The influence of formal similarity on creativity in name-based word formation: the case of personal name blends in German

1 Introduction

The paper investigates creativity in name-based word formation in German using experimentally elicited personal name blends (henceforth PN blends), such as Benjalene from Benjamin and Helene as a testbed. As Lipka (2007) notes, proper names were often excluded from studies on linguistic creativity. They were considered part of general world knowledge, belonging to encyclopedias but not dictionaries. However, as they appear as constituents of linguistic units that are often regarded as creative (see Dancygier 2011), proper names play an important role in studying linguistic creativity.

This paper adopts a usage-based approach to blending (see Kemmer 2010; Kjellander 2018). According to this approach, blends are words cognitively linked to pre-existing/previously encountered blends that are co-activated when a particular blend is used (see Kemmer 2010: 71). Blends are understood and produced by cognitively routinized schemas that emerge due to perceived similarities among previously encountered blends (see Kemmer 2010: 78). Therefore, factors such as the frequency of language users’ exposure to PN blends and their knowledge about the domains in which PN blends typically occur should influence the production and interpretation of novel (previously not encountered) PN blends.

Furthermore, creativity is defined as a gradual and bipartite phenomenon. According to the common definition of creativity (so-called standard definition), it encompasses originality (novelty) and effectiveness (usefulness, appropriateness) as two primary components (Runco and Jaeger 2012). In linguistics, creativity means that speakers create and use novel linguistic units to achieve some communicative goal (e.g., to express an attitude or a sentiment towards someone or something, to be witty, etc.). As for PN blends, creativity is operationalized as a degree of recoverability/recognizability of source names^[1] underlying PN blends and the grade to which PN blends are interpreted as novel/original compared to attested PN blends and coordinative structures.

The paper investigates how the grade of formal similarity (measured using the normalized indel similarity, i.e., the minimum number of insertions and deletions required to change one sequence into the other (Deza and Deza 2009: 293)) between PN blends and their source names influences the ratings of the PN blend’s originality and the recoverability of its constituents. Furthermore, given the usage-based definition of blending mentioned above, we explore which usage-based factors (the degree to which the participants were exposed to concrete exemplars of PN and lexical blends and the participants’ domain-based knowledge about typical domains in which PN blends occur) significantly predict the creativity of PN blends exhibiting different similarity grades. Finally, the paper explores how usage-based factors (linguistic experience with blends and domain-based knowledge) influence the similarity values of PN blends.

The data used to test the hypotheses were not obtained from dictionaries, databases, or secondary literature. Instead, this work is based on PN blends elicited experimentally in a production task. The results of the production task were then used in a rating task conducted with the participants who did not take part in the production task. In this regard, the PN blends under investigation are novel compared to attested PN blends (e.g., Brangelina or Bennifer) because we assume the participants have not encountered them before. Following the usage-based approach, it is assumed that the production and interpretation of such blends are influenced by the language users’ experience with corpus-based/attested PN blends. Most of the pertinent studies on name blending rely on corpus data and primarily focus on English (e.g., Dow 2018). However, as Wulff and Gries (2019) note, “much of our previous knowledge of blends is based on observational data that, ultimately, may entail the risk of being opportunistic data samples”. Therefore, the present paper adopts an experimental approach and contributes to experimental research on blending in general (see for lexical blends Arndt-Lappe and Plag 2013; Beliaeva 2014; Connolly 2013; Villalva and Minussi 2022; Wulff and Gries 2019) and the creativity in name-based word-formation in particular.

The paper is structured as follows: After discussing the general properties of lexical blending in German and PN blends in particular (Sections 2 and 3), it focuses on creativity in German personal name blends (Section 4). Next, the measurement of formal similarity used in this paper is explained (Section 5). Experiments are presented in Section 6 and discussed in Section 7. The paper ends with a conclusion (Section 8).

2 Blending as a word-formation pattern with focus on German

This section discusses the properties of lexical blending in German that are relevant to the current study. It focuses on German lexical blends and presents their status and the classification of German lexical blends used in this paper. Further aspects mentioned in the literature (mainly on English lexical blends), such as the order of constituents, the stress, and the prosodic properties of blends and their constituents, will, therefore, not be discussed here.^[2]

(Lexical) blending^[3] can generally be defined as a non-concatenative process of creating new complex words by combining two (or more) words into one. It minimally involves some subtraction on one input word (e.g., in English acquihire < acquire + hire) and maximally allows for the extreme subtraction of all but one of its input words (e.g., English tactical + air + navigation > tacan, see Renner 2023). Similarly, Friedrich (2008: 219) defines lexical blends in German as the intentional merging of two (or more) elements (in German mostly nouns, see Fleischer and Barz 2012: 94; Friedrich 2008: 227) into a new unit in terms of both form and content, which is accompanied by the reduction of some parts of at least one element.

Although it is attested as a word-formation phenomenon in many languages of the world (see Stekauer et al. 2012: 132), blending has often been regarded as a relatively marginal aspect of word formation (Adams 1976: 148). In German word formation, (lexical) blends are also considered a marginal phenomenon (see Elsen 2008: 123). There are almost no extensive empirical studies (an exception is Friedrich 2008). For instance, Fleischer and Barz (2012: 94), a standard reference book of German word formation, notes that lexical blending is a word-formation type that usually remains occasional, with only a few such word formations becoming established lexical units. Beliaeva (2014: iii) summarizes the general status of blending in morphology as follows: There are “two possible approaches to (lexical) blending: either to deny blends a place in regular morphology (as suggested in Dressler (2000), for example) or to find grounds for including them into general morphological descriptions and theories (as was done, using different frameworks, in López Rúa (2004b), Gries (2012), Arndt-Lappe and Plag (2013) and other studies”. While one group of theories claims that blending is a part of extra-grammatical word formation (Dressler 2000; Mattiello 2013), other theories state that blending can be considered as part of regular word formation. From the perspective of extra-grammatical morphology (see Dressler 2000), blending is a word creation technique that goes beyond regular word formation patterns (composition, derivation, and conversion) and emerges by intentional deviation from the word formation rules to achieve different communicative needs (Ronneberger-Sibold 2006). In extra-grammatical morphology, blending is defined as extra-grammatical compounding because blends are composed of several (mainly two) lexemes, though formally violating the rules of regular compounding (see Ronneberger-Sibold 2015b: 2201).

By contrast, according to Kubozono (1990), blending is part of the grammar because it does not have any characteristics that are not found in the grammar (see Bat-El 1996: 284). Similarly, Kelly (1998: 588) claims that the variability in English blend formations, though extensive, is far from random. Still, he does not elaborate on the relationship between regularity and variability. As for the status of German blends, Fleischer and Barz (2012: 147) argue that blending is a phenomenon between regularity and irregularity. According to Beliaeva (2019a), factors that enhance the creative and attention-catching properties of blends may decrease the predictability of their form. Recent approaches to lexical blending in English follow usage-based theories of language and stress the role of schemas and linguistic experience with concrete exemplars for the production and processing of blends (Kemmer 2010; Kjellander 2018). However, to the author’s knowledge, the influence of linguistic experience with exemplars of attested PN blends on experimental PN blends has not been investigated.

Following the usage-based approach, we assume that the formation and interpretation of novel blends are not only based on recalling concrete instances of previous use of blends but also on their formal, semantic, or both formal and semantic modification. Two general mechanisms describe how speakers modify existing patterns of language use to form novel units, i.e., how existing experience with concrete exemplars is modified. As Fradin (2015: 393) (see also Körtvélyessy et al. 2022: ch. 3.1.2) states, the formation of (lexical) blends is based on “two contradictory requirements, namely a) the shortening of the source lexemes in order to make the blend resemble a single lexeme, and b) the preservation of as many segments (Bat-El 1996: 66–67) or relevant phonological properties from the source lexemes as possible (Ronneberger-Sibold 2006) in order to maximize the semantic transparency of the blend”. The first requirement is also known as the principle of least effort (Zipf 1949)/language economy (Martinet 1953), stating that all effort should be least (Zipf 1949: 543). This means that new blend words that are brief, not redundant, easy to pronounce, and pleasing to the ear are preferably formed (see Lefilliâtre 2019; Mattiello 2013: 27, 2022a). According to Mattiello (2022b), this principle is evident in “the formation of overlapping blends, such as sext (<sex + text), in which the redundancy given by phonemic overlap between the source words is avoided by haplology” (Mattiello 2022b: 44). A similar example from German is Amtsschimmelpilz < Amtsschimmel ‘red tape’ + Schimmelpilz ‘fungus’ (Ronneberger-Sibold 2012: 119). The second requirement is a universal principle, also known as the principle of maximization (preserve as much of the source words as optimal for their recognizability). According to this principle, both source words should be preserved to retain as much material from the source words as possible, e.g., Amtschimmelschimmelpilz. This example shows that such cases are against the principle of language economy and least effort. As discussed in Section 4, a creative PN blend is an optimal combination of both principles, i.e., the name constituents should be recognizable but only to the extent that does not violate the need for creating or being interpreted as a novel word.

Plag (2018: 123) describes the formation of English lexical blends as follows:

It is always the first part of the first element that is combined with the second part of the second element (see also Bauer 1983). This can be formulated as a rule, with A, B, C, and D,^[4] referring to the respective parts of the elements involved: Blending rule AB + CD = AD. In general, blends that do not correspond to the structure AD are in a clear minority.

Gries (2012) distinguishes between blends in a narrow sense (i.e., AD, ABD, and ACD constructs) and clipped compounds (i.e., AC structures). Since this distinction poses a problem of classifying structures such as the BD, ACB, and ACDB types, some scholars prefer the prototype-based approach to lexical blends (see Renner 2023: 4 for an overview). Although it has been stated that German lexical blends comprise the initial part of the first and the final part of the second constituent (Elsen 2008: 121), more extensive studies on patterns underlying the formation of German lexical blends have been described using classifications different from the typology mentioned above for English blends (see, e.g., Ronneberger-Sibold 2012: 120). In this regard, there is a mismatch between lexical and PN blends in German because German PN blends have been analyzed using Plag’s typology (Filatkina et al. 2019) and not based on existing classifications of German lexical blends (see, e.g., Ronneberger-Sibold 2012: 120). For this reason, classifications of German lexical blends will not be further discussed.

Regarding their semantic properties, lexical blends are often used as expressive means in various domains, including slang, popular media, political terms, professional vernacular, company names, names of musical bands, and other cultural groups (Beliaeva 2019a). While four semantic patterns underlying lexical blends are mentioned in the literature (see Fradin 2015: 402–404), Friedrich (2008: 221) mentions only two types for German lexical blends: the determinative type, where one blend constituent is further defined by the other in the semantic paraphrase of the blend (such as in ABIcalypse < Abi (Abitur) + apocalypse), and coordinative blends (e.g., Schafziege < Schaf ‘sheep’ + Ziege ‘goat’; an example from Fleischer and Barz 2012: 150), in which the two lexemes are on an equal footing, and their semantic import is equivalent. This distinction is also relevant for PN blending in German. In the current study, only coordinative PN blends are considered.

3 Personal name blending in German

PN blends are defined in this paper as routinized, cognitively entrenched patterns of experience with formal, semantic, and pragmatic aspects of blend usage (Kemmer 2010: 78). Kjellander (2018) states that the formation of lexical blends is governed by a set of cognitive constraints that may hinder the process of lexical blending. For the formation of experimentally elicited PN blends, the schema transfer effect (Kjellander 2018) is significant. The schema transfer means that a schema, understood as a cognitive representation consisting of perceived similarities across many instances of PN blend’s usage, can fuel the formation of similar forms through the process of analogy (Kemmer 2010; Kjellander 2018). Therefore, the formation and interpretation of experimental PN blends are influenced by individual differences in the speaker’s frequency of exposure to attested PN blends, such as Bennifer or Brangelina, i.e., with language users’ linguistic experience with them. The exposure/experience is operationalized in the current study as the participants’ familiarity with PN blends and lexical blends (see Section 6).

As for their status in German, PN blends have been sporadically mentioned in previous studies. For instance, Koch (2010: 110) claims that lexical blending with names as constituents is very popular in German media and provides several examples from different domains, such as Klosolski formed from the names of football players Miroslav Klose and Lukas Podolski. He also mentions other types of PN blending, such as those formed by combining a name and a lexical unit (e.g., Jensationel from the first name Jens and the adjective sensationell ‘sensational’. Similarly, Schmid (2003: 271) claims that the amount of foreign and name-based elements in German blends is high. However, there is no systematic analysis of their structural, semantic, and pragmatic properties. Friedrich (2008: 233) mentions two main types, namely the group-specific nicknames (e.g., Cornsula as a witty nickname of sisters Cornelia and Ursula) and brand names. Blended brand names in German have been studied by Ronneberger-Sibold (see for an overview Ronneberger-Sibold 2015a). A similar classification has been proposed for English: “today, blending is popular for naming. Newly developed software like Linux (Linus/Unix) and celebrity couples like Brad Pitt and Angelina Jolie (Brangelina) are given blended names to signify that two entities are merged into a single unit” (DiGirolamo 2012: 231).

Extensive empirical studies are still scarce. Belosevic (2021) investigates so-called complex PN blends, such as Angela Merkohl < Angela Merkel and Helmuth Kohl from politics based on social media data and internet texts. Name-based blends comprise different name classes, such as personal names, place names, or brand names. This paper considers only the subtype of PN blends with personal names (first names in German) as constituents. Filatkina et al. (2019) provide a corpus-based comparison of English and German PN blends regarding their formal and semantic properties. Filatkina (2019) compares German PN blends with other types of name-based word formation in German based on the attestations from the German Reference Corpus.

PN blends have the status of personal names and can be considered as a subtype of nicknames. In addition, this paper assumes that experimentally elicited PN blends and corpus-based PN blends, such as Brangelina or Merkozy, form two subtypes of personal name blends in that experimental PN blends are novel (in terms of never previously being encountered) compared to attested/corpus-based blends that language users were exposed to different degrees.

In addition, although blending based on proper names may be regarded as less productive (in terms of type frequency) than lexical blends, PN blends have been attested in several languages (for French Sablayrolles 2015: 192; and for the blended nicknames used by Japanese junior high school students Barešová et al. 2020). Filatkina et al. (2019) compare formal and semantic properties of German and English PN blends based on studies of English lexical blends and confirm the patterned character of German and English PN blends previously postulated only for English lexical blends, namely that “particular patterns have been observed in terms of the selection of the words that are blended and the formal structure and the semantics of blends” (Beliaeva 2019a). Furthermore, Filatkina et al. (2019) provide a classification of PN blends in coordinative and determinative subtypes (in analogy to lexical blends, see Section 2). Determinative PN blends have one name constituent as a referent, such as Messidona < (Lionel) Messi and (Diego) Maradona, referring to the football player (Lionel) Messi by comparing him with Diego Maradona (i.e., Messi is like Maradona). The coordinative subtype focused in the present study has both name constituents as its referents, e.g., Merkozy < (Angela) Merkel and (Nicolas) Sarkozy. Besides an additive interpretation (Merkel and Sarkozy), there is corpus-based evidence of more abstract interpretations, such as ‘political cooperation’ or ‘rivalry’ (Belosevic 2021; Filatkina et al. 2019). The current study also assumes that experimental PN blends are not only formally but also semantically novel compared to coordinative structures (see Section 4). This semantic aspect of PN blends’ creativity is not the focus of the paper.

Finally, the study accounts for domain-specific properties of German and English PN blends in that they occur in different domains, mainly politics, celebrities, and sports. The paper considers this aspect an integral part of language users’ experience with blending (following a usage-based approach).

4 Creativity of personal name blends

Lexical blending is regarded as the most creative of all word-formation processes (Lalić-Krstin et al. 2024: 101; Renner 2015). Therefore, such blends are usually defined as creative word formation units (see, e.g., Gries 2006), as instances of wordplay (Renner 2015) or as creative neologisms (Lehrer 1996). Gries (2012: 145) defines lexical blends as not rule-governed, less productive than other word-formation processes, and more creative word-formation units than other word-formation phenomena. Villalva and Minussi (2022) understand blending as “an exercise of linguistic creativity, bound to no explicit constraints”. Fradin (2015: 393) states that

blends give the speaker something that compounds do not, namely the opportunity to show her capacity to play with language, which is a socially praised ability, creating an unconventional, witty semantic association between two (or n) lexical meanings packed in one word. The coinage of blends is part of the epilinguistic competence of native speakers, which also manifests itself through puns, spoonerisms, witticisms, and other language games.

However, it is still unclear which aspects of creativity are included in the definition of PN blends as creative word-formation units. Furthermore, the differences between types of PN blends (e.g., corpus-based and experimentally elicited PN blends) and groups of language users (e.g., those who were more or less frequently exposed to attested/corpus-based PN blends) have not been considered as factors that may influence the creativity of PN blends.

Since the term creativity has been used in various linguistic domains and disciplines, such as politics, psychology, or education (see Tin 2022: ch. 1 for an overview), this section defines creativity by considering the properties of PN blends. The paper takes one of the most prominent definitions of creativity, the so-called standard definition (Runco and Jaeger 2012), as a starting point. Initially developed for psychological creativity research, this definition is adopted for a usage-based approach to PN blends in German. Despite criticism, it “still appears to be the most balanced in terms of establishing the requirements for the achievement of creativity” (Corazza 2016: 260). Finally, the current study considers creativity as the ability of all language speakers and not as an ability of non-conventional people in their respective fields (see also Körtvélyessy et al. 2021).

According to Runco and Jaeger (2012), “for something to be creative, two elements are required: originality, or what some people might refer to as novelty or uniqueness, and effectiveness, which in creativity may go by another name, such as usefulness, fit, or appropriateness” (Runco and Jaeger 2012: 92; see also Abraham 2025 for main drawbacks of this approach to creativity). Similarly, Sternberg and Lubart (1998) define creativity as “the ability to produce work that is both novel (i.e., original, unexpected) and appropriate (i.e., useful, adaptive concerning task constraints)”. It is “an inherent human faculty, shared by all the speakers of a language and from the very earliest age” (Munat and Sorlin 2010: 155). A recent review of the literature on creativity (Puryear and Lamb 2020) yielded that many definitions include novelty and appropriateness as core components of creativity. As Hoffmann (2024: 142) puts it:

Some ideas, like using a thimble to drink water, might be fairly novel/original but will probably be considered inappropriate by most people. On the other hand, using a regular glass for drinking will seem perfectly acceptable, but is not original at all (Hoffmann, “Constructionist” 260; “Cognitive”). It is only when something is both original and appropriate that it is seen as creative (and consequently appreciated by the listener/reader; Giora, On; Veale) – e.g., when you are marooned on an island, and you use coconut shell halves to drink water from a well.

Although the bipartite definition can be regarded as the most common definition of creativity nowadays (Weiss and Wilhelm 2020: 3), “the question about which objective criteria can be applied to measure creativity in language arises” (Weiss and Wilhelm 2020: 4). The bipartite definition implies that a PN blend cannot be creative without being original and effective at the same time.

In this paper, the creativity of PN blends is investigated using two scales: one for the grade of originality/novelty of PN blends and one for the degree of their effectiveness (measured as the degree of recognizability/recoverability of source names). Notably, the paper accounts for the creativity of PN blends from the perspective of how they are interpreted by addresses and not from the production perspective of blend coiners. However, experimentally elicited PN blends are used as items in the rating task.

Creative PN blends are units with a degree of recoverability that allows for a reconstruction of source names but in such a way that it also enables a blend to be interpreted as a novel form compared to juxtaposed personal names (e.g., Brad and Angelina). The interpretation of PN blends as more or less creative is based on the language user’s previous experience/exposure to attested (corpus-based) lexical and PN blends. We also assume that the recoverability is based on the principle of maximization, while novelty follows the principle of least effort/language economy (see Section 2). In the following, the two aspects of PN blend’s creativity will be defined more precisely.

The operationalization of the effectiveness is related to the definition of creativity as something novel that functions as intended (see Silvia 2018). The intended function of PN blends encompasses the speaker’s intentions and how the addressees interpret this communicative intention. It can be assumed that the first general intention of coining PN blends is that addressees can reconstruct the name constituents underlying the blend. Recoverability of blend constituents means that “coiners of subtractive word formations must ensure their creation’s component parts can be recognized again. However, the secure way of doing this – simply including (nearly) the whole word – is not available since blends and complex clippings would then not exhibit the wit for which they are frequently put to use (esp. in advertising) because (i) no cunning wordplay would be involved and (ii) the blend would not be similar to both its source words anymore” (Gries 2006: 538). The second intention of the blend coiner is to verbalize so-called expressive meanings, see Filatkina et al. 2019). Specifically, the blend coiner has an additional, so-called perlocutionary intent (see Lehrer 2003: 380), namely to express an attitude towards the relationship between name bearers whose names are involved in the blend. Therefore, the present paper assumes that the effectiveness of PN blends encompasses recognizing the source names and inferring their expressive meaning (after the source names have been successfully identified).

Although Runco (2023) states that an idea is only creative when it successfully solves the problem and not only when it is original, previous studies (Pichot et al. 2022) show that everyday judgments of creativity may weight the novelty criterion more highly than appropriateness. Originality is closely related to the properties novel, unique, new, or unusual. It can be analyzed from the perspective of an individual language user or the perspective of changes in the language system. In this paper, we focus on individual language users and not on originality in the speech community. Note that both aspects cannot be separated and should be investigated in correlation. Originality includes the formal and/or semantic novelty of a linguistic unit compared to existing linguistic units that can be used instead of this unit in similar contexts. This definition of originality follows a usage-based approach because novel linguistic units are formed and interpreted based on the language users’ frequency of exposure to or experience with concrete instances of existing linguistic units.

Linguistic forms with a high originality grade are usually hard to link to linguistic experience (i.e., to some familiar form and/or meaning already cognitively entrenched in the language user’s mind). According to the bipartite definition of creativity, such units cannot be regarded as creative. For example, a “pill of pepper” is more original than “a piece of paper”, but the salient meaning conceptualized by the form “a piece of paper” is hard to infer from the form “a pill of pepper”. Similarly, too familiar units are not creative if conceptual novelty is lacking, such as in a single piece of paper versus a piece of paper where both forms refer to a qualitatively similar response (see Giora et al. 2004: 117).^[5] What is regarded as creative can be explained by the optimal innovation hypothesis (see Giora et al. 2004). Optimally innovative units (see Giora et al. 2004: 117) are linguistic innovations, such as “a peace of paper” that can be traced back to some salient meaning (i.e., “a peace of paper” < “a piece of paper”) to a particular extent while promoting new meanings at the same time (i.e., “a worthless peace agreement” for “a peace of paper”).

As for PN blends, the similarities and differences between an entrenched unit (e.g., Bastian and Renate) and a novel unit (e.g., Bastinate) must be assessable (see Giora et al. 2004: 116) for a novel unit to be creative (given an additional requirement that Bastinate conceptually goes beyond Bastian and Renate and verbalizes more abstract/non-additional meanings). For this reason, a PN blend such as Bre < Bastian and Renate is less creative than Bastinate because the name constituents are more complex to reconstruct (although the form is more novel than Bastinate).

The originality of experimental PN blends is defined in this paper from the recipients’ perspective (and not from the perspective of the blend coiner). It comprises three levels, two of which are focused on in the current study. First, on a general level, experimental PN blends are original compared to attested (corpus-based) PN blends. According to the usage-based approach, unattested/experimental PN blends (e.g., Renastian) are novel compared to attested/corpus-based blends (e.g., Brangelina) as the former are assumed to be never encountered. This aspect will be tested in the post-questionnaire of the rating task (see Section 6.2).

Second, experimental PN blends are original relative to the originality of their source names. That is, we assume that the participants would interpret a blend (e.g., Renastian) as more original than the sum of its name constituents, i.e., juxtaposed names, such as Renate and Bastian. The rating task was designed so that the participants were informed that the PN blend was formed based on a combination of two names. At the same time, the participants had to write down the source names that they think underlie a PN blend (see the task description in Section 6.2).

Finally, some blend forms are interpreted as more original than others (depending on different (extra)linguistic factors, such as the context or the domain of use). For example, the form Bre from Bastian and Renate may be interpreted as more original than Bastinate (however, it is probably interpreted as less effective since the source names are less recognizable than in Bastinate). This aspect of originality was not tested in the current study. All levels assume a particular degree of originality assigned to a PN blend.

Besides this formal perspective, the originality of PN blends can be defined in terms of their semantics. However, this aspect will not be the focus of the present study. It encompasses cases in which a PN blend is not interpreted in an additional way, e.g., Renastian referring to Renate and Bastian. By contrast, semantically original PN blends are assigned lexical meanings, such as when language users interpret Renastian as an ‘ideal relationship’ or ‘ideal couple.’

5 Measuring the similarity between PN blends and their constituents

The similarity between PN blends and their source names is measured by normalized indel similarities. This section defines indel similarity and discusses how it can be used to measure PN blends’ creativity.

Given that “a certain similarity with the source words is a requirement for a successful blend” (Hamans 2010: 465), formal similarities between PN blends and their source names are an important factor in assessing the interpretation of creative PN blends. Different similarity measures have been applied to lexical blends (Gries 2004b; Wulff and Gries 2019). However, to the author’s knowledge, the formal similarity between a blend and its constituents has never been tested for its influence on the creativity of PN blends. Although Gries (2004b: 656) proposed SI (similarity index) as an “objective quantitative measure of the similarity between source words and blends calculated by assessing the proportion of graphemes (or phonemes and articulatory features) each word contributes to the blend together with the proportion these graphemes/phonemes make up in the blend”,^[6] later he argues that the Levenshtein string edit distance is a more sophisticated measure than other measures, such as phonemic or graphemic similarity discussed in different studies (see Beliaeva 2019b for an overview; Kelly 1998 for phonological similarity in lexical blends; and Gries 2017 for applying the Dice coefficient as an alternative measure). Levenshtein string edit distance is based on several operations (insertions, deletions, and substitutions) needed to change one word into another (see Gries 2012). Importantly, it reflects dissimilarities between two words, i.e., the higher the distance, the more dissimilar the two words are (see Gries 2017: 35). Wulff and Gries (2019) calculated the average Levenshtein string edit distance (ASED) value by taking the average of the SED between source word one and the blend and the SED between source word two and the blend.

In this paper, the normalized indel similarity in the range [0, 1] between each source name and a PN blend is used as a measure because the average values of the Levenshtein string edit distance were similar for most items. Normalized indel similarity, therefore, provided a measure that could be operationalized for our research questions. The normalized indel similarity measure can be defined as a variation of the Levenshtein distance that only allows insertions and deletions (while the Levenshtein distance allows for substitutions and insertions and deletions, see Deza and Deza 2009). It was calculated with the Levenshtein-package in Python using Levenshtein module (more precisely Levenshtein.ratio)^[7] and following the procedure similar to those proposed by Wulff and Gries (2019) for lexical blends, i.e., by taking the average of the normalized indel similarity between the first source name and the PN blend and the normalized indel similarity between the second source name and the PN blend. The closer the ratio value to 1, the higher the similarity, as it took fewer insertions and deletions to get from one string to the other. For example, indel similarity between the blend Dencon (Denise and Constantin) and the name constituent Denise is 0.5 and between Constantin and Dencon 0.25. This means that Denise is more similar to the blend because it took fewer insertions and deletions from the name to the blend compared to Constantin. In the second step, we sum up the two values (0.25 + 0.5 = 0.75) and divide them by two to obtain the mean value (here: 0.375). Since the value of 0 means that the blend and the source names are completely dissimilar, the value of 0.375 indicates a relatively low similarity value that can be assigned to the blend Dencon. This mean value of the normalized indel similarity was then used in the rating task (see Section 6.2).

The formal similarities between both source names underlying the PN blend were not considered because it is assumed that in naturally occurring contexts, the selection of source names underlying PN blends is motivated by extralinguistic factors, such as the goal to take a stance toward the relationship between the name bearers rather than by formal similarity between their names. Gries (2012, 2017: ch. 2) mentions that it is necessary to compare the similarity scores of intentional blends to other similar formations to account for differences and to compare the properties of intentional blends against a random distribution. Since the present paper did not focus on similarity as a property of PN blends but rather used similarity as a starting point for analyzing the blend’s creativity, the values of normalized indel similarities between PN blends and their name constituents were not compared with other datasets and similar word-formation units.

As Gries (2004a, 2012) states, the intentional coinage of blends is based on the interplay between different types of similarity and recognizability. This means that the recognizability of blend constituents and the similarity between them and a PN blend are related, but they can also conflict with each other (Gries 2017: 30–39). Gries (2017: 30) distinguishes between how similar the source words of a blend are to each other and how similar they are compared to the blend structure. The similarity between the blend and its source words (measured by the number of syllables) increases the recognizability of the blend’s constituents, or it contributes to the conflict between them (such as in cases where both source words are clearly recognizable but the blend is not similar to either of the source words anymore and is not witty. Therefore, to be interpreted as creative, a PN blend must be balanced in terms of the recognizability of its constituents and the similarity between the blend and its source words. The low similarity (in terms of the syllable number) between a blend and its constituents makes a blend less witty/novel/original (Gries 2017). Since many different ways exist for the similarity between words to be operationalized (Gries 2017: 34; see similarly Gries 2012: 147), this paper does not use the syllable number to measure the similarity between a PN blend and its source names. Instead, the normalized indel similarity is used as a more precise measure.

The relationship between the normalized indel similarity between a PN blend and its source names on the one hand and the originality of a PN blend on the other can be described as follows: If higher indel similarity means that more material from both source names is retained in the blend (because it took fewer insertions and deletions of a source name to coin the blend), then PN blends with high indel similarity values will be interpreted as less original compared to those with low indel similarity values. Therefore, it can be hypothesized that the more the source names are retained in the blend (i.e., the higher indel similarity), the less original it is. Given that, in contrast to the relationship between the recoverability of blend constituents and similarity, the correlation between the similarity and the originality, to the author’s knowledge, has not been investigated in previous research on (lexical) blending, this hypothesis will be tested based on the results of the rating task.

6 Experiments

The paper investigates the following research questions using an experimental approach:

1. How does formal similarity between a blend and its name constituents influence its creativity?

2. Which factors significantly predict the creativity of PN blends bearing different similarity grades?

3. How do usage-based factors (linguistic experience with blends and domain-based knowledge) influence the indel similarity value of PN blends?

As for the first research question, a rating task was conducted to test the three hypotheses listed below. The classification of PN blends into two groups (high similarity vs. low similarity between the blend and the source names) is based on the results of the production task (see Table 2):

1. H1: PN blends with high normalized indel similarity values (e.g., Constise = 0.61) have higher ratings of recoverability than of originality and vice versa: PN blends with low normalized indel similarity values (e.g., Dencon = 0.37) are rated as more original than recoverable.

2. H2: PN blends with low normalized indel similarity values are rated as more original than the opposite group of PN blends and vice versa: PN blends with high normalized indel similarity values are easier to reconstruct than PN blends with low similarity ratings.

3. H3: The higher the number of correctly predicted source names, the higher the rating values of recoverability. The hypothesis is based on Lehrer’s (1996) assumption for lexical blends that “subjects who correctly identify the target words will rate that item more favorably than those who do not. Moreover, there will be a ranking, with the highest rating given by subjects who correctly identify both words, followed by those who identify one word, with the lowest rating given by those who can not identify either word” (Lehrer 1996: 367).

Research question (2) is investigated using inferential statistics based on the rating task. As for research question (3), the statistical analysis of the data from the production task is conducted. In the following sections, the experimental design and the results of the production study and the rating task are presented and discussed.

6.1 Production task

Before the rating task, a production experiment was conducted to obtain the stimuli for the rating task. In addition, the produced PN blends were tested for the factors that significantly predict the similarity grade between a blend and its source words (see research question 3). In this regard, two under-researched factors relevant to usage-based approaches to blending, namely the participants’ knowledge about the typical domains in which (attested) PN blends occur and their linguistic experience with concrete instances of lexical and PN blends, were tested.

The production study was conducted as a discourse completion task. The participants were asked to propose PN blends from the name pairs presented in a minimal context in four conditions (love couples, political rivals, neighbors, sports duo). An example of the task for the domain ‘love couples’ is provided below (translated from German):

The names of the couple are supposed to be on the wedding cake, but there isn’t enough space. Combine and shorten the two names given below to create a new name for the couple.

The conditions were selected based on the corpus study of PN blends in German (see Filatkina et al. 2019) yielding these four domains as typical of German PN blends. The task was constructed in such a way that the participants should produce only coordinative PN blends (see Sections 2 and 3).

6.1.1 Materials

The items (20 in total, five per domain) were constructed based on the list of the 300 most frequent German first names.^[8] Female and male first names were combined. The names were controlled for formal properties (syllabic length, ending vowel) based on the gender index for German first names (see Nübling 2017). Note that it was not possible to meet all criteria since the list mentioned above does not contain enough names that account for all aspects of the gender index. This especially holds for the syllable number of female names (see Denise: two syllables instead of three). However, names met most of the criteria. Both female and male names comprised three syllables but had a different stress position: female names had the non-initial stress and received the value +3. Male names had the initial stress position (value +1). Female names always ended in -e and had the value +3 for ending vowels (according to the gender index). Male names ended with a sonorant −n and had the value +1 for ending vowels. For female names, the distribution of vowels and consonants was equal (value +1), while male names had more consonants than vowels (value −2). The following name pairs were used in the production task: Sabine and Jonatan, Helene and Benjamin, Renate and Bastian, Simone and Florian, Denise and Constantin. The order of female and male names was controlled, i.e., the participants were always presented with a combination of male + female names. However, given that the tendency to maximize the similarity between the blend and each of its source words is one of the factors that influence the ordering of the source words (see Beliaeva 2019b), the participants were allowed to change the order of names if they wanted to do so. The order of name pairs was randomized.

6.1.2 Procedure and participants

All participants were exposed to all four conditions (within-subjects design). In each condition, the participants were presented with five name pairs (two first names), i.e., each participant was exposed to 20 items in total. The name pairs were identical in each condition. The participants were not explicitly required to be creative. The term (name) blends was not mentioned explicitly. Since creativity is constrained by background knowledge according to the Creative Cognition Approach,^[9] it can be assumed that the participants were constrained by their background knowledge of first names in German and the name structure to a different degree. Therefore, the produced PN blends should exhibit different degrees of normalized indel similarity.

The production experiment was conducted as a web-based experiment using the questionnaire tool SoSci Survey.^[10] The experiment consisted of four sections: demographics, instructions, test items, and post-questionnaire. Sociodemographic information (age, gender, and mother tongue) was collected in the pre-questionnaire. Furthermore, the participants were presented with test examples. In the post-questionnaire, the frequency of exposure of the participants to lexical blends and PN blends/their linguistic experience with lexical and name blending was assessed by asking about their familiarity with concrete exemplars of PN blends and lexical blends. Specifically, the participants were required to answer the following two questions on a three-point scale (in German): yes, to some extent, no: “Are you familiar with words such as Brangelina, Bennifer, or Merkozy?”, and “Are you familiar with words such as motel or brunch?”

One hundred students took part in the production study. Seven participants were excluded because they provided incomplete answers. Therefore, the results provided by 93 participants (mean: 24.11 years, median: 23 years, SD: 3.430298, min.: 18 years, max.: 34 years, male: 18, female: 74, other: 1) were analyzed. Twelve participants were bilingual, and 80 were German native speakers. One participant did not respond to the question regarding the native language. Participants received course credit in exchange for participation. Seventy-one participants (77 %) self-reported being familiar with blends (answers in the post-questionnaire for the experience with lexical blends yes or to some extent and PN blends yes or to some extent).

6.1.3 Results

In total, 1840 tokens (618 hapaxes) were produced. Items comprising phrases with und (“and”) were excluded from the analysis (for example Sabi und Jonas for the name pair Sabine and Jonatan) so that 1,577 tokens remained (433 hapaxes). The frequency of produced PN blends for each name pair was calculated in R (R Core Team 2023).

The normalized indel similarity was used to measure the formal similarity between a PN blend and its source names. The measurement follows the procedure described in Section 5. The minimum similarity value (0.17) was assigned to the item Seco (from Denise and Constantin). The maximum value was 0.71 for Renastian (Renate and Bastian) and Florimone (Florian and Simone). These values indicate that for Seco more insertions and deletions of source words were necessary to coin a PN blend than, for instance, Conise, where both PN blends are formed from the name pair Constantin and Denise. The mean value of indel similarities is 0.57, the median is 0.59, and SD is 0.11. In addition, the five most frequently produced PN blends have very high similarity values (Jonabine [Jonathan and Sabine]: 0.68; Conise [Constantin and Denise]: 0.58; Simorian [Simone and Florian]: 0.68; Renastian [Renate and Bastian]: 0.71; and Bastinate [Bastian and Renate]: 0.64).

A linear mixed-effect model with normalized indel similarity as a dependent variable and the following predictor variables: reported frequency of exposure to lexical blends, reported frequency of exposure to PN blends, and the domain in which the production of a PN blend is embedded (love couples, politics, neighbors, sport), was constructed to account for the research question 3 (How do usage-based factors [linguistic experience with blends and domain-based knowledge] influence the similarity values of PN blends?). The factors that significantly predict the similarity grade between the blend and the source names were tested.

The starting point of the model selection was a maximal generalized linear mixed model that included all predictor variables. The predictor variables were then reduced stepwise and assessed by model comparison based on log-likelihood values. Interactions between the significant predictor variables were not included because they did not significantly improve the model. The final model has the following formula:

indel_similarity ∼ domain + exposure_lexical_blends + (1|ID).

The final model comprised normalized indel similarity as a dependent variable and domain and exposure to lexical blends as predictor variables. In addition, random intercepts for participants were included in the model. The results of an ANOVA showed no significant effect of the exposure to PN blends (χ (1) = 1.5252, p > 0.05). The coefficients of the fixed effects,^[11] standard errors (SD) and t-values are given in Table 1. The reference level for the dependent variable was 0.17, which is the lowest value for normalized indel similarity.

Table 1:

Production task, fixed effects.

	Estimate	SE	t value
Intercept	0.579176	0.008226	70.407
domain_neighbors	−0.018032	0.005993	−3.009
domain_politics	−0.022064	0.005953	−3.707
domain_sport	−0.003547	0.005628	−0.630
noexposure_lexicalblends	−0.154765	0.048361	−3.200
exposure_lexicalblends_tosomeextent	−0.020140	0.032043	−0.629
Marginal R 2	0.05975748
Conditional R 2	0.4213533
Nr. of observations	1,577

The effects of predictor variables (domain and exposure to lexical blends) are negative for the lowest value of normalized indel similarity (dependent variable), which means that the higher the normalized indel similarity, the lower the chance that the blend comes from the domain neighbors, politics, or sports. Furthermore, the higher the normalized indel similarity, the lower the probability that the participant reported not having experience with lexical blends. The effects of the domain and the exposure to lexical blends on the production of PN blends with different similarity grades are visualized in Figure 1.

Figure 1:

Effects of predictor variables production task.

The figure shows that if the production of a PN blend is embedded in the context of love couples, the produced PN blend bears a higher normalized indel similarity between names and PN blends compared to the other three domains. Furthermore, language users who reported higher exposure to lexical blends are more likely to produce PN blends with higher normalized indel similarity values than the participants who reported lower exposure.

After obtaining the items, a rating task was conducted to test the hypotheses mentioned above. To the author’s knowledge, this is the first study dealing with the relationship between formal similarity and creativity in German PN blending, and it has a rather exploratory character.

6.2 Rating task

Based on PN blends elicited in the production task, a rating experiment was carried out to test whether and to what extent experimentally elicited PN blends with different degrees of indel similarities are interpreted as creative, i.e., to what extent are their constituents recoverable/recognizable and to what extent a PN blend as a whole is interpreted as original (see research question 2). In this regard, the study is similar to the experimental study by Lehrer (1996) on the factors that go into the rating of lexical blends. However, Lehrer investigated lexical blends and did not test explicitly for creativity but rather for how good or bad the lexical blends were. However, the hypothesis proposed by Lehrer (1996) regarding the correlation between the rating values and the number of successfully identified source names will also be tested in the present paper (see Hypothesis 3).

6.2.1 Materials

For each name pair, the PN blend from the production task with the highest and lowest normalized indel similarity was selected for the rating task (10 PN blends in total) regardless of the distinction between blends in a narrow sense and clipped compounds (see Section 2 and Plag 2018: 121).^[12] In terms of their semantics, all PN blends coined in the production experiment have coordinative meanings, according to the semantic classification of PN blends (see Section 3).

Table 2 shows the final list of stimuli for the rating task. Note that high and low similarities refer to normalized indel similarities between source names and PN blends.

Table 2:

Stimuli for the rating task.

name pair	High indel similarity	Low indel similarity
Sabine und Jonathan	Jonabine (0.68)	Neon (0.26)
Denise und Constantin	Denistantin (0.67)	Seco (0.17)
Renate und Bastian	Renastian (0.71)	Bre (0.21)
Benjamin und Helene	Helenjamin (0.69)	Majle (0.25)
Simone und Florian	Florimone (0.71)	Moflo (0.25)

As the table shows, items with low similarity values are shorter than PN blends with high similarity values. While the group of PN blends with low similarity values shows a tendency towards the economy of expression, the group with high similarity values shows a tendency towards semantic transparency (see Körtvélyessy et al. 2022: ch. 4 and Section 2). In the rating task, similarity was coded as a binary variable: “0” for items with high similarity values and “1” for items with low similarity values. The Supplementary Materials provide a complete list of normalized indel similarity values for all items from the production task.

6.2.2 Procedure and participants

The participants were asked to rate the creativity (operationalized in terms of originality of a PN blend and recoverability of its source names) of the items from Table 2 on two scales with values between 1 (clearly not recognizable/not original) and 5 (clearly original/clearly recognizable) because both originality and recoverability are a matter of degree. Specifically, the question was as follows (in German):

[NAMEBLEND] is a combination of two first names.

I think [NAMEBLEND] is:

absolutely not original/rather not original/neutral/rather original/absolutely original;

I can recognize the names underlying the [NAMEBLEND]:

absolutely not/rather not/neutral-I do not know/rather yes/absolutely yes;

If you can identify the name constituents underlying the [NAMEBLEND], please write down the names that you think the blend is made of.

The items were presented in contexts identical to the production task (four domains: sports duo, politicians, love couple, and neighbors). Additionally, a condition without context was included. The order of the questions was as follows: First, the ratings of recoverability should be provided. In this question, the participants were not informed about the source names of a PN blend because, in the second question, they were required to write down source names. In the final question, the originality of PN blends should be rated. In this question, both source names were presented to the participants.

In the post-questionnaire, the frequency of exposure to PN blends and lexical blends was tested using the same question as in the production task (two single-choice questions with a three-point scale: yes, to some extent, no, see Section 6.1.2), as well as the attention-check single-choice question asking about the sincerity of the provided answers (translated from German:

Did you carry out all the tasks as requested in the respective instructions? Please answer honestly – this answer has no consequences for you (the survey is anonymous)!

1) I completed all the tasks as requested in the instructions.

2) Sometimes, I clicked on something because I was unmotivated or didn’t know what I was doing.

3) I often clicked on something to finish quickly.

Of 20 participants, two reported that they sometimes provided random ratings because of fatigue. However, their ratings were not excluded from the analysis because the question was designed as a subjective assessment of their own performance, and the ratings provided by the two participants did not differ from those of other participants who reported having answered all questions appropriately.

A new group of participants that did not participate in the production task (20 students, seven male, 13 female, mean age: 26.25 years, median: 26 years, SD: 4.63 years, min.: 20 years, max.: 36 years) was recruited for the rating task. It must be noted that this is a pilot experiment with a small number of participants and a small number of blends, and larger numbers of participants and/or blends might have led to different results. The main aim of this experiment was to test whether similarity measures previously applied to lexical blends can be transferred to PN blends.

In a within-subjects design, the participants were presented with the items from Table 2. Most participants were native speakers of German (n = 16), two participants were bilingual, and two were non-native speakers of German. Regarding the exposure to lexical blends, 18 reported being familiar with lexical blends presented to them, and two reported being partially familiar with presented lexical blends. As for PN blends, 11 participants were familiar with presented PN blends, three were unfamiliar, and six were somewhat familiar (according to the scale used in the post-questionnaire, see Section 6.1.2). As in the production task (see Section 6.1.2), familiarity refers to the frequency of exposure to or experience with exemplars of lexical and PN blends.

6.2.3 Results

This section reports on the results of descriptive and inferential statistics for the rating task. The regression analysis was done using the lme4-package (see Bates et al. 2015). The effects of the fixed factors were plotted based on the effects-package (Fox 2003).

6.2.3.1 Descriptive statistics

The overall distribution of ratings for the originality of PN blends and the recoverability of source names are visualized in Figure 2.

Figure 2:

Overall distribution of originality and recoverability ratings.

The most frequent rating value for both aspects of creativity is 4 (rather original and rather recoverable) followed by 2 (rather not original/rather not recoverable). Recall that, according to Runco (2006), a creative unit is both effective and original. Therefore, the PN blends can be defined as creative (since the most frequent rating values are rather original and rather recoverable). The mean values are 3.17 for originality (median: 3, SD: 1.22) and 3.08 for recoverability (median: 3, SD: 1.41). The Spearman’s Rank Correlation between the overall ratings of originality and recoverability is positive and weak (rs = 1,053,185, p < 0.001). This means the higher the originality rating, the higher the recoverability ratings.

Regarding the group of PN blends with high similarity, the mean value for originality is 3 (median: 3.2, SD: 1.32), and for PN blends with low similarity, the mean for originality is also 3 (median: 3, SD: 1.12). Therefore, no similarity-based differences in originality ratings could be observed. As for the recoverability, the mean value for PN blends with high similarity is 3.69 (median: 4, SD: 1.34), and for those with low similarity, the mean value is lower: 2.48 (median: 2, SD: 1.21).

Finally, the participants with high exposure frequency (n = 11) and the participants who reported no or some exposure to name blends (n = 9) were compared.^[13] Participants who reported higher exposure frequency rated PN blends as less recoverable (total mean value 3.04, mean value for PN blends with a low similarity: 2.49, and mean value for PN blends with high similarity: 3.6) than the group of participants who reported some or no exposure to blends (total mean value for recoverability: 3.13, mean value for the recoverability of PN blends with low similarity ratings: 2.46, for PN blends with high similarity ratings: 3.8). At the same time, PN blends received overall higher ratings for originality from the participants with higher exposure to PN blends (mean value 3.22) compared to the participants without and with some exposure to name blends (mean value 3.11). Note, however, that further studies with a larger number of participants should confirm these results.

In what follows, the hypotheses mentioned above (see Section 6) will be discussed based on the results of the rating task. To test Hypothesis 1 (PN blends with high indel similarity values are more recoverable than original. By contrast, PN blends with low indel similarity values are rated as more original than recoverable), the median values of originality and recoverability for PN blends with high similarity values were compared using the paired one-tailed Wilcoxon test (Baayen 2008: 76). The results indicate statistical significance (V = 759, p < 0.001), i.e., the median value of originality ratings for PN blends with high similarity values is lower than the median value of their recoverability ratings. The paired one-tailed Wilcoxon test was also applied to PN blends with low similarity values. The results confirm that the median value of originality ratings is significantly higher than the median value of their recoverability ratings (V = 1,640, p < 0.001).

In contrast, Hypothesis 2 (PN blends with low similarity values are rated as more original than PN blends with high similarity values) was not completely confirmed since the differences in median values of originality ratings between both groups are not statistically significant (paired one-tailed Wilcoxon test: V = 1,535.5, p > 0.05). The second part of this hypothesis, namely that PN blends with high indel similarity values are easier to reconstruct than PN blends with low indel similarity ratings, is true because the differences in median values of recoverability are statistically significant (V = 485.5, p > 0.05), i.e., the median values of recoverability ratings for PN blends with high similarity values are higher than the median values of recoverability ratings for the opposite group.

To test Hypothesis 3, the participants were asked to rate the grade of recoverability of constituents and write down both name constituents of which they think the name blend consists. The Spearman’s Rank correlation between the rating values of recoverability and the number of correctly predicted source names is statistically significant (p < 0.001), moderate, and positive (rs = 0.6110702), which means that the higher the recoverability ratings, the more likely it is that a participant rated both source names correctly.^[14]

6.2.3.2 Inferential statistics

A mixed-effects linear regression model (Baayen 2008: ch. 7) was fitted to explore which factors significantly predict the ratings of originality of PN blends and the recoverability of source names (see research question 2). The exposure was coded with “0” if a participant answered both questions (regarding the familiarity with PN blends and lexical blends, see above) with yes; otherwise, if one of the answers or both answers were no or to some extent, the exposure was coded with “1”.

The predictor variables (reported exposure, domains, and the number of correctly recognized source names) were added stepwise to the null model, which included ‘ID’ (participants) and items as random intercepts. Participants also received random slopes for normalized indel similarity. The model with the recoverability of source names as a dependent variable has the following final formula:

rating_recognizability ∼ domain + constituents_recognized + (1 + normalized_indel_similarity |ID) + (1|item)

The model contains domains and the number of correctly recognized source names (constituents_recognized) as predictor variables, random intercepts for participants, and random intercepts and random slopes for normalized indel similarity. The model comparison was done based on likelihood ratio tests. The model with the reported exposure to blends as a predictor did not lead to an improvement of the model, as shown by an ANOVA, which compared the model with and without this predictor variable (χ (1) = 1.2947, p > 0.05). The coefficients of the fixed effects,^[15] standard errors (SD) and t-values are given in Table 3. The reference level for the dependent variable (rating values for the recognizability of source names) was 1 (clearly not recognizable).

Table 3:

Rating task, fixed effects recoverability.

	Estimate	SE	t-value
Intercept	1.9193	0.1843	10.411
domain_neighbors	0.8215	0.2436	3.373
domain_neutral	0.7358	0.2391	3.078
domain_politics	0.8449	0.2380	3.550
domain_sport	0.7636	0.2423	3.151
constituents_recognized	0.8411	0.1200	7.011
Marginal R 2	0.3150241
Conditional R 2	0.5163009
Nr. of observations	200

The effects of predictor variables are positive for reference level 1, which means that the higher the rating value for recognizability, the higher the probability that the domain from which the PN blend comes is neighbor, neutral context, politics, or sports duo. In contrast, lower rating values for recognizability are more likely from the domain love couples. Furthermore, the higher the number of correctly predicted source names, the higher the recoverability ratings. The effects of the predictor variables are illustrated in Figure 3.

Figure 3:

Fixed effects recoverability.

We see a positive correlation between the predictor variables at hand and recoverability ratings as a dependent variable in that the higher the rating values for the recoverability of source names, the higher the number of correctly predicted names and the higher the chance that the PN blend comes from the domain neighbors, neutral context, politics, or sports.

The model with ratings for the originality of PN blends as a dependent variable has the following final formula:

rating_originality ∼ constituents_recognized + (1 + normalized_indel_similarity |ID) + (1|item)

The model contains the number of correctly predicted source names as a predictor variable. Participants and items were entered as random intercepts. In addition, random slopes and random intercepts for normalized indel similarity were included in the model. The model comparison was done based on likelihood ratio tests. The model with domains as a predictor variable failed to converge. The model with reported exposure to blends as a predictor variable did not lead to an improvement of the model as shown by an ANOVA, which compared the model with and without this predictor variable (χ (1) = 0.931, p > 0.05). The coefficients of the fixed effects,^[16] standard errors (SD), and t-values are given in Table 4. The reference level for the dependent variable was 1 (clearly not original).

Table 4:

Rating task, fixed effects originality.

	Estimate	SE	t value
Intercept	3.0428	0.1415	21.509
constituents_recognized	0.2741	0.1269	2.161
Marginal R 2	0.02769214
Conditional R 2	0.3393081
Nr. of observations	200

The table shows the positive effect of the number of correctly predicted source names on the ratings of originality. The effect of the predictor variable is visualized in Figure 4. As mentioned above, the effect of the number of correctly recognized source names (none, one source name, or both) is positive: the higher the originality ratings, the more accurately the source names were identified.

Figure 4:

Fixed effects originality.

Looking back at the results of the inferential statistics, the analysis shows that the grade to which PN blends are interpreted as creative is predicted by the number of correctly recognized source names and domains (recoverability ratings). The originality ratings are predicted only by the number of correctly recognized source names.

7 Discussion

Several research questions have been investigated using descriptive and inferential statistics. The results can be summarized as follows:

As for the first research question (How does the formal similarity between a blend and its name constituents influence the creativity of PN blends?), the correlation analysis confirms Hypothesis 1 that PN blends with high similarity values (see Table 2) are rated as more recoverable than original and vice versa: PN blends with low similarity values were rated as more original than recoverable. The reason is that the blends with high similarity ratings follow the maximization principle. In contrast, the other group follows the principle of least effort, where redundancy is avoided, and brevity is preferred (see Section 2). As the correlation analysis shows, in contrast to recoverability ratings, the originality ratings of PN blends with low and those with high similarity ratings are similar (Hypothesis 2 partially confirmed). Finally, the correlation between the rating values of recoverability and the number of correctly predicted source names is statistically significant and positive (Hypothesis 3), which confirms the results obtained for lexical blends (see Lehrer 2007): the higher the recoverability ratings, the more likely the participants rated both source names correctly.

The factors significantly predicting the creativity of PN blends (see research question 2) include the number of correctly recognized source names and the participants’ domain-based knowledge about PN blends. The higher the rating values of the recoverability of source names, the higher the number of correctly predicted names and the higher the chance that the PN blend comes from the domain neighbors, neutral context, politics, or sports. Furthermore, the more source names are correctly recognized, the higher the originality ratings. This confirms the hypothesis that being only novel/original is not sufficient for a linguistic unit to be interpreted as creative. Moreover, a unit must also be interpreted as communicatively successful, and this is only possible if language users interpret a novel unit as meaningful. In the case of PN blends, this includes reconstructing the name constituents underlying the blend.

Both usage-based factors (the participants’ domain-based knowledge about PN blends and the reported exposure to lexical blends) significantly predict the similarity values of PN blends elicited in the production task (see research question 3). PN blends with higher indel similarity values are more likely to come from the domain of love couples than from neighbors, politics, or sports. Given that love relationships are the most common purpose for creating PN blends (see, for German, Filatkina et al. 2019, and for English, DiGirolamo 2012), the result shows that the participants associated name pairs with love relationships more frequently than with other types of relationships (e.g., politics). They preferred to ensure that both name constituents will be recognized in the blend by enhancing the similarity between the blend and the source names (i.e., reducing the number of insertions and deletions necessary to change the source names in order to create a blend). By contrast, other relationships may not be interpreted as a case of PN blending. As a consequence, the blend coiners deviated from the source names. The structures they formed for other relationships obey the principle of least effort (short and not redundant blends without overlap) rather than the principle of maximization. Furthermore, the higher the indel similarity, the higher the chance that the blend was produced by a participant who reported being familiar with blends (i.e., having experience with them). Given the two general principles underlying the blend formation (the principles of least effort and maximization, see Section 3), we can assume that the familiarity with concrete exemplars of PN blends motivated the participants to maximize the possibility that the PN blend can be understood rather than to create economical (brief and not redundant) PN blends where relevant phonological properties of source names are not preserved. In other words, the participants (blend coiners) with higher exposure to blends consider communicative success/effectiveness more important than originality.

Finally, the rating task yielded that both aspects of creativity are mostly rated as rather original and rather recoverable (rating value at 4) with similar mean values (3.17 for originality and 3.08 for recoverability, see Section 6.2.3.1). The correlation between the ratings of originality and recoverability is statistically significant, which confirms the definition of creativity underlying the current study, namely that both aspects of creativity are related.

Regarding the implications of the current study on word-formation theories, the results can be interpreted from the perspective of usage-based approaches to blending. By including the exposure to lexical and PN blends and the domain-based knowledge about PN blends, the paper tested how two usage-based factors predict the creativity of PN blends with different similarity grades and how they influence the similarity values of produced PN blends. The statistical analysis shows that linguistic experience does not significantly influence creativity ratings compared to domain-based knowledge, which influences recognizability ratings. In contrast, the production of PN blends with different similarity grades is influenced by both usage-based factors. Since the participants who reported higher exposure tend to produce PN blends with higher similarity values, they are less prone to producing formal variations of entrenched structures. This is so because PN blends with higher similarity values bear the AD structure (see Table 2), and this is a canonical form of PN blends in German (Filatkina et al. 2019). Although it can be concluded that language users re-use cognitively entrenched exemplars of PN and lexical blends to create novel PN blends, the results are rather explorative, given the small number of participants in the rating task, and should be verified in further experimental studies.

8 Conclusions

The paper aimed to account for how the normalized indel similarity between the PN blend and its name constituents is related to the interpretation of experimental PN blends as more or less creative (original and recoverable). Furthermore, the paper investigated the influence of two usage-based factors (frequency of exposure and domain-based knowledge) on how they predict the indel similarity grade between PN blends and their constituents and the creativity of PN blends bearing different similarity grades. To this end, two groups of PN blends obtained in a production study were divided into blends with low and high indel similarity values (see Table 2). They were used in the rating task.

It must be noted that other factors are likely to influence how language users assess the creativity of PN blends bearing different similarity grades and the similarity values of produced PN blends. Further experiments should consider variables such as different structural types of blends, phonotactic rules, or contexts in which PN blends are embedded.^[17] A comparison with lexical blends and corpus-based PN blends should provide further insights into the role of formal similarity in the production and interpretation of creative word-formation units.