Our sample consisted of 810 undergraduate students (73 % women) of the National University of Mar del Plata, Argentina. Their ages ranged from 20 to 40 years old (mean = 24 years). All of them gave their informed consent to participate in this research. Every participant was a native Argentinean Spanish speaker.
The concepts were extracted from the data base built by Cycowicz, Friedman, Rothstein, and Snodgrass (1997). The following criteria were used for selection of the concepts: (a) these concepts and the corresponding pictures are frequently used in psychological experiments and tests about semantic memory, (b) these concepts correspond to those included in the expanded norms for experimental pictures in Argentinean Spanish-speaking population (Manoiloff, Artstein, Canavoso, Fernández, & Segui, 2010), and (c) most of them are included in McRae et al.’s norms (precisely, 229 concepts). Each concept corresponds to a single noun in Argentinean Spanish. In case of polysemic terms, a key was added to clarify the target meaning.
The norms include 400 concepts belonging to 22 semantic categories from the living and nonliving things domains. Next, we specify these categories and the number of exemplars that compose each of them (this number is indicated inside the parentheses next to each category name). Living things domain (129): animals (93), vegetables (12), fruits (15), and plants (9). Nonliving things domain (227): accessories (19), weapons (4), tools (33), constructions (17), house parts (10), clothing (17), utensils (29), furniture (14), vehicles (17), devices (13), objectsFootnote 1 (27), containers (16), and toys (11). Salient exceptions (44): food (6), musical instruments (14), body parts (19), and nature (5). Food, musical instruments, and body parts are exceptional cases because, according to the category-specific deficits literature they do not behave neither as nonliving nor living things (see Mahon & Caramazza, 2009). We also added the category nature as a salient exception because it includes concepts such as cloud and moon that are nonliving but, at the same time, are not manufactured by men. The concepts were also chosen to span a wide range of familiarity values, although a minimum value of familiarity was obviously required so participants could give useful information.
The norms were collected over a period of 3 years at the National University of Mar del Plata (Argentina). Concepts were distributed in groups of 15 in different spreadsheets in such a way that categories were homogeneously represented. Each participant listed features for only one set of concepts. Care was taken to avoid including similar concepts in the same spreadsheet. By “similar concepts” we mean concepts that belong to the same semantic subcategory. For example, cat and ant appear together in a spreadsheet, but ant and spider do not, because both of them are not just animals, but also arthropods (same phylum).
Participants were asked to list the features that describe the things to which the presented words referred. They were provided with 15 blank lines per concept to write down their corresponding features. They were instructed to list different types of features, such as those regarding internal parts and physical properties (their appearance, sound, smell, or touch). They were also encouraged to think about where, when, and what for they use the object at issue, and to consider the category to which it belongs. Two examples were provided, one for each domain. The instructions that were employed are presented in Appendix A. In every case, 30 participants listed features for each concept. Participants were not given time limit; they took approximately 20–30 min to complete the task.
An individual file was created for each spreadsheet, assigning one page to each concept. Because a very large number of participants contributed to the construction of the norms, their spontaneous answers obviously consisted in a quite large variety of ways of expressing the same features. For instance, to characterize the concept sun some participants wrote “yellow,” whereas others wrote “is yellow” or “its color is yellow.” The variety of spontaneous features is even wider in Spanish than, for instance, in English because in that language adjectives can be feminine or masculine and singular or plural, for they vary in concordance with the noun they qualify. To deal with variability, it was necessary to do an extensive work to ensure that features that conveyed the same meaning were recorded identically, both within each concept and among them. It was relevant as well to ensure that features that had dissimilar meanings were recorded with different labels. These recording procedures constituted a process named unification. This data treatment implies the adjustment of most of the features produced by participants, but it must be executed avoiding the alteration of the original content of those features. There are at least two important reasons to unify the features. First of all, the norms intend to capture the regularities in the production of semantic features; therefore, the wide variety of spontaneous formulations of those features must be reduced. Otherwise, the vast information provided by the norms would be useless, and its analysis would be impossible. In the second place, the unification of the features is mandatory in order to correctly compute many feature variables: on the one hand, variables such as production frequency (i.e., the number of participants who wrote a certain feature within a specific concept) would be erroneously calculated if features were not unified inside each concept. On the other hand, variables such as distinctiveness (that depends on the quantity of concepts in which a certain feature is listed) would not be correctly calculated if features were not unified among all concepts. Considering these reasons, we respected the recording criteria proposed by McRae and colleagues (2005); but we had to add new criteria to perform this process successfully. Next, we report the most important criteria we employedFootnote 2:
All features consisting in adjectives were written as singular and masculine independently of the number and gender of the corresponding concept.
Quantifiers (e.g., “generally” or “usually”) were eliminated, because the information provided by these words is expressed by the production frequency of the feature.
To identify the features that referred to a subtype of a concept, we used the expression <can be> (e.g., for the concept apple, <can be red> and <can be green> were used).Footnote 3
The features constituted by a quantifier adjective preceding a noun such as “has four legs,” were divided into two separate features: <has four legs> and <has legs>. This decision was taken because two bits of information are contained in features like these, and we intended to preserve both.
Disjunctive features (such as “is red or black” in the case of ant) were also divided (in this example, into <is red> and <is black>). However, if a feature conveyed a conjunction (such as “is black and yellow,” in the case of the concept bee), it was not divided.
In some cases, some words were added to the features. For example, an indefinite article (“a” or “an”) was added to the features that referred to superordinated categories (for instance, “animal” was transformed into <an animal>), and the expression “used for” was incorporated into the features that referred to a function (for example, the feature “to carry things” was transformed into <used for carrying things>).
Every feature that consisted in a verb was conjugated in the indicative mood of the present tense (e.g., “roar” was transformed into <roars> in the case of the concept lion).
The word “has” was added to every feature that made reference to the possession of a certain part or object; and that word replaced any synonym of it, such as “possesses,” and any other word that conveyed a similar meaning, such as “with” (e.g., in the case of the concept lion, the features “possesses a mane,” “with a mane,” and “mane” were all replaced by <has a mane>).
Measures and statistics
In the following paragraphs, the measures contained in the norms and the statistics calculated from them will be described.
The total number of features produced by participants was 21,630. However, it is important to remark that we only included those features that were produced by at least five participants, as McRae and colleagues (2005) did.Footnote 4 We did not take into account those features whose production frequency values were lower than 5, because we considered that they were not representative of the knowledge that the community has about the concepts at issue. Consequently, only 3,064 features were kept, 766 of which are distinctive. The mean of the features produced by each participant was 5.82 (SD 2.25; Max. 17–Min. 1).
Four files were elaboratedFootnote 5: a concept–feature file, a concept–concept matrix, a feature–feature matrix, and a significantly correlated features file. Next, we will describe each of these files.
File 1. concept–feature
The first and second columns of this file correspond to the concept name in Spanish and English; the two following columns correspond to the semantic-category name in both languages; then, the feature name in both languages is shown. The following column corresponds to the variable feature type, according to the coding scheme proposed by Wu and Barsalou (2009) in an updated version sent personally by the last author. These researchers considered five major categories: taxonomic categories (C), situation properties (S), entity properties (E), introspective properties (I), and miscellaneous (M). These categories are represented in the column named WB_Label with a capital letter as indicated above. In the present norms, 520 features were coded as taxonomic categories, 1,064 as situational properties, 1,383 as entity properties, and 97 as introspective properties. Wu and Barsalou (2009) also included a more detailed feature type classification, which is expressed in lowercase after the hyphen (to see the complete coding scheme, go to Appx. B). Some features conveyed information that was related to more than one classification category; in spite of this, in general terms we decided to allude to just one of these categories, following McRae and colleagues’ (2005) criterion. Nonetheless, there were some exceptions, such as the following: the features that alluded to quantities (such as <has two wings>) were codified as E-quant + the corresponding feature category (in this example, E-excomp), and those that included negations (like <cannot fly>) were codified as I-neg + the corresponding feature category (in this example, E-beh). After that, four columns were included with the amount of each type of feature within the concept at issue according to Wu and Barsalou’s (2009) major categories. (The miscellaneous category was excluded, because no single feature in our norms corresponds to it.)
The following columns correspond to production frequency (Prod_Freq)—that is, the number of participants who wrote that feature within the concept at issue; ranked production frequency (Rank_PF)—that is, the rank according to production frequency of the feature at issue with respect to the rest of the features within the concept; total production frequency (Sum_PF), which expresses the sum of the production frequencies of that feature across all concepts in which it appears; and CPF, which indicates the number of concepts in which that feature appears. Two measures related to CPF, which reflect whether or not the feature is shared among concepts, are also included: a qualitative binary variable (Disting) that indicates whether or not the feature is distinguishing by considering whether it appears only in one or two concepts, or in more than two; and Distinctiveness, a quantitative continuous measure that indicates the position of the feature in a range that goes from truly distinguishing features to highly shared ones (Devlin et al., 1998; Garrard, Lambon Ralph, Hodges, & Patterson, 2001). This last variable was calculated, as McRae and colleagues (2005) indicated, as the inverse of the number of concepts in which the feature appears in the norms (i.e., 1/CPF). In concert with these last researchers, this variable was calculated considering all the concepts included in the norms, instead of only taking into account the concepts that constitute a particular semantic category. Cue validity, which was calculated as the production frequency of the feature divided by the sum of the production frequencies of that feature for all the concepts in which it appears, was also added. It is important to note here that, unlike McRae and colleagues, we included taxonomic features in this last calculus. The reason for this is that this calculus does not include other features (as it is the case, for example, of intercorrelational strength), so it does not mix taxonomic features with other kind of features. Consequently, we provide researchers with information about taxonomic features that could be of interest without generating interference in the other features` measures.
We also included another feature property with (as the ones that were mentioned previously) demonstrated influence in different cognitive processes, named relevance (Rel) (Marques, Cappa, & Sartori, 2011; Sartori & Lombardi, 2004; Sartori, Lombardi, & Mattiuzzi, 2005; Sartori, Gnoato, Mariani, Prioni, & Lombardi, 2007). This variable is closely related to the distinguishing/nondistinguishing variable, distinctiveness and cue validity, because the four of them are measures of feature informativeness. In spite of this tight relationship, which explains the high correlation that exists among these variables, their importance differs in diverse kinds of tasks (as was shown in Sartori et al., 2005). This is the reason why we decided to include all of them in the Spanish norms. Relevance integrates two different components: a local one that may be interpreted as production frequency or dominance, which expresses the importance of a certain feature for a particular concept, and a global one that may be interpreted as distinctiveness, which expresses to what extent the feature at issue contributes to the meaning of the rest of the concepts. To calculate the values of this variable, we used the equation employed by Sartori et al. (2007):
$$ kij = lij\times gj = xij\times log\left(I/Ij\right) $$
In this equation, kij represents the relevance value of a feature j for a concept i, lij stands for the local component of relevance, and gj represents its global component. lij, which is equivalent to xij, is the production frequencyFootnote 6 of feature j over concept i. gj is equivalent to log (I/Ij), where I stands for the total number of concepts that constitute the data base at issue, and Ij represents the number of concepts of that data base in which feature j occurs.
Regarding these last four variables, it is essential to consider that as the norms only include a limited number of concepts, they cannot reflect with complete accuracy the actual distinctive quality of the features, because those features can also be defining of concepts that were not included in the norms. To solve this limitation, Devereux and colleagues (2014) decided to include related concepts for each of the categories that conformed their norms, in order to avoid having just one concept within certain semantic categories (e.g., they included at least two kinds of flowers). However, this proposal does not insure to solve the problem completely because the norms still constitute a limited sample of the universe of existing concepts. Moreover, the criteria they used to define what they considered to be a “related concept” is not clear. A different proposal regarding this problem was presented by Sartori et al. (2005), who focused on relevance in particular. These researchers highlighted that this variable may be greatly influenced by the total number of concepts contained in a certain normative database. To investigate this influence, they compared the relevance values of diverse features when computed in sets of different sizes (containing 50, 100, 150, and 254 concepts, respectively). The results they obtained indicated that the relevance values of the features that were calculated using the smaller samples predicted with high accuracy the relevance values that were obtained when the 254-concept normative database was used. Despite the fact that these findings reveal the stability of the relevance values from sample to sample, the original problem remains unresolved: the available databases do not exhaust the vastness of our conceptual knowledge.
Some physical characteristics of features’ and concepts’ names were also included, such as feature length including spaces between words (Feat_Lenght_Including_Spaces), and numbers of letters (Length_Letters), phonemes (Length_Phonemes), and syllables (Length_Syllables) of the concepts. Another reported concept variable is familiarity (Familiarity), which was extracted from the Argentinean psycholinguistic norms (Manoiloff et al., 2010).
Other variables refer to the concept’s feature composition. We included here: the number of features used to define the concept, including every produced feature (even those produced by just one person) (Total_Feat); the amount of features produced by at least five people, including taxonomic features (5_Feat_Tax), and excluding them (5_Feat_No_Tax). The reason to include these variants of the variable at issue is that they have been used as a measure of semantic richness but authors do not always agree in the criteria used to delimitate which features to consider (e.g., Pexman, Lupker, & Hino, 2002; Pexman et al., 2008; Yap, Pexman, Wellsby, Hargreaves, & Huff, 2012). Some authors exclude taxonomic features because they consider these features convey a different type of information than the rest of them. That is why we decided to give the interested reader the three options.
In this file, we also included four concept variables derived from the feature–feature matrix. The first is intercorrelational strength (Intercor_Strength_No_Tax) of the concept’s features, which is the strength with which a target feature (e.g., <is golden>) is correlated with the rest of the features of certain concept (e.g., bell). It is calculated by adding the features’ shared variances (i.e., r
2) with the rest of the features of the same concept. For this calculus, we considered a level of significance of p ≤ .05, which corresponds to a |r| > .164 (that is a 2.7 % of shared variance; Sheskin, 2007). The second is intercorrelational density (Density_No_Tax), which is the sum of r
2 for the concept’s significantly correlated features. This is a measure of the degree with which a concepts’ features are intercorrelated. Whereas intercorrelational strength is a feature variable, intercorrelational density is a concept variable. Both variables were calculated excluding taxonomic features. This decision was taken because we considered that other kinds of features have an asymmetrical relation with taxonomic features and are included directly in the definition of the taxonomic category itself. For example, to say that something is <an_animal> includes the idea that it can have hair (<has_hair>) or legs (<has_legs>).
The last two variables are the number of significantly correlated feature pairs in concepts excluding taxonomic features (Num_Correl_Pairs_No_Tax) and the percentage of significantly correlated feature pairs excluding taxonomic features as well (%_Correl_Pairs_No_Tax).
Other potentially relevant concept variables regarding their feature composition were included. To calculate some of these variables, taxonomic features were excluded: number of distinguishing features (Num_Disting_Feats_No_Tax), percentage of distinguishing features (Disting_Feats_%_No_Tax), mean distinctiveness (Mean_Distinct_No_Tax), and mean cue validity (Mean_CV_No_Tax).
Another concept variable, derived from the concept–concept matrix, was mean correlation (Mean_Corr). This variable is similar to the notion of normalized centrality degree (Freeman, 1979), according to which concepts are considered as vectors of features and concept proximities are a resultant of the number of shared features. These relations can then be represented as a two dimensional semantic network. Consequently, the normalized centrality degree would be the calculus that links the actual relations among concepts with their potential relations, expressed as a percentage.
File 2. concept–concept matrix
This matrix is composed of the 400 concepts and reflects the semantic distances between every pair of concepts according to their featural composition. Semantic distances were calculated by establishing the correlation between concepts using the geometric technique of comparing two vectors in the n-dimensional Euclidean space by the (least) angle between them. Parallelism (i.e., a cosine equal to 1 or –1) represents the maximum similarity, and orthogonality (a cosine equal to 0), the maximum difference. The computation of that angle (or actually its cosine) was made in the usual way, computing the ratio between the “component wise” inner product and the product of the respective Euclidean norms. It is worth mentioning that the idea of measuring the semantic distance through the construction of two vectors from a set of features that defines a certain concept was originally proposed by Kintsch (2001).
As a result of this calculation, we generated a mode-1 squared matrix (Borgatti & Everett, 1997) considering the semantic distances between pairs of concepts. To verify the validity of this matrix, a method to analyze emergent clusters (Johnson, 1967) was applied, using Ucinet 6 (Borgatti, Everett & Freeman, 2002). These clusters are depicted in Fig. 1. It is important to note that the aim of this figure is just to illustrate the clustering. The full information regarding the concepts’ correlations can be found in the concept–concept matrix.
As can be seen in Fig. 1, concepts that belong to the same semantic category are clustered together. In other words, the more features that overlap, the greater the proximity among concepts. For example, animals, means of transport, clothing, and musical instruments are clearly clustered, whereas concepts such as anchor or bird nest, which do not share many features with other concepts, do not show links with any of them. It is worth noting that this plot was built considering only those correlations ≥.4. The clusters’ sizes might slightly vary if this cutting point were modified.
File 3. feature–feature matrix
To construct this matrix, following McRae et al.’s (2005) criterion, only the features that were shared among at least three concepts were included. This allowed to avoid spurious correlations between features. Of the 1,315 total features produced, 186 were selected considering the criterion mentioned above. The final squared matrix in which the statistical co-occurrence was calculated contained 17,205 feature pairs, which were derived from the multiplication of 186 by 186 and subtracting 186 (because a squared matrix includes the combinations of each feature with itself), then divided by 2, due to the symmetric nature of the matrix.
Correlations were calculated using the CORREL function of Google Spreadsheets service. From the resulting squared matrix, the determination coefficient was calculated by squaring each correlation value and multiplying them per 100. As a result, we obtained a second squared matrix.
From this last matrix, we extracted some of the previously mentioned variables: features’ intercorrelational strength (Intercorr_Str_No_Tax), concepts’ density (Density_No_Tax), the number of significantly correlated feature pairs (Num_Corred_Pairs_No_Tax), and the percentage of significantly correlated feature pairs (%_Corred_Pair_No_Tax).
File 4. significantly correlated feature pairs
This file includes the r
2 value of each pair of features that was significantly correlated, excluding taxonomic features.