2.1 On simile (vs literal comparison)

We share Israel et al.’s (2004: 125) understanding of similes as a category subsuming a range of explicit figurative comparative constructions, which in English include but are not limited to those with the prepositions/conjunctions like and as.^[6] Other properties proposed as criterial for distinguishing similes, especially from literal comparison, include:

1. the source and target entities are sufficiently unlike (cf. Miller 1993 [1979]: 373; Ortony 1993 [1979]);

2. the source entity is a model/representative/paragon of the property, i.e., exhibits the property to a high degree (Barnden 2015; Israel et al. 2004; Moon 2008; Wikberg 2008);

3. the source NP is non-referring; it is used attributively as a type representing the conceptual background for the description/evaluation of the referential target (Bredin 1998; Wikberg 2008: 133).

To summarize, in similes the target entity is described or evaluated by comparison with another “sufficiently unlike” source entity. The source entity is construed as a model of the property concerned, and the speaker intends the source NP to be interpreted attributively, rather than referentially, i.e. as representing the conceptual background for the description/evaluation of the referential target. None of these properties stand if construed as based on discrete binary oppositions.

A key semantic property distinguishing similes, e.g., This lawyer is like a shark, from literal comparisons like A barracuda is like a shark (examples from Barnden 2015: 49) is the figurative feel of similes that comes from mentally comparing “the unlike” (cf. Miller 1993 [1979]: 373). However, disparity/similarity are non-binary notions and are subject to construal (Israel et al. 2004: 126; Wikberg 2008: 127). Luckily, this can be accommodated into a framework like cognitive linguistics, which is comfortable with fuzzy categories and encyclopedic semantics (we return to this shortly).

Further, whereas in literal comparison source–target associations are not meant to go beyond what is stated, e.g. He is as tall as his mother (it is not implied that the mother is particularly tall), in simile, targets are associated with sources with intense features (Israel et al. 2004: 126–127). The sources exhibit the property to an extreme degree, allowing the speaker to convey her “superlative evaluation of the target” (Israel et al. 2004: 128). Sources in similes involve entities construable as a model, a representative (Hao and Veale 2010: 642; cf. also Ortony 1993 [1979]), or a paragon for the property in question^[7] (Israel et al. 2004: 127), e.g. he’s thin as a rake or even She’s as thin as Twiggy. In Barnden (2015), deploying sources with intense features into similes (and metaphors) is called “source-based exaggeration”. This feature needs to complement the “unlikeness” criterion if a construction of comparison is to be interpreted as prototypical. A non-exaggerative comparison like (of plutonium) shaped into a ball more or less as big as a grapefruit (COCA) does compare two unlike entities, but with no superlative evaluation of the target, it is a step closer to literal comparison.

Finally, similes may also differ from literal comparisons interpretively. From his philosophical angle, Bredin (1998: 74–75) draws a distinction between similes and what he calls “ordinary” (i.e. literal) comparisons by claiming that similes are predicative comparisons in which the predicate describes the subject, while literal comparisons are symmetrical comparisons with referentially independent subjects and predicates. Paraphrasing Bredin, in a predicative comparison, by asserting or denying the target’s likeness to something else we are enriching our knowledge of the target. In a symmetrical comparison, the target and source refer independently to two different things, whose likeness is being asserted or denied. This would seem like a sound basis for the distinction. However, in his plea for a more explicit examination in psychological processing studies of how similes (My lawyer is like a shark/like a shark I saw at the zoo yesterday) and be-form metaphors (My lawyer is a shark) link up with referential vs predicative interpretation, Barnden claims that “in the be-form metaphor case, but not in the simile case, a very strong default is to take a simple indefinite source term such as a shark ‘predicatively’, and not to even consider the referential possibility other than in highly exceptional circumstances” (2012: 274 [emphasis ours]). Theoretically, this leaves room for similes with referentially specific sources and further blurs the boundary between simile and literal comparison.

Furthermore, referentiality itself is a tortured linguistic term. With Givón (2001), we assume that referentiality is not a logical relationship between referring expressions (NPs) and entities existing in the Real World. It is a matter of mapping from linguistic expression to entities in the Universe of Discourse. The Universe of Discourse is a realm “established by a particular speaker who then intends entities in it to either refer or not refer” (Givón 2001: 439). Thus, a horse and it are meant to refer to an actual horse in the Universe of Discourse in She’s looking for a horse; it escaped last Friday and are only meant as the non-referential type in She’s looking for a horse; it had better be white (Givón 2001: 439). Referentiality is also not a binary notion. There is a continuum between clearly referring and clearly non-referring interpretations based on how strong the speaker’s intention is to refer to a specific individual (Givón 2001: 449). In He bought books Past Tense reference (realis) suggests a referential interpretation of books, but Plurality imparts the irrelevance of the individual reference of books. How does this inform our view of similes? While it may well be that most similes involve non-referring uses of source NPs, non-referentiality cannot be assumed or even read off from the grammatical coding of sources; rather, we may speak of degrees of referential strength, which is a combination of the grammatical devices employed on source NPs and speaker’s intent. To give one example, grammatically definite NPs tend to be interpreted referentially (Givón 2001: 441) – and we may reasonably assume the same applies to proper nouns, like Twiggy in She’s as thin as Twiggy. But, although the speaker arguably intends to refer to the individual by that name, Twiggy is actually invited into the Universe of Discourse to stand as the paragon of thinness, to represent the type “extremely thin people”, and is not quite referential. By recruiting this knowledge, i.e., by asserting the target’s likeness to Twiggy, we are enriching our knowledge of the target, leaving the source entity in the conceptual background.

Summing up, with Barnden (2015: 49) and others (Carston and Wearing 2011: 297–298; Israel et al. 2004: 126), and in line with our cognitive-linguistic orientation, we assume a non-discrete distinction between similes and literal comparison.^[8] The simile is a fuzzy category (Lakoff 1987; Rosch and Mervis 1975; Rosch 1977), whose prototype is represented by any construction of comparison which is/where:

1. figurative in the sense that the source and the target are construed as conceptually distinct entities (cf. human target vs environment source in He is deaf as a stone); they are even more figurative if the adjective itself is understood figuratively (hard for ‘unfeeling’ in hard as stone);

2. the source concept represents the model or paragon for the property (a status defined contextually, e.g. when a person is described as subtle as a cockroach crawling across a white rug [COCA]);

3. the source concept represents a non-referential type, intended to be interpreted as conceptual background relative to which the referential target entity is understood or evaluated;

4. connected to 3, the act of comparison is asymmetrical because the target and source of comparison differ in salience. In Langacker (1987: Ch. 3.1.2), comparison is understood broadly as a fundamental, complex cognitive event occurring in a variety of domains. It allows novel experience to be interpreted with reference to old experience, and therefore implies an inherent asymmetry between the old experience as the standard, and the new experience as the target of comparison. According to Langacker (2008: 58), “[t]he categorizing structure lies in the background, taken for granted as a preestablished basis for assessment, while the target is in the foreground of awareness as the structure being observed and assessed”.

To these, we add two properties to capture the difference between similes perceived as familiar/formulaic and thus as PUs, and those perceived as novel, on-line occurrences:

1. similes become conventionalized in a community of speakers and perceived as part of the phraseolexicon;

2. similes are most likely processed as units in their canonical, unmodified form.^[9]

A construction of comparison meeting all these criteria is the prototype of the simile category ([man] hard as a stone). Some similes are less prototypical either because a property is present to a lesser degree (e.g. conceptual distinctness) or because it is absent (e.g. non-referentiality of the source). Criteria 5 and 6 will not be discussed further. More comments on conceptual distinctness (criterion 1) are in order because it connects to criteria 2 (paragon-status of sources), 3 (non-referentiality of sources), and 4 (difference in salience in the source-target pair).

Conceptual distinctness is a fuzzy notion and has less to do with the ontological status of entities designated by the lexical items (e.g. human vs environment) than with the distinctness of the conceptual domains relative to which those entities are construed in context.^[10] For instance, two human entities may be construed as sufficiently distinct if one is a non-entity from the conceptual domain of mundane, ordinary life and the other a cultural icon construed against the razzle-dazzle of the world of the entertainment industry (my sister and Twiggy).^[11] In cognitive linguistics, the meaning of linguistic expressions is characterized relative to conceptual domains and there is a whole matrix of domains from which any can be recruited for the contextually fitting understanding of the linguistic expression. According to Langacker (1987: 158–161), a domain is central for the interpretation of a linguistic expression if it meets four criteria, of which the first three are pertinent here. A central domain is (a) characteristic for the designation of the linguistic unit, i.e. unique to the class of designated entities and thus sufficient to identify a class member; (b) generic, i.e. the knowledge is not restricted to specific exemplars but applies to whole classes of entities, (c) conventional, i.e. it represents knowledge that is shared within a community of speakers^[12] and (d) intrinsic, i.e. it makes no essential reference to external entities. When it comes to source entities, understanding similes like hard as stone or busy as a bee involves accessing conventionally shared knowledge about classes of entities like natural objects and animals. Understanding a simile like thin as Twiggy in turn, is based on the conceptual domain representing conventional knowledge about a specific exemplar, and only by extension about the generic category the exemplar represents. Finally, for semantically motivated similes, the knowledge recruited from the domains evoked by source NPs will be more or less characteristic (even if not unique to them), i.e. it will allow the construal of sources as typical hosts for the “extreme” properties involved in comparison. Hardness is not a unique property of stones nor is diligence a property unique to bees, but this knowledge is sufficiently characteristic of members of those classes. The bottom line is, the paragon nature of the source entity (prototypicality criterion 2) is part of what is characteristic knowledge about the source entity. The typically non-referential nature of sources (prototypicality criterion 3) resonates with genericness as the criterion for domain centrality; generic knowledge about sources is knowledge that pertains to the whole class/type and this general knowledge, accessed through source NPs, is deployed to characterize rather than to refer to specific entities known to the utterer. Finally, the difference in saliency (prototypicality criterion 4) between target and source entities stems from the function of source entities and their domains in acts of comparison. They provide background knowledge that allows for a target’s more vivid description or evaluation.

In (6a–e) below we list some modifications of the similes thin as (a) stick/tissue/paper, proposing a loosely linear progression from the most prototypical as-simile in 6a to the completely literal comparison in 6e. We ignore here the prototypicality criteria 5 and 6, since none of these creative examples are likely to be perceived as conventional or to be processed holistically. For easier reference, all examples include subscripts for the new source (S’), the target (T), and the property (P):

2.2 PU and simile modifications

PU modifications can be defined as deliberate, creative, and idiosyncratic ad hoc changes of the canonical PU structure and/or meaning which produce semantic, stylistic, affective, or pragmatic effects in discourse (Omazić 2015: 35). According to Omazić, PU modification rests on finding harmony between a set of contradictions involving some of the definitional properties of PUs: familiarity vs novelty; entrenchment vs creativity; stability vs variability; figurative vs literal meaning; automated vs creative production (2015: 36). There are now a number of corpus, discourse, and cognitive studies proving that PUs are variable and quite abundant in corpora (Geeraert 2016; Langlotz 2006; Naciscione 2010; Omazić 2002, 2003, 2007, 2015; Vo 2011). It has also been found that modifications occur with any type of PU, even semantically opaque ones (Langlotz 2006; Naciscione 2010; Omazić 2003).

Interestingly, according to Moon (1998), similes in particular tend to be lexicalized. Wikberg (2008: 129) attributes the lack of creativity to the relatively low frequency of the as-simile in the BNC corpus he explored.^[13] Nevertheless, Wikberg does acknowledge some innovations: “However, white as snow appears once as white as newly fallen snow, which might be a more realistic version today” (Wikberg 2008: 135). Contrary to Moon (1998), Omazić (2015) found in her BNC-based study of PU modification that “idioms of comparison” are among the most frequently modified PUs. In Omazić’s work, “idioms of comparison” involve constructions of comparison involving as (i.e. as-similes) and like. Hao and Veale (2010) examined a corpus of authentic ironic similes like as welcome as a root canal without anesthesia, where sources express properties ironically opposed to those coded in the adjectives. Veale finds the syntactic form of as-similes to be a scaffolding for creativity that can be exploited “with remarkable freedom”:

There is something appealingly democratic and unpretentious about similes. Not only are they pervasive in language, they are at home in any register of speech and any genre of text, from tabloid newspapers to romantic poetry (Fishelov 1992). Conveniently, most languages provide a wealth of pre-fabricated similes that are as well-known to native speakers as the adjectival features they serve to exemplify. [...] But just as importantly, languages like English make it easy for speakers to mint their own similes on the fly, by imposing low barriers to creation. (Veale 2012: 329)

There are, therefore, already a number of corpus-based studies exploring variation/modification in English as-similes. But to the best of our knowledge, no study has explored the specific type of modification examined here. We will rely on aspects of a model of PU modification proposed by Omazić. In her extensive work on the topic (2002, 2003, 2007, 2015; Omazić and Čačija 2020), Omazić sets up a dynamic model that accounts for principles and constraints on PU modification (cf. also Langlotz 2006). A closer examination is beyond the scope of this paper, but we lay out below the four elements of the model and briefly explain those that will be pertinent here:

1. Personal constraints: individual propensity for PU modification based on personality type, flexibility with language, etc.

2. Modification principles and constraints (inspired by Blending theory [Fauconnier and Turner 2002]), which include:

   1. constitutive principles of PU modification (essentially a roadmap for how to arrive from the canonical to the modified PU, involving notions like inputs, projections, emergent structure, etc.);

   2. modification principles (semantic, grammatical, and lexical constraints on what is possible; they set limits to how far one can go in modifying PUs);

   3. vital relations (modifications can be successful if they rest on salient, vital relations between elements of the base-form and the new context, like change, time, space, cause-effect, etc.).

      In brief, the principles and constraints under (2) ensure that the modified PU is recognizable as being based on the canonical form, that it is grammatically coherent, that it is interpretable in the new context, and that users can grasp its relevance.

3. Contextual constraints: type of medium, type of discourse and genre, etc.

4. Functional constraints: a PU modification needs to fulfil a purpose to be effective since purposeful modifications create new meaning, provoke, impress, entertain, etc.

This study is about a type of simile modification that is clearly purposeful, possibly highly author-style/personality driven, and discourse sensitive. It is paradigmatic (featuring lexical change), and involves syntagmatic structural realignment, if merging the original source and property lexemes can be seen as part of the modification process itself. If indeed it can, we could argue that our modifications weaken Moon’s (2008) claim that, because of the basic structure of as-similes, there is little room for syntagmatic variation. The type of modification examined here is also easily recognizable and interpretable since the “old” source of comparison is not discarded after the new one is introduced. Assuming that the innovator must have had some familiarity with the original simile to bend it, we may consider the original simile as part of the established phraseolexion, regardless of how close or distant from the as-simile prototype it may be. That being the case, the general principles and constraints applicable to modifying PUs are likely to play a role here too. The modification constraints that may be specific to this type of modification, like the syllabic size of adjectives and original source nouns, will be addressed in Section 4.2. We will not explore the functional or author/personal constraints since these mainly involve processing issues concerning the speaker’s motivation/production and the hearer’s reception of modifications. Neither will we work with the global constitutive principles or Blending theory-specific constructs like “vital relations”. We will use quantitative and qualitative methods to examine:

1. the relationship between some distributional and formal features of what we will define below as base-form similes and the Matryoshka-type modification/Matryoshka-type similes (from here on M-modification or M-similes);^[14] and

2. the semantic and discourse-related features of the M-similes themselves.

A. Distributional and formal features of base-form similes

The following distributional and formal features of base-form similes will be explored for their association with M-modification:

1. The overall token frequency of what we call “base-form similes” (B- similes). Base-form similes are here defined as all our corpus-retrieved similes that fit the [X_NP ... (as) Adj as Y_NP] schema and act as a potential scaffolding for M-modification.^[15] Note that B-similes are defined in purely structural terms, which means they include canonical similes and those that already count as modifications because of lexical substitutions in the source or property slots e.g. white as snow vs white as the moon (COCA) vs white as the frost balls on a Christmas tree (COCA); silent as a tomb vs quiet as a tomb.^[16] This allows us to avoid making arbitrary decisions as to which B-similes are part of the canon. Instead, we merely register the objective measure of frequency to establish if there is a difference in token frequency between the two subsets of B-similes, viz. the B-similes that have been found to M-modify and the B-similes that have been found not to M-modify. From here on, we will be referring to these as “two subsets of B-similes”. The reasonable assumption is that B-similes with higher token frequencies may have lost their luster and may be generally more prone to M-modification than B-similes of lower token frequencies. Some support for this comes from claims that frequency of exposure breeds familiarity and diminishes information value: “Core use is not always best suited for the purpose of verbalizing a person’s feelings, thoughts, and experiences, nor does it adequately convey the meaning which the discourse situation requires. As a frequently recurrent item, core use sounds more common and conveys less information in comparison with the infrequent unusual stylistic instantiations” (Naciscione 2010: 39, [emphasis ours]).^[17]

2. Token frequency of cognate noun-adjective compounds (CNACs) (e.g. blood-red). The reasonable assumption is that preexisting CNACs may have precipitated M-modification by strengthening the entrenchment, and weakening the expressivity of B-similes, i.e. of the original source – property pairs. This assumption could be supported by Norrick’s (1987) claim about striking semantic similarities between the meaning relations expressed in similes and CNACs (see also Moon 2008: 7; for a different view see Novoselec and Parizoska 2012); and by that of Moon (2008: 32) who, arguing a different point, submits that highly frequent CNACs may make the simile sound familiar even when the simile is not very frequent itself.

3. The family size of alternatives. This is defined as lexical variation of sources, i.e. the sum of B-similes with similar or identical meanings where the same property is matched with a number of different source NPs, e.g. white as X = white as a sheet/snow/a lily/the moon/bone. We will refer to such semi-schematic similes with open source slots as “B-simile templates”, and sometimes, for simplicity, only as “templates”.^[18] The reasonable assumption is that B-simile templates with bigger families of sources already show a tendency to modify and may also be prone to M-modification.^[19]

4. The frequency of source NPs in general (i.e. in the corpora used), and the frequency of their use as sources of comparison in the B-simile databases. Whether or not the token frequency of each complete B-simile, i.e. its “overuse”, has any bearing on the existence of its M-modification, higher token frequencies of B-simile components, in particular source NPs, may have rendered those sources too worn out to be useful. The reasonable assumption is that B-similes with source NPs that are generally more frequent in corpora and/or more frequent in our B-simile databases are more likely to M-modify. This assumption is partly justified by Mancuso and Laudana (2019), who report a positive correlation between the frequency of manipulated idioms and the frequency of lemmas: idioms which more often occur in corpora in a manipulated form involve frequent words. It remains to be seen whether the same applies to M-modification.

5. The syllabic size of adjectives in the two subsets of B-similes. Since the adjective needs to become the second CNAC member in the M-simile, there may be restrictions on its size. The assumption is that B-similes with shorter adjectives are more likely to be associated with M-modification than those with longer adjectives. Admittedly, Moon (2008) established that the majority of adjectives in her corpus of similes were monosyllabic anyway. It remains to be checked whether this is the case in our databases, and whether there might still be a difference between the syllabic sizes of adjectives in the two subsets of B-similes.

6. The syllabic size of source nouns within syntactically simple source NPs in the two subsets of B-similes. Since the source noun needs to become the first CNAC member in the M-simile, the assumption is that M-modification is more likely with short source nouns.

Since we could not afford a semantic and discourse analysis of the entire B-simile databases, only M-simile tokens will be analyzed for their lexico-semantic and discourse-related features.

B. Lexico-semantic features of M-similes

B.1 Lexico-semantic features of old sources, new sources, and targets

1. The semantic domains of the old source entity, the new source entity, and the target entity (e.g. Environment, Tools, Animals, etc.). We also consider here: (a) whether M-similes show uneven affinities to particular lexemes from the same semantic domain for their old sources. For instance, although the Human Body domain is exploited more than some other domains, not all of its lexemes are (blood is common in M-modification: blood-red as... but lymph is not used at all); (b) the syntactic complexity of new and old sources (cf. the simple old source and the complex new source in as bone-dry as a bale of wastepaper compacted by feelings of helplessness [COCA]).

2. The match between the semantic domains of old sources and targets, new sources and targets, old sources and new sources, and among all three nominal entities at once (e.g. no match in: Anxiety_i as flour_j-fine as sand from Aram_k [Google]; where _i = Emotion/State/Disposition, _j = Food and Drinks, _k = Environment).

B.2  Lexico-semantic features of properties implicit in the three nominal entities

1. The semantic domain of the properties implicit in old sources, new sources, and targets (e.g. Temperature, Color, Textural Smoothness, etc.). We also consider here whether M-similes show a preference for particular adjectives from the same semantic domain. For instance, although the Color domain is heavily exploited, not all of its lexemes are (equally frequently) used (white, black and red are common in M-similes, e.g. lily-white as Norway [Google], jet-black as Zoro’s costume [iWeb], blood-red as a vampire [COCA], beige is not found at all).

2. The match between the properties of old and new sources, old sources and targets, new sources and targets, and among the properties of all three entities at once (e.g. a stranger_i ... steel_j cold as ice_j [Google]; where _i = Emotion, _j = Temperature)

Our semantic categories, i.e. domains, are inspired by Dixon (2005), Givón (2001), and Moon (2008). However, they should not be considered discrete since they are also a matter of construal. For instance, lobster could be categorized under both Animal, and Food and Drinks, but it will be tagged as Food in lobster-red as a sunburn (Google), since lobsters only turn red after cooking.

C. Discourse-related features of M-similes

1. Which text varieties are typical hosts for M-modifications?

   Because of the complex and ever-changing landscape of register/genre/text types, especially those inherently Internet-based (e.g. blogs, cf. Giltrow and Stein 2009), the categories we will use in addressing this context-based feature of M-modification should be seen as our tentative proposals pending a more dedicated discourse-oriented analysis.

4.1 M-similes: general descriptives

Since M-similes are the main focus of this study, we first present in Table 1 general descriptives for M-similes found in the three corpora.

Type frequency of M-similes is the total number of different M-similes corresponding to the B-similes in our databases. Token frequency of M-similes indicates the sum of all tokens of M-similes corresponding to the B-similes in our databases. Thus, blood-red as a raw steak (COCA) and lobster-red as a sunburn (Google) count as two M-simile types, going back to red as blood and red as lobster, respectively. The two instances of M-modification of red as blood attested in COCA – blood-red as a raw steak and blood-red as a vampire – count as two tokens of one M-simile type. Note that we treat as tokens of a single M-simile type all the M-similes that share the old source and property (above blood and red), regardless of whether they also share the new source. In the blood-red M-similes, the two tokens do not share new sources (a raw steak vs a vampire), but in pitch dark as night (COCA) and pitch-dark as the night (COCA) they do. The type/token ratio is the averaged measure of dispersion of M-simile tokens across M-simile types. Values closer to 0 imply that a smaller number of M-simile types accounts for all attested M-simile tokens; values closer to 1 suggest greater variety, viz. the tokens of M-similes are spread out across more M-simile types. Note that this ratio needs to be interpreted relative to the overall number of types and tokens. Theoretically, where there is only one type and one instance of a target structure in a corpus, the T/T ratio would be the ideal 1, but in reality, there would be no variability. Where, however, there is at least some (albeit small) number of types and tokens (as is the case here in iWeb and COCA), then even within such small pools of examples, we may speak of lower or higher T/T ratios, relatively speaking. In our databases, iWeb has a low T/T ratio because many M-simile tokens (N = 19) belong to a single M-simile type, i.e. that based on the B-simile clear as crystal. For such instances we might tentatively propose that low T/T ratios suggest that new tokens of M-similes may (also) be built by analogy to familiar exemplars rather than from productive schemas like [(as) CNAC as NP].^[25] All in all, with the exception of Google, the type and token frequencies of M-similes are quite low.

4.2 Distributional and formal features of B-similes & M-modification

4.3 Semantic features

The four semantic aspects of M-similes announced in Section 2.2 will be discussed in two subsections:

1. semantic domains of the three nominal entities: old source, new source, and target, and the existence of (mis)matches between/among them (Section 4.3.1)

2. semantic domains of the properties for which the old source, new source, and target are compared, and the existence of (mis)matches between/among them (Section 4.3.2)

Recall that this analysis was done at token level and only on non-duplicate M-similes from all three corpora (N = 289).^[31] Also, unlike in Section 4.2, no comparison was attempted between the semantic features of B-similes that have, and those that have not been found to M-modify.

4.4 Discourse-related features

In observing the kinds of text where M-similes can be found, we do not commit to any definitive distinction between individual register or genre types, or even generally between registers and genres. We adopt in principle the explanation of registers and genres as different perspectives on text varieties (Biber and Conrad 2009: 2). Both are defined functionally, by considering the communicative purposes of different varieties of language used in different situations: the register perspective focuses on a careful analysis of linguistic patterns associated with the situation of use or specifically, with particular purposes, topics, degrees of interactiveness, and mode whereas the genre perspective focuses on “the conventional structures used to construct a complete text within the variety” such as conventional ways to open or close a letter (Biber and Conrad 2009: 2).

Our discourse categories (Tab. 22) are a mix of what could be considered registers and genres. For instance, fiction (under creative writing) could best be analyzed as a register significantly associated with linguistic features typical of narrative production (Biber et al. 2000: 148). Blogs, on the other hand, qualify as a genre, at least in their original role as personal online diaries. However, none of the categories are discrete, and (sub)categories that carry different names may show increasing overlaps. A recent volume dedicated to Internet genres emphasizes the “volatility and chameleon-like properties of Internet genres”, adding that “[t]here is a constant and fast proliferation of genres–and of forms of communication that are candidates for being a genre. [...] Existing genres quickly differentiate into sub-species” (Giltrow and Stein 2009: 9). Blogs, for instance, started out as personal online diaries, but have now evolved into professionally-edited platforms for company advertising, delivering news, and instructional resources (so-called edublogs). As for their role in journalism, according to Britannica, “To meet increasing consumer demand for up-to-the-minute and highly detailed reporting, media outlets developed alternative channels of dissemination, such as online distribution, electronic mailings, and direct interaction with the public via forums, blogs, user-generated content, and social media sites such as Facebook and Twitter”.^[35] Essentially, any content created and shared (also publicly commented) via various social media or other services on a fairly regular basis can now be called a blog.^[36] In sum, the categories proposed below represent text varieties loosely based around communicative purposes like describing, commenting, evaluating, advertising, etc., purposes which make them suitable hosts for M-similes. The texts varieties cover largely similar topics (home, cooking, travel, food, celebrities, sports, culture, business, relationships, nature, music, etc.) and most have a notable degree of interactiveness (especially those in 4 through 8).

The M-similes found in the most prolific category of creative writing and non-fiction prose are ideally suited to the purposes of the variety, viz. to develop characters, vividly describe entities, including abstract feelings and the setting (see 24–25). Most examples were found in fiction (N = 85), then poetry (N = 18), creative non-fiction (N = 5), non-fiction prose (N = 4), and song lyrics (N = 3). According to Moon (2008: 30), “the strong association in BoE^[37] between as-similes and fiction suggests a convention of story-telling, with similes belonging to its sublexicon as a kind of descriptive commonplace […]”. The next most robust category is interpretive material featured in newspapers and magazines (see example 26), i.e. opinion and commentary articles found on the websites of media outlets. Note that, consistently with earlier claims (Moon 2008: 21), M-similes were not found in dry, informative, matter-of-fact news reports.^[38] Creative M-similes also made a notable appearance on company/commercial websites (see example 27). A web presence is a now a must in running a business. Via their websites, companies relay important information to the readers. Amongst other things, they describe and advertise their products and services, sometimes relying on nifty (M-)similes to catch the eye of potential customers. The remaining categories represent some of the most interactive and most informal text varieties, which may involve emotionally tense situations or discourses of conflict (Naciscione 2010: 42); it is here that users mint M-similes to provide fresh and exciting perspectives, characterizations and evaluations, exchange wit, opinions, insults, questions and answers, share stories and experiences.^[39] There is virtually no limit to the topics discussed on non-commercial websites/blogs of online communities (communities of people with shared interests), personal blogs (understood as online personal diaries), forums/discussion sites (websites, or sections of a website^[40] which allow people to post messages and create open group conversation in the form of posted messages), social networks like Twitter and Facebook, and question and answer sites (Quora): they cover celebrities, food, home furnishing, personal accessories, nature and environment, relationships and emotions, spirituality, theater, travel, gaming and technology, etc. (see examples 28–31):

1. The sky was pitch-black as ink. (iWeb: fiction)

2. The color red is related to the undead. Decomposing corpses often acquire a ruddy color, and this was generally taken for evidence of vampirism. Thus, the folkloric vampire is never pale, as one would expect of a corpse; his face is commonly described as florid or of a healthy color or dark, and this may be attributed to his habit of drinking blood. (The Serbians, referring to a redfaced, hard-drinking man, assert that he is “blood red as a vampire.” (COCA: nonfiction prose)

3. That is necessary to keep profits on track in an industry where margins can often seem as wafer-thin as a slice of supermarket ham (Google: newspaper)

4. Old Growth Imperial Stout is ebony dark as the night skies around winter solstice (Google: commercial website)

5. I’m sorry but it was too late. To Sheree its crystal clear as the Maui waters that Bob hasn’t changed (iWeb: RealityTea blog for Reality TV shows)

6. Prince you are beautifully masculine, composed and you are as porcelain smooth as an Irish stout on a cool summer’s evening. (Google: blog)

7. My experience with the HTC One is that this phone is as rock solid as Mt. Everest, in both build and operation (iWeb: Brighthand forum)

8. He is pitch dark as the winter’s night. Tomorrow at 10:24 approved stallion Feel Good will start in the finals for 7 year olds (Google: Facebook)