If Birds Have Sesamoid Bones, Do Blackbirds Have Sesamoid Bones? The Modification Effect With Known Compound Words

Spalding, Thomas L.; Gagné, Christina L.; Nisbet, Kelly A.; Chamberlain, Jenna M.; Libben, Gary

doi:10.3389/fpsyg.2019.01570

ORIGINAL RESEARCH article

Front. Psychol., 09 July 2019

Sec. Psychology of Language

Volume 10 - 2019 | https://doi.org/10.3389/fpsyg.2019.01570

This article is part of the Research Topic Words in the World View all 18 articles

If Birds Have Sesamoid Bones, Do Blackbirds Have Sesamoid Bones? The Modification Effect With Known Compound Words

$\r\nThomas L. Spalding*$ Thomas L. Spalding^1*

Christina L. Gagné¹

Kelly A. Nisbet²

Jenna M. Chamberlain¹

Gary Libben³

¹Department of Psychology, University of Alberta, Edmonton, AB, Canada
²Department of Linguistics and Languages, McMaster University, Hamilton, ON, Canada
³Department of Psychology, Brock University, St. Catharines, ON, Canada

Three experiments investigate how people infer properties of compound words from the unmodified head. Concepts license inference of properties true of the concept to instances or sub-types of that concept: Knowing that birds generally fly, one infers that a new type of bird flies. However, different names are also believed to reflect real underlying differences. Hence, a different name creates the expectation that a new bird differs from birds in general, and this might impact property inference. In these experiments, participants were told, Almost all (Some, Almost no) birds have sesamoid bones, and then asked, What percentage of blackbirds (birds) have sesamoid bones? The results indicate both inference and contrast effects. People infer properties as less common of the compound than the head when the property is true of the head, but they infer them as more common of the compound than the head when the property is not true of the head. In addition, inferences about properties true of the head are affected by the semantic similarity between the head and the compound, but properties not true of the head do not show any semantic similarity effect, but only a small, consistent effect of contrast. Finally, the presentation format (Open vs. Closed compounds) affects the pattern of effects only when the spacing suggests the existence of a permanent name.

Introduction

Much research on compound words (words that consist of two or more free morphemes, e.g., snowball or hogwash) focuses on the processing involved in accessing the words (see, e.g., Libben, 1998; Libben and Jarema, 2006, for reviews) for use in specifically language-related tasks. There is much less work on how compound words are used more broadly in human cognition. In this paper, we investigate how compound words function in, and contribute to, human cognition more broadly. In particular, we are interested in what support compound words provide to conceptual (or categorical) inference. It is well accepted that a major function of concepts in human thinking is to provide the ability to infer properties from the concept to members of the category named or referred to by the concept (see e.g., Murphy, 2002; see also, Osherson et al., 1990). Thus, if one knows that birds have sesamoid bones, then one can make a reasonable inference that a particularnew bird has sesamoid bones, even though one has no other information that would specifically indicate this fact for this particular bird. Similarly, one might make a reasonable inference that blackbirds, in general, have sesamoid bones, if birds, in general, are believed to have sesamoid bones. Importantly, such inferences are probabilistic and defeasible, rather than strictly logical deductions (see Osherson et al., 1990), such that, for example, specific information about a particular item or sub-group may override the inference (e.g., knowledge that penguins cannot fly rules out the inference that penguins can fly, just because they are a kind of bird), as can the typicality of the sub-group with respect to the group (e.g., property inference is more likely from bird to robin than from bird to, say, turkey, even in the absence of knowledge about the specific property). However, the point of categorical inference is precisely that the category allows one to infer properties where there is no specific knowledge about the item or sub-group that can be brought to the question.

On the other hand, it is equally well accepted that people expect that different names for things reflect real underlying semantic differences. Indeed, the literature contains three proposed principles of human cognition (Synonomy Avoidance, Carstairs-McCarthy, 2010; Principle of Contrast, Clark, 1993; and Mutual Exclusivity Principle, Markman, 1989), each of which shares this core notion of expectation that different names reflect real underlying differences. While these principles are generally framed in terms of when a new name is, in some sense, justified, and how people make that decision, it is also the case that when a new name is presented as established, this leads to an expectation of some real differences from things that already have other names. Importantly, this expectation seems to be quite general, arising even when there is no specific knowledge of an existing difference. Taking these principles into account, then, one might expect that birds and blackbirds should be assumed to differ in significant ways, and this expectation might then affect the process of inferring properties from the head to the compound word. Thus, perhaps although birds generally do have sesamoid bones, blackbirds might not.

Clearly, then, we have two well-attested principles, which seem to work in opposite directions in terms of property inference: That categories license inference of properties to new sub-categories or category members, and that new labels, for example, of a sub-category, indicate property differences (i.e., a lack of licensed inference) for new sub-categories or category members. It is important to understand how these two principles operate together, as property inference is a major communicative function: Property inference (or lack thereof) creates expectations of novel objects, people, situations, and so on, based on what is known or said of an existing set, so that one is in a better position to deal with that novel object, person, or situation. This predictive function is also critical in numerous areas of applied work, for example in natural language processing, where understanding the expectations of a user is critical to the success of the system. As one example, research aimed at improving information retrieval engines (e.g., Baldwin et al., 2010; Wang et al., 2010) by identifying possible alternative terms that might be used to facilitate information access is one example where it is critical to understand what humans expect about the meaning of new terms. Similarly, understanding property inference can contribute to research on natural language question answering (for an overview, see Hirschman and Gaizauskas, 2001). For example, Wang et al. (2011) presented methods for identifying discourse structure for online forum data (see also Wang et al., 2010). The dialogue acts within these threads have various structures such as Question–Question, Answer–Answer, Question–Additional Information, and the meaning of terms in these dialog acts depends, in part, on the structure of those dialog acts (e.g., similar expressions can carry somewhat different meanings and referents, depending on whether they are embedded in an answer or in a request for additional information). Discovering the factors that influence the way in which the expression is used to refer to a referent and the properties of that referent can aid the development of NLP procedures used to automate the identification of dialogue acts. In addition, systems that attempt to automatically extract sentiment must be built keeping in mind the ways in which people infer properties of objects as existing labels are used and new labels are introduced (e.g., Maynard and Funk, 2012; Dragos et al., 2018). In sum, understanding how property inference operates in light of these two contrasting principles has important scientific, and also applied, consequences.

We begin by reviewing what is known about how people extend properties of concepts to novel combined concepts, and then turn to the question of whether people extend properties to known compound words in the same way. The question of when and which properties of a combined concept become available during conceptual combination has been one of the core questions within the conceptual combination literature. Research on this topic initially examined whether properties of the constituent concepts are available prior to properties of the whole concept. Early research by Springer and Murphy (1992) found that people were faster to verify properties that were true of the phrase (e.g., peeled apples are white) than properties that were true of the head concept, prior to modification (e.g., peeled apples are round). Gagné and Murphy (1996) found that discourse context did not alter this pattern.

More recently, the question of property inclusion has been examined in the context of examining whether the availability of properties differs for the head noun concept (e.g., ducks have webbed feet) relative to a modified concept (e.g., baby ducks have webbed feet). When using novel combinations, this work reveals a robust set of effects called the modification and inverse modification effects (see Spalding and Gagné, 2015, for a demonstration of both modification and inverse modification effects, but for modification effects using other property verification tasks see also Connolly et al., 2007; Jönsson and Hampton, 2008, 2012; Gagné and Spalding, 2011, 2014b; Hampton et al., 2011). In particular, properties generally true of the unmodified head noun become less true of the modified head (modification effect), while properties generally false of the unmodified head become less false of the modified head (inverse modification effect). Thus, for example, purple candles are judged less likely to be made of wax than candles, but purple candles are also judged more likely to have teeth than candles. The modification effect is robust over a range of specific verification tasks, including ratings of likelihood of the truth/plausibility of a property for a category (e.g., Connolly et al., 2007), true/false decisions about the property’s relation to the concept and the response times to make those decisions (e.g., Gagné and Spalding, 2011), and estimates of the percentages of category members for which the property is true (e.g., Spalding and Gagné, 2015). Also, the modification effect is very robust over a wide range of property typicality, including properties that seem to be nearly definitional of the head, such as being animate for lamb (see e.g., Jönsson and Hampton, 2008, 2012; Hampton et al., 2011). This robustness over various kinds of properties is unexpected by those theories where prototypicality of features should be a determining aspect of property verification, such as prototype theories of conceptual combination (e.g., Hampton, 1991), but also theories of the semantics of compound words which differentiate between “levels” of properties (e.g., the skeleton vs. body distinction in Lieber, 2004).

Although the modification effect was initially used to test hypotheses about whether or not properties of a combined concept are directly inherited from the constituent concepts (e.g., Connolly et al., 2007), there are some findings to suggest that this effect might not actually be driven by the process of conceptual combination (i.e., constructing a new concept based on the conceptual “contents” of the modifier and head) per se, but rather by reasoning about the combined concepts. Gagné and Spalding (2011, 2014b, 2015), Spalding and Gagné (2015), Gagné et al. (2017) present evidence that the modification effects primarily arise due to meta-knowledge of modification, and particularly to reasoning about the expected relation of the meaning of the combined concept and the head. In particular, they argue that the effects are largely driven by the expected level of contrast (i.e., matching or mismatching features) between the combined concept and the head, rather than by conceptual knowledge of the individual constituent concepts (as would be expected by e.g., Hampton, 1987, 1991; but also by many approaches to the semantics of known compound words such as Lieber, 2004, 2009; see Gagné and Spalding, 2015, for a discussion). For example, modification effects arise even when the modifier is a non-word, and thus cannot contribute any semantic or conceptual information about what properties are appropriate for the combined concept (e.g., Spalding and Gagné, 2015).

In short, the literature on the modification effect strongly suggests that (a) the inferential function of concepts does indeed extend to modified versions of those concepts, rather than only to individual members of the category picked out by the concept, but that (b) the modification and inverse modification effects are driven by people’s expectations about the nature, purpose, and use of modification. Thus, it seems likely that the general principle that a different name implies other, underlying differences (Synonomy Avoidance, Carstairs-McCarthy, 2010; Principle of Contrast, Clark, 1993; and Mutual Exclusivity Principle, Markman, 1989) should lead to modification and inverse modification effects with known compounds. However, in the existing literature, the “different names” created by modification are not well-established, but rather are novel. This novelty could have two different kinds of influence on the modification effect. It could be that novel names are simply seen as less established or less permanent names, and therefore they might lead to smaller modification effects (i.e., more likelihood of property inference), or it could be that the novelty makes the contrast more immediately salient, and therefore leads to larger modification effects (i.e., less likelihood of property inference). Furthermore, because of the novelty of the modified concepts, the existing literature is unable to investigate the way in which these principles interact with the semantic knowledge that is inherent in a category of things with a well-established name, or, indeed, with the simple fact of the well-established name. Nevertheless, there is good reason to believe that much of the processing and semantics of compound words is similar to that of novel conceptual combinations (see, e.g., Gagné and Spalding, 2014a). Thus, we expect to see modification effects (and inverse modification effects) with known compounds.

In the current experiments, we investigate the seeming conflict between the basic cognitive principles that we infer properties based on category membership and that the use of different names for things implies real underlying differences (Synonomy Avoidance, Carstairs-McCarthy, 2010; Principle of Contrast, Clark, 1993; and Mutual Exclusivity Principle, Markman, 1989), in three experiments. In particular, we investigate the extent to which people are willing to infer properties from head nouns to compounds under various conditions that should affect the extent to which people believe the compounds to be well-established as different names, and thus should bring the principle that different names imply different properties more into conflict with the principle that properties can be inferred from categories to more specific sub-sets. Experiment 1 investigates whether the modification and inverse modification effects occur for known, transparent compound words and further investigates whether semantic knowledge of the compounds and their relation to the head nouns plays a role. Experiment 2 replicates Experiment 1, and in addition investigates the effect of spacing, under the assumption that the lack of spacing would indicate the compounds as more established, permanent, unique names for existing sub-categories. Experiment 3 is similar to Experiment 2, except that we replaced the modifier of the compound with a non-word, in order to find out if the structural cue (spacing) would affect property inference when the “compound” is not a known category.

Experiment 1

This experiment investigates how people infer properties of transparent compound words based on their knowledge of those properties’ relationship with the head noun, using the method from Spalding and Gagné (2015). This method asks participants to estimate the percentage of members of a category (e.g., buds) or subcategory (e.g., rosebuds) that have a particular property.

To directly manipulate the truth value of the property, we used blank predicates (i.e., properties that use relatively unfamiliar terms but are relevant for the concept in question, e.g., biological predicates are used for animal categories, see Osherson et al., 1990). The likelihood of the property for the unmodified noun was manipulated by telling participants that Almost All, Some, or Almost No members of the head noun concept had it. Critically, because we used the same blank properties in the Almost all, Some, and Almost No conditions, not only is the semantic content of those properties relatively unfamiliar and/or unrelated to the compounds, but to the extent it is familiar, the semantic content is controlled across the likelihood manipulation.

Based on previous work showing that processing compound words appears to involve many of the same processes as the processing of novel combined concepts (e.g., Gagné and Spalding, 2009), we predict that in making a decision about whether a property of the head is true of a compound word, people will show a pattern similar to that previously demonstrated for novel modifier-noun pairings: Modification effect for properties true of the head and inverse modification effect for properties false of the head.

Methods

This and the following experiments were reviewed for ethical content and approved by the Research Ethics Committee at the University of Alberta, and all participants provided written consent.

Participants

Sixty-two participants took part in the study. Participants in this and all following experiments were undergraduate students enrolled in a very large first year psychology class and obtained partial class credit for participating. In this participant population, approximately 95% are between 18 and 24 years of age and approximately 70% are female. All participants in this and the following experiments self-identify as native English speakers. In this, and the following experiments, target sample size is determined by expected effect size and complexity of design. The number of actual participants is determined by the target sample size and availability of participants in the Departmental participant pool. In this experiment, we set a target of 10 participants per condition. Two extra participants were included (in our pool, if extra participants attend a session, they must be run).

Materials and Design

Experiment 1 used 96 transparent compounds (e.g., snakeskin) selected from a previously categorized set of items (Ji et al., 2011). Statements to be predicated of the heads of the compounds were then selected to match each compound word. The truth of these statements, relative to the compound words or the head nouns of the compound words, were expected to be unknown by the participants. The unknown predicates were taken from previous experiments on the modification effect (Spalding and Gagné, 2015) and from Wikipedia searches for uncommon words related to the head noun. The compounds and the predicates are presented in the Appendix. The design is a 2 (Modification: modified vs. unmodified) by 3 (Likelihood: almost all, some, almost no) crossed factorial design.

Procedure

On each trial, participants were first shown a statement regarding how often an unknown property is true of an unmodified noun using one of three quantifiers: Almost all, Some, or Almost no. For example, participants might see “Almost all birds require graminoids in their diet.” The participants were instructed to treat this statement as true. Participants were then asked a follow up question about how many members of the unmodified noun or the modified noun category have that same property. They were asked to respond on a scale of 1–100. For example, they would be asked either “What percentage of birds require graminoids in their diet?” or “What percentage of blackbirds require graminoids in their diet?”. The lists were counterbalanced so that each participant saw either the compound word or the head noun in the question.

Results and Discussion

Data Analysis

The descriptive statistics are shown in Table 1. We analyzed the data using linear-mixed effects (LME) regression models in which subject and item were entered as random effects, and Modification (modified vs. unmodified) and Likelihood (Almost all, Some, Almost no) were entered as fixed effects, using the mixed and contrast commands in Stata (StataCorp, 2017). The mixed function outputs coefficients (i.e., estimates) for simple effects at the first level of the other categorical variables and at the mean of the other continuous variables in the model (see Table 2). For testing our hypotheses, these coefficients are not directly interpretable because they represent the simple effect of a variable at the first level of other variables. The relevant statistical tests for addressing our research questions concern interactions and simple effects, which are reported in the following text. The contrast function in Stata was used to conduct these analyses. We report the tests conducted on these fixed effects. Tests of simple effects (the effect of a factor at one level of another factor) were conducted to follow up on statistically significant interactions, because in the case of significant interactions, the main effects are not informative. Some statistics packages report LME main effects as F-tests and the simple effects as t-tests. However, because the degrees of freedom are indeterminate for such tests in linear mixed effect models, Stata uses chi-square and Z-scores, respectively.

TABLE 1

Table 1. Mean (SE) judged percentage of category members having the test property by level of Likelihood from Experiment 1.

TABLE 2

Table 2. Experiment 1 Mixed Model Coefficients.

Results

The analysis indicated a significant interaction between Modification and Likelihood, X²(2) = 435.0, p < 0.001. Analysis of the simple effects indicated that the Almost All and Some conditions both led to a significant modification effect, z = 26.9, p < 0.001 and z = 10.1, p < 0.001, respectively, while the Almost No condition led to a significant inverse modification effect, z = -2.8, p < 0.01. That is, the property was judged less likely for the compound, if the property was presented as true of almost all or some of the head category members, but more likely for the compound, if the property was presented as true of almost no members of the head category. Clearly, known compound words give rise to a very robust modification and inverse modification effect, as predicted.

These results show that modification and inverse modification effects that have been reported for novel phrases extend to lexicalized compounds. However, there are several more specific points that should be noted. First, the modification effects are much larger than those found in previous studies with novel phrases. For example, in the current experiment, modification in the Almost All condition reduced the likelihood of the property by some 24 percentage points, while Spalding and Gagné (2015) found a reduction of only 2 percentage points, using exactly the same experimental paradigm and participant population. This is consistent with the notion that the degree to which the compound word is seen as a unique, permanent, established name is likely to make people believe more strongly that there are real, underlying differences between the things named by the head and by the compound. Second, there is a strong asymmetry between the true and false features (i.e., between the size of the modification and inverse modification effects), suggesting that the existence of a known compound affects the features presented as generally false of the head noun much less than those presented as generally true of the head.

To investigate the pattern of asymmetry between the true and false features further, and to investigate the role of the semantics of the existing compound word, we performed a post hoc analysis, in which we entered the semantic similarity between the known compound and the unmodified head, as measured by Latent Semantic Analysis (LSA; Landauer and Dumais, 1997), into an LME regression with the Likelihood factor (Almost All, Some, Almost No), using only the percentage estimates for the Modified condition (see Table 3). Given that the compound was never seen in the unmodified condition, it would not be meaningful to include these items when examining the impact of the similarity between the compound and head. That is, the unmodified condition is when the head is presented for the property judgment, thus those data points reflect the judgment made when the participant is not presented with the compound word, and hence the semantic similarity between the compound and head is not meaningful in these cases (e.g., if the participant judges the likelihood that birds require graminoids in their diet, they will not have seen blackbird in the experiment at all, and hence the relationship between blackbird and bird is completely irrelevant). We observed a significant interaction between the LSA measure (e.g., the association between the word birds and the word blackbirds) and Likelihood, χ²(2) = 38.2, p < 0.001. Further analysis indicated a significant slope for LSA in the Almost All and Some conditions, z = 5.8, p < 0.001 and z = 3.0, p < 0.005, respectively. However, the semantic similarity between the compound and the unmodified head did not affect the estimates in the Almost No condition, z < 1. See Figure 1. In short, the modification effects were significantly smaller for compounds that were more semantically similar to their heads in the Almost All and Some conditions, but semantic similarity between the compound and the unmodified head had no effect in the Almost No condition.

TABLE 3

Table 3. Experiment 1 Mixed Model Coefficients with LSA: Modified condition only.

FIGURE 1

Figure 1. Effect of semantic similarity (LSA) on judged likelihood of property in Experiment 1.

To further examine whether property inference is influenced by degree of semantic transparency, we also used semantic transparency ratings by human participants from a database containing semantic transparency ratings for over 8000 compounds (Gagné et al., in press). The judgment that is relevant for the current experiment is the rating between the head noun and the compound. Participants were asked to judge on a scale from 0 to 100% how much the head noun retained its meaning in the compound (e.g., How much is the meaning of birds retained in the meaning of blackbirds). This information was available for 71 items. The mean rating was 79% (SD = 13) and ranged from 52 to 96%. This judgment was entered into a model that also included Likelihood and was restricted to only the modified concept condition (see Table 4). There was an interaction between transparency rating and Likelihood, χ²(2) = 49.99, p < 0.0001. Analysis of this interaction indicated a significant slope for the transparency rating in the Almost All and Some conditions, z = 7.77, p < 0.0001 and z = 3.32, p < 0.002, respectively, with the slope for transparency judgments being steeper in the Almost All condition than in the Some condition, χ²(1) = 13.83, p < 0.001. As transparency increased, ratings for the properties increased for the modified items, meaning that more transparency would correspond to smaller modification effects. However, transparency judgments did not affect the estimates in the Almost No condition, z = 0.19, p = 0.852. See Figure 2.

TABLE 4

Table 4. Experiment 1 Mixed Model Coefficients with Semantic Transparency judgments: Modified condition only.

FIGURE 2

Figure 2. Effect of semantic similarity (human ratings) on judged likelihood of property in Experiment 1.

Discussion

Although the specific properties being tested are unrelated to participants’ existing knowledge of the known compound, properties presented as true or somewhat true of the head are sensitive to the overall known semantic difference between the compound and the head. Thus, Experiment 1 shows clearly that, in addition to an expectation of some contrast due to the existence of the compound as a name that contrasts in some way with the head, the semantics of the specific compound word also contribute to the modification effects with known compound words when Almost All or Some members of the head noun have that property.

However, when the properties are presented as false of the head (i.e., Almost No members of the head noun have the property) the participants are insensitive to the degree of semantic difference between the compound and the head, indicating that the inverse modification effect with such features might provide us with a kind of baseline measure of the pure effect of expectation of contrast driven purely by the fact of the different name (i.e., the fact that there is an existing compound that contrasts with the head). We propose that this difference stems from meta-knowledge about the relationship between properties and concepts. True properties are assumed to be related to the semantics of the concept and false properties are not (see e.g., Murphy and Medin, 1985, for discussion of what features are relevant to a given concept). That is, our concepts, in general, tend to be organized around things that are true of them (e.g., we tend to think of birds, for example, more in terms of the fact that they usually fly, have wings, and have feathers, rather than in terms of the fact that they do not usually explode or earn PhD’s).

Thus, true properties are influenced by the actual (pre-existing) semantic similarity between the compound and head because this similarity is used as one source of information about the expected level of contrast when making property judgments in the Almost All and Some conditions. On the other hand, in the Almost No condition, semantic similarity is not seen as being relevant due to the meta-knowledge that false properties are not generally associated with the concepts. We return to this point in the section “General Discussion.”

Experiment 2

Experiment 1 showed very robust modification and inverse modification effects, with a strong asymmetry between the size of the two effects. In this experiment, we attempt to replicate these effects. Further, we examine whether property inferences for known compounds are affected by presenting the compounds with an open structure (e.g., black bird) or a closed structure (e.g., blackbird). This comparison will allow us to determine whether a compound (i.e., closed) structure encourages people to view the modified concept as being more distinct from the head noun concept than does a phrase-like structure, when the items are known compounds. On one hand, given that (in English), a closed compound structure is associated with more established compound words (see e.g., Kuperman and Bertram, 2013), if these effects are primarily driven by the notion that anything with a separate, existing, established compound name should have semantic differences from the head noun, one might expect that the open presentation would decrease the expectation of those semantic differences (and thus, an open presentation should attenuate the modification and inverse modification effects). On the other hand, simply presenting these well-known compound words with a space is, perhaps, unlikely to overcome the participants’ knowledge that these are, in fact, established compound words. Thus, spacing might not be influential, because the participants might be unlikely to believe that the inserted space indicates a novel phrase.