Schmidtke et al., (2018a). The method for the possible relations task for novel compound words was identical to the method reported by Schmidtke et al., (2018a). In the possible relations task, participants were asked to pretend that they were learners of English who knew the meanings of the individual components of each compound and had to figure out the meaning of the whole word. For each trial, participants were shown a compound word in isolation. Afterwards, they were presented with 16 possible meanings for the novel compound based on the possible relations. For example, for the novel compound dogsong, participants would be asked to choose between ‘song CAUSES dog’, ‘song CAUSED BY dog’, ‘song HAS dog’, ‘song BY dog’, ‘song MAKES dog’, ‘song FROM dog’, ‘song MADE OF dog’, ‘song IS dog’, ‘song USED BY dog’, ‘song USES dog’, ‘song LOCATED dog’, ‘dog LOCATED song’, ‘song FOR dog’, ‘song ABOUT dog’, ‘song DURING dog’, and ‘song BY dog’ (Gagné and Shoben, 1997). Each participant selected a relational interpretation for each compound. At the beginning of the experiment, participants were given examples of relations and were given four practice trials. Both experiments (existing and novel compounds) were administered on the Amazon Mechanical Turk (MTurk) data collection platform. The experiments took approximately 25 min each to complete.

Possible relations task with self-paced reading

We combined a self-paced reading task with the possible relations task. This task was performed for four different lists by different groups of participants: the two counterbalanced lists of existing compounds embedded in sentences, and the two counterbalanced lists of novel compounds embedded in sentences. Each trial began with a fixation cross appearing for 1 s. A sentence then appeared one word at a time, with each word appearing in the center of a screen in Arial font with a letter height of 50 pixels. Participants progressed to the next word in the sentence by pressing the space key, and the latency of each button press was recorded. After reading each sentence, participants performed the same possible relations task for the embedded compound as was used in the possible relations task in isolation. Participants read a total of 4 practice sentences and 100 critical trials. Each participant completed one list (either List A or List B) of existing or novel compounds. Each experiment took approximately 40 min to complete. The self-paced reading task was programmed and administered using PsychoPy and was hosted online by Pavlovia. The online repository for the experiment, which includes the PsychoPy experiment code, is available at the following GitLab link: https://gitlab.pavlovia.org/danschmidtke/spr_compounds.

Participants

For the possible relations task with novel compounds presented in isolation, 30 consenting participants (ten female, 20 male) from the USA were recruited from Amazon Mechanical Turk. The median age of participants was 30 (IQR = 26.5). Participants were paid USD $1.50 to complete the experiment. The same task was hosted on the Amazon Mechanical Turk platform and was released to participants from the USA for the task involving existing compounds presented in isolation (refer to Schmidtke et al., 2018a for more details). The median participant age for this sample was 33 years (IQR = 17.5).

For the self-paced reading tasks, a desired minimum sample size of 40 participants per task was determined by consulting reference tables reported in Brysbaert and Stevens (2018). Based on the reference tables, the desired sample size was set at a minimum of 40 participants for each experiment. This ensured the self-paced reading experiments were sufficiently powered to detect an effect size of d = 0.4 for the interaction between Context type and Entropy of conceptual relations. Participants from the USA and Canada were recruited from the Prolific online platform. For the novel compound task 47 consenting participants were recruited (23 female, 22 male, 1 gender fluid, 1 undisclosed) with a median age of 34.5 years (IQR = 18). For the existing compound task 55 consenting participants were recruited (27 female, 26 male, 2 undisclosed) with a median age of 30 years (IQR = 18). Participants were paid USD $2.50 to complete the experiment. All participants were native English speakers. Participants across experiments were matched on gender, age and education level (see online Supplementary Materials S2). All studies were approved by the McMaster University Research Ethics Board (protocol 2020-4923).

Dependent variables

In Analysis 1, entropy of conceptual relations was the dependent measure. We computed entropy of conceptual relations based on this dataset. Entropy of conceptual relations is defined as \(H = -{\Sigma } p_{i} \log _{2}{p_{i}}\), where pi is the probability of a relation within the respective distribution of chosen possible relations for a given compound. For example, let us assume the compound toolbox had the following distribution of possible relations judgments; ‘box HAS tool’ (selected 10 times), ‘box FOR tool’ (7), ‘tool LOCATED box’ (5), and ‘box USED BY tool’ (3). The probability distribution for these relations would be .4, .28, .2, .12, which would result in an entropy value of 1.87 bits. To reduce noise and prevent relations that may have been chosen accidentally from affecting the results, only relations that were chosen more than once for a compound were included in the probability distributions. Figure 4 in the Appendix visualizes the language-wide distribution of conceptual relations, i.e., the probability distribution of conceptual relations collapsed across all compounds.

In Analysis 2, we analyzed log-transformed reading times, defined as the latency of the button press for the critical novel compound word in milliseconds (ms). Descriptive statistics for the raw and log-transformed reading latencies are provided in Table 3, along with descriptive statistics for all entropy values.

Table 3 Descriptive statistics for the dependent and independent variables used in the self-paced reading studies
Full size table

Independent variables

In Analysis 1, the presentation type of the compound (three levels: isolation, non-supportive context and supportive context) and the type of compound (two levels: existing and novel) served as the critical categorical predictor variables.

In Analysis 2, the self-paced reading data for existing and novel compounds was analyzed using two separate statistical models because the set of available lexical control variables differ for each compound type. There were two critical independent variables common to both analyses. The first is entropy of conceptual relations computed from data from compounds presented in context during the self-paced reading task, which we henceforth label as Entropy-c. The second is Context type (2 levels: supportive context vs. non-supportive context). We also fitted separate models including entropy values derived from the possible relations task in isolation, which we refer to as Entropy-i).

Lexical predictors

Frequency per million counts were obtained for the left constituent and right constituent as standalone (monomorphemic) words from the 1.9-billion-word Wikipedia corpus (Davies, 2015). We also included word length (in characters) in models for both analyses. Only the analysis of existing compounds included the frequency per million count of the compound word, also obtained from the Wikipedia corpus (Davies, 2015). For existing compound words, we also considered two semantic transparency measures from the compound word corpus made available by Schmidtke et al., (2021): left-whole semantic transparency and right-whole semantic transparency. Both semantic transparency measures capture the amount of meaning similarity between the whole compound word and the left constituent (left-whole transparency, e.g., air and airforce) and the right constituent (right-whole transparency, e.g., force and airforce). The semantic transparency estimates were provided by human raters on a scale from 1 to 7 (1 = not related, 7 = highly related). Table 3 provides distributional characteristics of all original and transformed lexical variables and reading times for existing and novel compound self-paced reading studies.

Separate correlation matrices of independent variables used in the self-paced reading task for existing and novel compounds are provided in Tables 4 and 5. The condition number was computed to assess the level of harmful collinearity that might be present between all lexical measures. The condition numbers for both existing (κ = 2.27) and novel compounds (κ = 1.26) indicate that there was no harmful collinearity present in the data (condition number > 30 indicates strong collinearity).

Table 4 Correlation matrix of lexical variables for existing compounds
Full size table
Table 5 Correlation matrix of lexical variables for novel compounds
Full size table

Statistical considerations

Linear mixed-effects regression models were used in Analysis 1 and Analysis 2. Mixed-effects regression takes into account group-level effects of multiple predictors on the dependent variable (fixed effects) in addition to individual variation between different participants and different items (random effects) (Baayen, Davidson, & Bates, 2008). In Analysis 1, Entropy of conceptual relations values served as the dependent variable. Context type (3 levels: isolation, non-supportive context and supportive context) and Compound type (2 levels: existing and novel) served as critical independent variables. Control variables included the frequencies of the left and right constituents. Random intercepts were included for items (compound).

In Analysis 2, we fitted linear mixed-effects regression models to log-transformed reading times. Separate models were fitted to log-transformed reading times for existing and novel compounds data. Random intercepts were included for participants and items for both models. Dummy coding was used for Context type in Analysis 2 (2 levels: non-supportive context and supportive context) with ‘non-supportive context’ set as the reference level. In addition to the lexical variables listed in Table 3, control variables included the position of the compound in the sentence, list type (2 levels: A and B), and the randomized trial number of the sentence. All frequency-based measures were log-transformed prior to inclusion in models.

All analyses were conducted in R, the open-source software for statistical computing (R version 4.0.5; R Core Team) using RStudio (RStudio Team, 2021a). The lme4 package (Bates, Mächler, Bolker, & Walker, 2015) was used to compute the mixed-effects regression models with restricted maximum likelihood (REML) estimations. P-values for models were obtained using the lmerTest package (Kuznetsova, Brockhoff, & Christensen, 2017), which uses Satterthwaite’s degrees of freedom method. All continuous measures were z-transformed prior to model inclusion. We refitted linear mixed-effects regression models after removing outlying residuals exceeding ± 2.5 standard deviations from the mean (Baayen & Milin, 2010). Effect size estimates for mixed-effects models were extracted from models in R using the effects package (Fox et al., 2019).

For Analysis 1 and 2, additional parameters were included to a model if the more complex model showed a statistically significant improvement in model fit. Improvement in model fit was assessed by conducting a Chi-square test on deviance statistics of models (Akaike information criteria). Omnibus effects were calculated for all analyses, with Type III model comparisons, using the car package in R (Fox & Weisberg, 2011). An alpha level of .05 was used for all statistical tests across all analyses. Post hoc contrasts in Analysis 1 were computed using the emmeans package in R (Lenth, Singmann, Love, Buerkner, & Herve, 2018). The Bonferroni p value correction adjustment was computed for the statistical tests applied to the post hoc contrasts. All plots were generated using the ggplot2 package (Wickham, 2011).

Results and discussion

For the possible relations task for novel compounds in isolation, 28 participants contributed a total of 3000 responses. We applied the same clean-up procedure as reported in Schmidtke et al., (2018a) for the possible relations data for existing compounds. To ensure that we did not consider responses that were made randomly, we removed all data points from any participants who selected the same relation for more than 50% of all trials. This resulted in the complete removal of data from five participants. In addition, 22 trials across all participants were removed due to no response being logged. The resulting data set consisted of 2279 judgments from 23 participants (17 male, 6 female).

The original data set for the self-paced reading task with novel compounds comprised 4700 trials (100 sentences, 47 participants). We eliminated all data from three participants who selected the same relation for more than 50% of trials. We also removed all trials from one participant who responded below 33 ms for critical compound trials for more than 50% of trials. In addition, the responses from three trials were not logged due to software error. The final data set contained 4297 trials from 43 participants (List A: 22 participants, List B: 21 participants).

For the self-paced reading task with existing compounds, the original data set included 5500 trials (100 sentences, 55 participants). We removed all trials from two participants who selected the same relation for more than 50% of trials. The responses for 47 trials were not logged due to software error. The final data set contained 5253 trials from 53 participants (List A: 28 participants, List B: 25 participants). For the analysis of reading times we excluded responses in the bottom 1% (below 133 ms for both tasks) and top 1% (above 4032 ms for the novel compound task and above 2758 ms for the existing compound task) of the reading time distribution.

Analysis 1

The goal of Analysis 1 is to determine the extent to which contextual information influences the ways in which relational meanings are generated for novel and existing compounds in different contexts. We analyzed two aspects of conceptual combination across compound types and contexts: semantic ambiguity and semantic flexibility.

Semantic ambiguity

We expect to see differences in the uncertainty, or ambiguity, of relational meanings for compounds across contexts. The possible relations task for compounds that are presented in isolation demands an interpretation for a compound in the absence of any contextual semantic cues that may help provide a more circumscribed meaning for the compound. We therefore predict that participants will be most uncertain about meanings of the compounds in the isolated condition, as reflected by highest entropy values. Since Shannon Entropy captures uncertainty in a system, greater entropy would indicate that participants tend to converge less on the set of possible relations for a compound word, signaling that the meaning for that compound is more ambiguous. Furthermore, we expect that entropy of conceptual relations will be lower for compounds embedded in any sentence context. We predict that after reading compounds in a sentence, participants will converge more on the same relations, reflecting a narrow hypothesis space of possible relational meanings of compounds. We expect that compounds presented in supportive contexts will have the lowest entropy values. This is because contexts with supportive semantic information will reduce the ambiguity of the meaning of a compound, leading to greater convergence in selecting a particular relational interpretation.

We also predict that, overall, participants providing relational interpretations for novel compounds will converge less on the same relational interpretations in comparison to participants providing responses to existing compounds. This prediction is based on the reasoning that when participants select relational interpretations to existing compounds, they will be able to draw on stored semantic knowledge about the compound based on prior exposures to the word. For novel compounds, a meaning has not been established, leading to a greater range of possible interpretations.

We tested these predictions in a mixed-effects regression model that examined the effects of compound type (existing and novel) and presentation type (isolation, supportive context and non-supportive context) on entropy of conceptual relations. We also explored the interaction between both variables to examine the possibility that novel and existing compounds are differentially affected by presentation type. Descriptive statistics of entropy of conceptual relations are provided in Table 6. The omnibus effects of the results of the mixed-effects model for entropy of conceptual relations is provided in Table 7. The results of the full mixed-effects regression model is reported separately in Appendix Table 11. The results are plotted in Fig. 1.

Table 6 Descriptive statistics for entropy of conceptual relations
Full size table
Table 7 Omnibus effects in the analysis of entropy of conceptual relations
Full size table
Fig. 1
figure 1

Box plots of entropy of conceptual relations broken down by compound type and presentation type. Each point represents a compound. Jitter added for readability

Full size image

The main effect of compound type was significant: as expected, entropy of conceptual relations was significantly greater on average for novel compounds [M = 1.85 bits, SE = 0.04 bits] compared to existing compounds [M = 1.58 bits, SE = 0.04 bits, t = 6.28, p <.001]. This result indicates that, on average, participants tended to converge more on the same relational interpretations for existing compounds versus novel compounds.

The main effect of presentation type was also significant and indicated that entropy of conceptual relations differed significantly across presentation types. As expected, entropy was significantly lower when the compounds were presented in supportive contexts [M = 1.38 bits, SE = 0.05 bits] compared to when they were presented in isolation [M = 2.28 bits, SE = 0.04 bits, t = -20.4, p <.001]. Entropy for compounds that were presented in non-supportive contexts was also lower [M = 1.49 bits, SE = 0.05 bits] compared to when they were presented in isolation [t = -17.44, p<.001]. These main effects suggest that participants were more variable when picking relational interpretations for compounds presented in isolation compared to compounds presented in a sentence frame.

There was a significant interaction of presentation type and compound type. Contrasts showed that this interaction was driven by differential entropy patterns in non-supportive vs. supportive conditions across compound type (see bottom two panels in Fig. 1). Entropy was greater for novel compounds compared to existing compounds in non-supportive contexts [t = -3.59, all p s <.001] (this contrast was also significant for the isolation condition), but there was no statistically reliable difference between compound types in supportive contexts [t = -0.84, p = .4]. In addition, entropy was greater for novel compounds in non-supportive contexts than when in supportive contexts [t = 3.89, p<.001], but this contrast was non-significant for existing compounds [t = 0.11, p = .99]. In summary, these results indicate that when embedded in semantically supportive contexts, participants tended to be equally systematic for both novel and existing compounds when choosing conceptual relations. Furthermore, when providing judgments to compounds that are embedded in sentence contexts, participants tend to be especially variable in choosing a relational meaning when exposed to a novel compound in a context that is non-supportive.

Semantic flexibility

The prior analysis showed that the meaning uncertainty of relational meanings of compounds behaved differently depending on presentation type and compound type. However, an analysis of entropy values is not ideal for examining the qualitative differences in meanings between novel and existing compounds across presentation conditions. We refer to the degree to which the same compounds may elicit different meanings across contexts as semantic flexibility. In this analysis, we used an index of overlap (Arita, 2017; Koch, 1957) to estimate semantic flexibility across the three presentation conditions (isolated, supportive context, and non-supportive context). The index of overlap for a compound is defined as \(\frac {(\frac {T - S}{n - 1})}{S}\), where n is the number of conditions in which that compound is presented, T is the grand sum of the total unique possible relations for each individual condition, and S is the total number of unique possible relations collapsed across all conditions. In keeping with the entropy measure, we restricted the dataset to include all relations that were selected more than once for a compound.

The semantic flexibility metric ranges from 0 to 1, where values closer to 0 mean that there is less meaning overlap for a compound across conditions, i.e., more semantic flexibility. Values closer to 1 indicate that there is more meaning overlap or less semantic flexibility: the same possible relational interpretations tend to be selected for a compound in each of the three presentation conditions. To provide an example of greater semantic flexibility, consider the novel compound grassladder. When embedded in a sentence with a supportive context – The girl is climbing up the grassladder in the yard – participants selected the following three relational interpretations: ‘ladder BY grass’ (selection probability = .12), ‘ladder LOCATION IS grass’ (.12), and ‘ladder MADE OF grass’ (.76). None of these relational interpretations overlapped with those selected for the same novel compound embedded in a non-supportive context – I’ll bring the grassladder into the house later. – which were ‘ladder FOR grass’ (.89), and ‘ladder USES grass’ (.11). Thus, this novel compound elicited a different set of relational interpretations in each context, yielding a value of .17 using the metric outlined above.

Existing compounds had significantly higher relational meaning overlap across conditions (M = .44, SD = .16) compared to novel compounds (M = .24, SD = .09), t(157.92) = 10.77, p<.001 (Fig. 2). This finding indicates that relational meanings tended to be more flexible for novel compounds compared to existing compounds in each presentation condition. In sum, not only were participants less certain about the relational meanings for novel compounds compared to existing compounds, especially when presented in isolation, participants also tended to choose qualitatively different sets of relational meanings for novel compounds compared to existing compounds.

Fig. 2
figure 2

Box plots of meaning flexibility by compound type. Each point represents a compound. Jitter added for readability

Full size image

Analysis 2

The RICE account proposes that conceptual relations compete for selection during compound word processing (Spalding et al., 2010). The goal of Analysis 2 is to examine whether this aspect of the RICE account holds when existing and novel compounds are processed during sentence reading. An effect of entropy of conceptual relations (either entropy-c or entropy-i) on the processing speed of both existing and novel compounds during sentence reading would confirm that relational competition is a general property of compound word processing during sentence reading. Based on Schmidtke et al. (2016; 2018a), we expect the effect of either entropy measure, if found, to be inhibitory for both compound types: greater entropy is expected to be associated with longer reading times. Furthermore, we expect that the entropy values derived from the possible relations tasks that were implanted in the self-paced reading task (entropy-c) to elicit a stronger effect than the entropy values derived from the experiment in which compounds were presented in isolation (entropy-i). We base this expectation on the reasoning that (i) the probability distribution of possible relations for compounds presented in the self-paced reading would give rise to an entropy value that is attuned to the context that the word appeared in, while this would not be possible for compounds presented in isolation, and (ii) the entropy-c measure is calculated from the same participants that read the compounds.

In addition, we tested whether the context in which a compound is presented influences the extent to which conceptual relations are engaged. Based on Gagné and Spalding (2004b), though meaning supportive contexts are expected to facilitate word processing, the presence of context itself is not expected to eliminate the engagement of conceptual relations. We examined this hypothesis for novel and existing compound processing by testing the interaction between each measure of entropy of conceptual relations (entropy-c and entropy-i) and context type. A significant entropy of conceptual relations (entropy-c or entropy-i) × context type interaction would indicate that the level of competition between conceptual relations during reading is moderated by context, but the absence of a significant interaction effect would suggest, in support of Gagné and Spalding (2004b), that the strength of the competition effect is not different across supportive and non-supportive contexts. The omnibus effects of the results of the mixed-effects models for existing and novel compounds are provided in Table 8. The full results of the respective models are reported separately in Appendix Tables 12 and 13.

Table 8 Omnibus effects in the analysis of reading times (log) in the self-paced reading tasks for existing and novel compounds
Full size table

There was a significant main effect of entropy-c on novel compound reading times [\(\hat {\beta }\) = 0.018; SE = 0.008; p = .036], and on existing compound reading times [\(\hat {\beta }\) = 0.015; SE = 0.005; p = .004]: in both studies greater entropy led to longer reading times. The effect size, defined as the amount of change in reading times estimated for the contrast between the minimum and the maximum value of entropy-c, was 44 ms for novel compounds and 22 ms for existing compounds.

There was a significant main effect of context on novel compound reading times [\(\hat {\beta }\) = 0.03; SE = 0.014; p = .035] and existing compound reading times [\(\hat {\beta }\) = -0.021; SE = 0.01; p = .036]. Novel compounds that were embedded in supportive contexts were read 17 ms slower than novel compounds embedded in non-supportive contexts. However, the opposite trend was true for existing compounds: existing compounds were read 8 ms faster in supportive contexts than when in non-supportive contexts. The partial main effects of entropy-c and context type on existing and novel compound reading are visualized in Fig. 3 (to aid interpretability, plots depict back-transformed values of reading times, in milliseconds).

Fig. 3
figure 3

The partial main effects of entropy-c (scaled) and context on reading times of novel compounds during self-paced reading

Full size image

The main effect of entropy-i did not reach statistical significance in reading times for novel compounds [\(\hat {\beta }\) = -0.008; SE = 0.01; p = .428] and existing compounds [\(\hat {\beta }\) = 0.01; SE = 0.007; p = .124]. A chi-square goodness of fit on deviance statistics of existing and novel compounds showed that the addition of entropy-i did not improve model fit [existing compounds: χ2(1) = 2.27, p = .13; novel compounds: χ2(1) = 0.63, p = .43]. The interactions of context type with entropy-c or entropy-i did not reach statistical significance for existing and novel compound reading times. The inclusion of interactions between either entropy measure and context did not improve model fit [existing compounds: χ2(2) = 0.41, p = .52; novel compounds: χ2(2) = 1.72, p = .19], so we did not include these interactions in our final model.

The results of Analysis 2 demonstrate that there is a processing cost to novel compound word reading when no single possible relational interpretation dominates over all other relational meanings. This finding indicates that novel compound word processing involves constructing multiple relational meanings and that these meanings compete with one another. Further, the same effect of entropy-c was present in existing compound word recognition and was also not moderated by context, suggesting that for both existing and novel compounds, the competition effect is reliable and operates independently of context. We discuss these results in the General discussion.

Finally, we report the effects of predictors of compound word processing that are not relevant to the current study. There was an effect compound length on both novel [\(\hat {\beta }\) = 0.057; SE = 0.01; p <.001] and existing compound types [\(\hat {\beta }\) = 0.029; SE = 0.005; p <.001]: as expected longer compounds were read more slowly. For existing compounds, the effect of compound frequency was significant [\(\hat {\beta }\) = -0.012; SE = 0.005; p = .026] and was in the expected direction: more frequent compounds tended to be read faster than infrequent compounds. We also found a significant effect of left-whole semantic transparency [\(\hat {\beta }\) = 0.013; SE = 0.005; p = .019]. Higher semantic similarity between the meaning of the left constituent and whole form led to longer processing times. The direction of the effect is surprising given the amount of evidence showing that high semantic similarity is associated with faster reading times (Schmidtke et al., 2021; Gagné et al., 2019; Kim, Yap, & Goh, 2019). One possibility for this result is that greater semantic similarity imposes a discrimination cost on processing whereby the closely related meanings of the left and whole word make it difficult for the cognitive system to reduce the uncertainty between the outcomes associated with each word under comparison (see also Schmidtke, Van Dyke, & Kuperman, 2018). Since this result is not the focus of the present study, we do not discuss it further.

General discussion

A series of experiments reported herein were designed to test two hypotheses about the roles of relational competition and context on the processing of existing and novel compounds. The first hypothesis is that both novel and existing compound word reading times should be sensitive to a common metric of competition between conceptual relations. This hypothesis was confirmed by the self-paced reading time data: entropy of conceptual relations was a significant predictor reading times to novel and existing compounds, such that greater competition between a compound word’s possible conceptual relations leads to longer reading times. The second hypothesis is that relational processing is competitive irrespective of whether compounds are encountered in context or in isolation. The self-paced reading data supports this hypothesis. We report that entropy of conceptual relations predicted reading times to compounds embedded in sentences, and that this effect was equally strong in sentences that supported a particular compound word meaning versus those that did not.

Our results were presented across two analyses. In Analysis 1, we first examined differences in the kinds of relational interpretations generated by participants across three presentation conditions (isolation, supportive sentence contexts and non-supportive contexts). This analysis provided a preliminary quantitative overview of differences in the ambiguity and flexibility of relational information between existing and novel compounds. In Analysis 2, we turned to the effect of conceptual relations on processing times during sentence reading. In what follows, we discuss the results of each analysis in turn.

Differences in semantic ambiguity and flexibility across novel and existing compounds

Semantic ambiguity

Context serves to reduce meaning ambiguity of compound words. Analysis 1 showed that entropy of conceptual relations was lower for both novel and existing compounds after they were read in sentences compared to when presented in isolation. This result indicates that participants tend to choose a more restricted set of relational interpretations for compounds when they are encountered in a sentence context. This aligns with the conclusion drawn by Gagné et al., (2005) that the space of possible semantic interpretations of a compound is affected by sentence context.

Analysis 1 also showed that whenever there was a reliable difference in entropy of conceptual relations between compound types, entropy was always greater for novel compounds. This result confirms that, as expected, relational meanings for novel compounds are more diffuse in comparison to existing compounds. A straightforward explanation for this pattern is that participants draw upon stored semantic knowledge when selecting plausible conceptual relations for existing compounds in the possible relations task (for an extensive discussion on how individual print experience and compound frequency affect flexibility and precision of relational knowledge see Schmidtke et al., 2018a). Perhaps the purest support for this hypothesis is in the comparison between the entropy of novel and existing compounds in the isolated condition: entropy of conceptual relations was significantly lower for existing compounds even when there was no contextual information to narrow the space of possible meanings. In sum, even in the absence of supporting contextual information, participants tend to rely upon existing semantic knowledge more readily when deciding upon the relational interpretations of existing compounds compared to novel compounds.

Interestingly, the results of Analysis 1 indicated that possible relations judgments of existing and novel compounds were differentially affected by context type. That is, supportive vs. non-supportive contexts yielded differences in entropy values for novel compounds but not for existing compounds. When selecting meanings for existing compounds, participants were equally systematic when the same compound was encountered in a context that aided a narrow interpretation of the compound vs. a context that did not provide specific semantic support for a plausible meaning.

This was not the case for novel compounds. Participants were less variable in selecting a plausible conceptual relation for novel compounds after they were read in supportive contexts compared to when encountered in non-supportive contexts. Only when novel compounds were presented in meaning supportive contexts was the degree of convergence upon relational meanings equivalent to that of existing compounds. Since the meanings of novel combinations are not established in memory, readers may rely more heavily on available contextual information to help choose a plausible meaning. Altogether, these results suggest that relational meanings of novel compounds are more sensitive to contextual information than existing compounds.

Semantic flexibility

Were there qualitative differences in the relational meanings for novel and existing compounds across presentation types? The results of the semantic flexibility metric in Analysis 1 suggest so. Participants tended to choose overlapping sets of relational interpretations when existing compounds were presented in isolation or in either context type, but were more likely to choose different sets of relational meanings for novel compounds in each of the presentation formats. These results further support the idea that the types of relational meanings ascribed to novel compounds tend to be more flexible and sensitive to the contexts in which they are embedded.

In sum, Analysis 1 showed that relational meanings of novel compound words are more ambiguous in comparison to lexicalized compounds. Furthermore, compared to existing compounds, participants tended to select different sets of meanings for novel compounds across contexts. We argue that the relational meanings of existing compounds are tied to the meanings of their stored lexical representations in memory, but that conceptual combination of novel compounds relies instead on sources of available semantic information from the surrounding context. These findings are consistent with the extensive literature on the critical role of contextual information during novel word learning (Adelman, Brown, & Quesada, 2006; Chaffin, Morris, & Seely, 2001; Johns, Dye, & Jones, 2016; Pagán & Nation, 2019; Snefjella, Lana, & Kuperman, 2020).

Competition between conceptual relations affects novel and existing compound word processing in context

In support of the RICE hypothesis, the results of Analysis 2 provide evidence for relational competition during the processing of both novel and existing compounds: greater entropy among possible conceptual relations inhibited reading times for both novel noun-noun combinations and lexicalized compounds during sentence reading. These results represent two new contributions to the field of compound processing. First, as far as the authors are aware, the effect of entropy of conceptual relations (Schmidtke et al., 2016; 2018a) has not been previously shown to affect novel compound word processing. This study therefore demonstrates, via a common measure, that relational competition is a general property of novel and existing compound word reading. Second, this is the first occasion that competition between conceptual relations has been shown to affect compound processing during sentence reading rather than in isolated word processing.

Linking the results of Analysis 1 and 2 together, it is possible to conclude that relational competition occurs even when the same compounds did not always elicit a fully overlapping set of competing relational interpretations across different contexts. Recall that in Analysis 1 we found that the sets of relational meanings that participants selected tended to overlap across different presentation formats for existing compounds, but did not overlap as much for novel compounds. We therefore argue in favour of the account posited by Gagné et al., (2005) that relational competition is an automatic process, even when the types of activated possible interpretations are particular to the context in which a compound appears. In sum, despite each sentence frame giving rise to different sets of possible meanings for the same compound, ambiguity in the resulting sets of meanings were equally predictive of lexical processing in each condition.

Contradicting the results of the present study, as stated at the outset of this study, the entropy effect was not reliably found in eye movements during sentence reading (Schmidtke et al., 2021). There are two possibilities for these conflicting results. The first possibility concerns differences in task demands. Although both studies involved sentence reading, the eye-tracking study reported by Schmidtke et al., (2021) was a naturalistic reading task that recorded a very different behavioral measure of processing. That is, instead of serial presentation of words and a key-press response time, participants silently read sentences that were presented fully on a screen while eye movements were tracked. Participants were also given an additional task in the current study. Participants were required to make a possible relations judgement after the sentence, which may have caused participants to pay more attention to conceptual relations.

The second possibility is that detecting an effect of entropy during sentence reading might depend on whether the entropy values for compounds are derived from judgments made only when the compounds are embedded in the same sentences that are used in the reading task itself (entropy-c). Only the entropy-c measure, and not the entropy-i measure, reliably predicted processing speed for novel and existing compounds. Although the entropy-c measure was predictive of reading times, as noted above, there was no interaction between context type. This result leads us to speculate that the competition effect is observable when the corresponding measure of meaning uncertainty is attuned to sentence context, and that the size of the effect is not necessarily moderated by the amount of semantic support in the preceding context. Therefore, a pressing question is whether entropy-c affects eye movements during the reading of existing and novel compounds embedded in the same sentences used in this study. This would help disentangle the two possibilities for lack of a significant effect of conceptual combination in the eye-movement record.

General processing differences across context type

Novel compounds were processed around 200 ms more slowly than existing compounds overall. However, there was a difference in the direction of the effect of context between compound types. Supportive contexts aided the processing of existing compounds, such that compounds in supportive contexts were processed significantly faster than the same compounds embedded in non-supportive contexts. Conversely, the overall processing of novel compounds was significantly slower in the supportive context condition relative to the non-supportive context condition. We interpret our result as a manifestation of the Noticing Hypothesis (1991). That is, readers spend more time when reading novel compounds since readers devote greater attention to figuring out the possible meaning of unfamiliar words, and devote even more effort inferring the meaning of the novel compound when it is possible to integrate the meaning more readily with the preceding sentence context. We note that this pattern of results for novel compounds did not conform with the results presented by Gagné and Spalding (2004b), which we speculate may be a consequence of task differences. In Gagné and Spalding’s study, participants were asked to provide an acceptability judgement after reading the entire sentence. This response type was a measurement of the time taken to process a relational interpretation that was readily provided to the participant and not the processing time on the compound word as it is encountered in the sentence itself. Therefore, in Gagné and Spalding’s study, the faster response time in the supportive context condition relative to the neutral context may have reflected post hoc evaluation of the consistency of the meaning of the relational phrase with the preceding sentence without capturing the online processing associated with integrating a plausible novel compound meaning with the preceding sentence, as outlined in the Noticing Hypothesis.

Conclusions

In conclusion, this study provides confirmation of the RICE hypothesis (Spalding et al., 2010), demonstrating, via a common measure, that possible conceptual relations compete for selection during the processing of existing and novel compound words during sentence reading. The amount of semantic support that is present in the preceding sentence context differentially constrains the flexibility and ambiguity of the sets of possible relational meanings for novel and existing compounds. Despite these constraining factors, competition between relational meanings remains a stable and consistent predictor of reading times of novel and existing compounds during sentence reading.