Compounds and multi-word expressions in French

French compounds differ from Germanic compounds in two important aspects. First, while Germanic compounding complies with the Right-hand Head Rule (e. g. English postage stamp, German Briefmarke, Dutch postzegel), French, like other Romance languages (see the chapters by Masini (Italian) and Fernández-Domínguez (Spanish) in this volume), has a general tendency towards lefthand headed compounding (e. g. timbre-poste lit. stamp-post). Second, whereas languages such as Dutch and German establish a clear demarcation between compounds and lexicalized phrases on the basis of formal criteria (spelling, prosody, linking elements, loss of adjectival inflection in [A N] compounds), French compounds are not easily distinguishable from syntactic expressions, and true compounds in Germanic languages often correspond to syntactic multi-word units in French (e. g. English admission ticket vs. French billet d’entrée (lit. ticket of entrance)) (Zwanenburg 1992: 222; see also the chapters by Booij (Dutch), Schlücker (German) and Bauer (English) in this volume). Contrary to Germanic languages, French has no distinctive word stress, only phrase stress. Moreover, whereas Germanic compounds may present linking elements (e. g. Dutch zonnebril, German Sonnenbrille ‘sunglasses’), these do not occur in French. Furthermore, the spelling of French multi-word units is characterized by many inconsistencies and irregularities: many combinations can be spelled with or without a hyphen (e. g. bébé(-)éprouvette ‘test-tube baby’ (lit. baby(-)test tube), porte(-)monnaie ‘coin purse’ (lit. carry(-)money)) or even as one word (e. g. portefeuille ‘wallet, billfold’ (lit. carrysheet) (Lehmann/Martin-Berthet 2008). Spelling of complex lexical units as one word occurs (e. g. vinaigre ‘vinegar’ (lit. wineacid)), but it is far from being the rule (cf. French vin rouge vs. German Rotwein), and the French spelling rules are systematically updated by orthographic reforms.1 Finally, many French compound-like expressions have

internal inflection markers (e. g. beaux-arts 'fine arts'), while these are generally attributed to syntactic formations.
As a result, none of the formal criteria typically applicable in Germanic languages 2 allow for a straightforward differentiation between compounds and (lexicalized) multi-word phrases in French, and, accordingly, the term 'compound' is not always used in a consistent way in the literature on French morphology. As a matter of fact, 'compounding' is often used to refer to various types of complex lexical units regardless of the formation process (morphological or syntactic) (for an overview, see, for example, Van Goethem 2009 andVilloing 2012).
Van Goethem (2009) illustrates this in the domain of [A N] units. The Dutch compound zuurkool 'sauerkraut' (lit. sour-cabbage) can be distinguished from the lexicalized phrase zure regen 'acid rain' and the non-lexicalized syntactic phrase zure kers 'sour cherry' on the basis of its spelling (written as one word), its stress pattern (prominent stress on zuur in zúurkool while zúre kérs has double stress and zure régen has prominent stress on the noun regen, cf. De Caluwe 1990: 17) and the lack of inflection of the adjectival component zuur in the compound (cf. Booij 2002: 314). In French, however, these criteria do not apply and Van Goethem (2009) concludes that, leaving aside some exceptions that do not conform to regular modern French syntax (e. g. rouge-gorge 'robin' (lit. red-throat) and grandmère 'grandmother', cf. Van Goethem 2009: 246 f.), French [N A] and [A N] units are phrases and not compounds, whatever their spelling may be: whether written as two separate words (e. g. premier ministre 'prime minister'), hyphenated (e. g. cordon-bleu 'master chef' (lit. cord-blue)) or even as one single word (e. g. vinaigre 'vinegar' (lit. wineacid)).
In this paper, we will turn the focus to [N 1 N 2 ] units, but before doing so we will present the different approaches to complex lexical units in French and show previously written as separate words (with or without a hyphen), for example chauvesouris 'bat' (lit. bald-mouse), millepattes 'centipede' (lit. thousand-legs), passepartout 'pass key' (lit. pass-everywhere), portemonnaie 'coin purse' (lit. carry-money) and véloski 'skibob' (lit. bikeski). (Internet: www.lalanguefrancaise.com/guide-complet-nouvelle-orthographe, last access: 18.4.2017). 2 In this respect, English may be considered to occupy an intermediary position: the traditional distinctive criterion applicable to English is the stress pattern, compounds being typically characterized by fore-stress (e. g. black bírd vs. bláckbird, cf. Bauer 2004 and this volume), but even this criterion is not always straightforward and many mismatches can be observed: as shown by Bauer (2004), a lexicalized phrase such as prímary school has first-element stress (or compound stress), whereas first-áid, with the two components hyphenated and unified, has second element stress (or phrase stress). These inconsistencies also apply to [N N] formations: péanut oil, for instance, has fore-stress, whereas olive oil may have end stress (cf. Bauer 1998, this volume;Giegerich 2009aGiegerich , 2009b. how true morphological formations (i. e. compounds) can be distinguished from multi-word phrases (Section 2). At the end of this section, the possible benefits of a constructionist approach to the issue will be highlighted. Section 3 will concentrate on [N 1 N 2 ] lexical units, which turn out to be the most problematic case in French since it is not easy to determine whether this formation belongs to syntax or morphology. In Section 4, a specific subtype, that of subordinative [N 1 N 2 ] units, will be examined because the latter most severely challenge this morphology-syntax divide. Whereas Fradin (2009) considers these formations to be true compounds, we will show that this only holds for the classifying subtype, and not for the qualifying one. Section 5, finally, will be devoted to a constructionist account of qualifying subordinative [N 1 N 2 ] formations, followed by the conclusion in Section 6.

Complex lexical units in French: four approaches
The notion of compounding generally has a more extensive scope in French morphology than in the literature on Germanic languages. Van Goethem (2009) identifies three different approaches. The common view is 'non-restrictive' in the sense that it includes all kinds of complex lexical units, regardless of whether they are formed in morphology or syntax (2.1). According to the 'scalar' approach (2.2), compounds are considered the endpoint of a scale of 'lexicalization' (used here to refer to the process of becoming a lexical item). The 'restrictive' or 'lexicalist' approach (2.3) aims to establish a clear demarcation between compounds and multi-word phrases. In what follows, we will outline these three different approaches. In 2.4, finally, we will add a fourth perspective and briefly show how complex lexical units can be accounted for from a Construction Grammar perspective.

The non-restrictive approach
In their overview article of multi-word expressions, Hüning/Schlücker (2015: 454 ff.) convincingly show that (syntactic) multi-word expressions and wordformation units (i. e. compounds) share a set of properties. Both are complex expressions with (potential) status as a lexical unit, and both expressions typically serve as linguistic signs for specific concepts (i. e. they have a 'naming function', cf. also Schlücker/Hüning 2009). Lastly, lexicalized phrases and compounds may have compositional or non-compositional semantics and may contain constituents with metaphorical semantics.
In French, formations such as [N de N] (e. g. fil de fer 'iron wire' (lit. wire of iron)), [N à N] (e. g. verre à vin 'wine glass' (lit. glass to wine)), [N à Det N] (e. g. sauce à l'ail 'garlic sauce' (lit. sauce to the garlic)), [A N] (e. g. Moyen Âge 'Middle Ages') and [N A] (e. g. poids lourd 'heavyweight' (lit. weight heavy)) (Fradin 2003: 199;Booij 2010: 172) are constructed by means of syntactic rules, as manifested through the presence of prepositions, determiners and adjectival inflection. Nevertheless, like compounds, they are productively used in name formation and it is therefore not surprising that the notion of compounding is often extended to all kinds of complex lexical units with a naming function, regardless of the formation rules. This approach can be illustrated by Mathieu-Colas's (1996)

The scalar approach
A second approach is to establish a scale of lexicalization ranging from free syntactic phrases over (semi-)lexicalized phrases to true compounding. Such a scale contains, by definition, a large transition zone in which it is not easy to decide whether we are dealing with syntactic phrases or with compounds.
This idea of a scale of lexicalization of complex units can be found in studies by Gross (1988Gross ( , 1996, who argues that lexicalized phrases and compounds can be distinguished from free syntactic phrases by means of semantic and syntactic parameters of lexicalization ('figement'). Semantically, lexicalized phrases and compounds such as fait divers 'novelty, piece of news, news item' (lit. fact diverse) are typically characterized by 'non-compositionality', in contrast to free syntactic phrases such as fait évident 'obvious fact', which have compositional semantics. Syntactically, in lexicalized [A N] or [N A] expressions the adjective loses the possibility of 'actualization' (1) and of 'predication' (2) (cf. Gross 1996: 31-34 On the basis of these parameters 3 , Gross (ibid.) distinguishes between different degrees of lexicalization. Cordon solide 'solid rope', cordon électrique 'power cord' (lit. cord electric) and cordon(-)bleu 'master chef' (lit. cord(-)blue) illustrate three different degrees of lexicalization: cordon solide is a free syntactic noun phrase ('groupe nominal libre'), cordon électrique is considered a semi-lexicalized noun phrase or compound ('un groupe nominal ou nom composé semi-figé') and cordon(-)bleu is called a lexicalized compound ('un nom composé figé'). However, as rightly observed by Corbin (1992: 36), Gross still uses the term 'compounds' ('mots composés') to refer to all lexicalized and semi-lexicalized combinations: both cordon électrique and cordon-bleu are called 'noms composés', whatever the differences may be in structure or degree of lexicalization. In other words, similar to the non-restrictive approach, the notion of compound is still applied to all structures with a naming function, including syntactic expressions.

The restrictive or lexicalist approach
In a modular approach to grammar, it has to be accepted that phrasal multi-word expressions and compounds, notwithstanding significant similarities, are different, the most crucial distinction being the fact that they are constructed according to the rules of different components of the language system (syntax vs. morphology).
A third theoretical tradition in French morphology, whether or not inspired by the 'lexicalist' approach in Generative Grammar (Di Sciullo/Williams 1987) and represented by Benveniste (1974), Corbin (1992Corbin ( , 1997, Zwanenburg (1992), Fradin (2003Fradin ( , 2009) and Villoing (2012), among others, follows this view and argues that a clear distinction should be made between compounds and lexicalized phrases. Although both strategies may have the same naming function, they obviously fit into different parts of grammar, compounds belonging to morphology and phrases to syntax.
These authors argue, for instance, that [N Prep N] combinations such as pomme de terre 'potato' (lit. apple of ground) and sac à main 'handbag' (lit. bag to hand), commonly considered compounds in French, should be analyzed as lexicalized syntactic phrases since they respect the general principles of word order and syntax in French.
The most extreme position can be found in Di Sciullo/Williams (1987), who claim that French does not have any compounds at all: It now appears that French (and no doubt Spanish) lacks compounding altogether. Once we have subtracted fixed syntactic phrases (idioms) such as timbres-poste and phrases reanalyzed as words (syntactic words) such as essui-glace <sic>, there are no candidates left. (ibid.: 83) Corbin (1992Corbin ( , 1997 is less restrictive and preserves the term 'compound' to refer to lexical units of the type [N 1 N 2 ] (e. g. timbre-poste 'postage stamp') and [V N] (e. g. essuie-glace 'windscreen wiper') because they are formed according to lexical composition rules, specific to the lexicon and different from syntactic rules. Corbin (1997) uses the notion of 'polylexematic units' ('unités polylexématiques') as a general term for covering both compounds and lexicalized phrases. However, both naming strategies are distinguished on the basis of the 'division of labor principle' between morphology and syntax. According to this principle, also labeled the 'Lexical Integrity Hypothesis' (LIH hereafter), syntax has no access to morphological operators or infralexical units and, conversely, morphology has no access to syntactic operators: 4 Les règles syntaxiques n'ont accès ni aux opérateurs morphologiques ni à des unités infralexicales. Les règles morphologiques n'ont pas accès aux opérateurs syntaxiques. (ibid.: 83) On the one hand, this implies that affixed polylexematic units such as fil-defériste 'high wire walker' (lit. wire-of-iron-ist) belong to morphology, since syntax cannot attach affixes. On the other hand, polylexematic units containing a syntactic operator, a preposition as in verre à vin 'wine glass' (lit. glass to wine) or a determiner as in hors-la-loi 'outlaw' (lit. outside-the-law), necessarily belong to syntax. 5 In other words, polylexematic units are exclusively formed either by syntax or by morphology, and the idea of a scale is thus rejected: En vertu du partage des tâches entre les modules d'une grammaire, les séquences engendrables syntaxiquement ne le sont pas morphologiquement et réciproquement. (ibid.: 84) 6 On the same grounds, Fradin (2009: 418) excludes expressions such as sans papiers 'person without identity papers, illegal immigrant' (lit. without papers) and pied-à-terre 'pied-à-terre, holiday cottage' (lit. foot-on-ground) from true compounding because they correspond to phrases that can be generated by syntax (cf. Il s'est retrouvé sans papiers 'he ended up without (identity) papers' and Le cavalier mit pied à terre 'the horseman dismounted' (lit. put foot on ground)). He relabels Corbin's proposal as 'Principle A': Whereas in Corbin's (1997) view, only [N N] and [V N] configurations can be considered true compounds, Fradin (2009) concludes that not two but four productive compounding patterns can be retained in French: auteur-compositeur 'author-composer') and [N N] subordinates (e. g. pois son-chat 'catfish' (lit. fish-cat)). Villoing (2012: 36) adds to this a particular subclass of [A N] compounds with a color adjective as head (e. g. bleu-ciel 'sky blue' (lit. blue-sky)). She argues that all these formations should be considered true compounds because they all display syntactic anomalies: -VN compounds: the absence of a determiner between the verb and the noun, and a diverse range of semantic relations between the verb and the noun (ouvre-boîte 'can opener' (lit. open-can)), -coordinated NN (horloger-bijoutier 'jeweler-watchmaker' (lit. watchmaker-jeweler)) and AA (aigre-doux 'sweet and sour' (lit. sour-sweet)) compounds: the absence of a coordinating conjunction between the constituents, -all other NN compounds (poisson-chat 'catfish' (lit. fish-cat), pause-café 'coffee break' (lit. break-coffee)): hyponymic interpretation, -AN compounds (bleu-ciel 'sky blue' (lit. blue-sky)): the presence of an adjectival rather than a nominal head.
(paraphrased from Villoing 2012: 36) Villoing (2012: 30) specifies that French native compounding 7 'is prototypically formed of two lexemes of the current lexicon of French, without any linking element; the internal order of constituents is XY, where X is the governing element'. Furthermore, the composing lexemes belong, by definition, to the major word classes (noun, verb, adjective), and are uninflected. This implies that 'no constituent is marked by inflection: no modality, tense, person or aspect marking on the verb in VN compounds, no number on the N, and no gender or number on adjectives, disregarding cases of agreement' (ibid.: 31 f. This view implies that many other multi-word units that are often considered compounds do in fact belong to syntax and, therefore, need to be analyzed as lexicalized phrases. According to Villoing (ibid.: 35 f.), the following French multi-word units should not be analyzed as compounds: -Complex units composed of non-lexemes, such as complex prepositions and complex conjunctions: e. g. par-dessus 'from above', de sorte que 'such that' 10 -Lexicalized syntactic constructions, namely NPs (3), PPs (4) and VPs (5) that behave like lexical units: 7 Villoing (2012) distinguishes native compounds from neoclassical compounds, which have different properties: the latter are 'prototypically composed of two bases of Greek or Latin origin that are not syntactically autonomous in French, connected by a linking element; the internal order of constituents is YX, where X is the governing element' (Villoing 2012: 30) (e. g. ludo-thèque 'game library', homi-cide 'manslaughter', cyno-céphale 'dog head'). 8 However, Villoing (2012: 34) rightly observes that some compounds actually display inflected forms of the lexeme: for instance, many [V N] compounds include a plural N, orthographically and/or phonologically marked (e. g. presse-fruits 'fruit press' (lit. press-fruits), protège-yeux 'eye protector' (lit. protect-eyes)). Villoing argues that this plural inflection is not the result of syntactic marking, but of inherent and semantically motivated inflection. 9 This approach, in line with Corbin (1992), Villoing (2009), Bonami/Boyé (2003, 2014 and Fradin (2009), among others, implies that the V in French [V N] N compounds (e. g. ouvre-boîte 'can opener') is not an inflected form of the verb (imperative or present indicative), but a stem of the lexeme. 10 Although Zwanenburg (1992)

A constructionist perspective to complex lexical units
It can be concluded from the preceding overview that the term 'compounding' is not used consistently in the French linguistic tradition and often covers much more than, strictly speaking, morphological complex lexical units. Hüning/ Schlücker (2015) point out the commonalities and differences found between compounds as word-formation units and syntactically formed multi-word expressions. In spite of the differences, both patterns may serve the same purpose and even enter into competition to do so. printing (by) laser) (cf. also Section 3.1). These facts indicate that in French, too, the boundary between compounds and syntactic multi-word expressions is fuzzy and the data are suggestive of a lexicon-syntax continuum. This non-modular view of language is precisely a basic assumption of Construction Grammar (cf. Goldberg 1995Goldberg , 2006Croft 2001;Booij 2010;Hoffmann/ Trousdale (eds.) 2013, a. o.). Crucial to this model is the concept of 'constructions': these are conventional pairings of form (referring to syntactic, morphological and phonological properties) and meaning (including semantic, pragmatic and discourse-functional properties) and are considered the fundamental units of the linguistic system. All levels of grammatical description involve such form-meaning pairings -not only words as in the Saussurean tradition -and constructions vary in size, degree of schematicity and complexity (cf. Goldberg 2009), the minimal linguistic construction being the word in Booij's (2010) model of Construction Morphology. Furthermore, constructions, both syntactic and morphological, are linked to each other by (vertical) inheritance relations and also by (horizontal) connectivity links (Norde 2014; Norde/Morris 2018). As a consequence, language can be considered a complex network of constructions. Substantive constructions (e. g. petit mais vaillant 'small but tough', position clé 'key position') are instances of semi-schematic constructions (e. g. [Adj 1 mais Adj 2 ], [N 1 clé]), which -in turn -inherit properties from more general schematic constructions (e. g. [Adj 1 CONJ Adj 2 ], [N 1 N 2 ]). Moreover, constructions may also inherit properties from multiple-parent constructions via so-called 'multiple inheritance' (cf. Trousdale 2013; Trousdale/Norde 2013). 11 It is not surprising that many recent studies in the field of multi-word expressions are in the constructionist vein. In this approach, it can be assumed that both compounds and phrasal structures with a naming function can act as conventionalized form-meaning pairings or 'constructions', and we should accept the existence of what Booij (2010: 190) calls 'lexical phrasal constructions': these are syntactic formations that should be stored as lexical units in the mental lexicon, such as fil de fer 'iron wire' (lit. wire of iron) and moulin à vent 'windmill' (lit. mill at wind). These formations demonstrate that there is no strict boundary between the lexicon and syntax, or, as Booij (ibid.: 191) puts it, 'syntax permeates the lexicon because syntactic units can be lexical'.
Compounds and phrasal structures are not only closely linked in the constructional network; they may also compete or interact with each other. The process of 'multiple inheritance' may even produce hybrid constructions that inherit properties from parent constructions belonging to different domains, such as morphology and syntax. We believe that these insights from Construction Gram-11 The idea of 'multiple inheritance' could be seen as the synchronic representation of the complexity of language change. Diachronic developments do not always follow linear pathways from one source construction to another target construction; a complex interplay between different sources and processes is often at stake (cf. De Smet/Ghesquière/Van de Velde's (eds.) 2013 volume On multiple source constructions in language change). mar are useful to account for problematic cases that cannot be univocally classified as morphological or syntactic constructs, such as French [N 1 N 2 ] subordinatives. In Sections 3 and 4 we will therefore focus on these particular cases and in Section 5 we will propose an analysis in line with the constructionist insights.

French [N 1 N 2 ] sequences: compounds or phrases?
In Section 2.3, we observed that both Fradin (2009: 428 f.) and Villoing (2012: 36) admit that the boundary between morphological and syntactic units in French is most difficult to apply in the case of [N 1 N 2 ] formations. We will therefore now concentrate on Fradin's arguments to retain only [N 1 N 2 ] coordinates and subordinates as true French compounds, at the expense of other types of [N 1 N 2 ] sequences, namely so-called 'two-slot nominal constructs' and identificational [N 1 N 2 ] constructs (3.1). In Section 3.2, we will focus on subordinate [N 1 N 2 ] formations and show that their status is more problematic than acknowledged by Fradin (2009). Fradin (2009) distinguishes between four types of [N 1 N 2 ] sequences: coordinates, subordinates, two-slot nominal constructs and identificational constructs; the first two are assigned to morphology and the others to syntax.

Fradin's (2009) typology of [N 1 N 2 ] sequences
First, two types of [N 1 N 2 ] coordinates can be distinguished: in (6) each N has a distinct referent and the compound's denotatum is the sum of these referents; the compounds in (7), however, denote a unique referent combining properties of both N 1 and N 2 (ibid.: 429 f.): Bosnie-Herzégovine 'Bosnia-Herzegovina' physique-chimie 'physics-chemistry (as a teaching discipline)' (7) chanteur-compositeur 'singer-composer' hôtel-restaurant 'hotel-restaurant' As also argued by Villoing (2012: 36), the absence of a coordinating conjunction between the constituents excludes these sequences being generated by syntax, and they should therefore be considered true compounds. Unlike coordinate compounds, subordinate compounds only denote the referent expressed by N 1 (i. e. the head noun), while N 2 (i. e. the modifier) refers to one of its salient properties. According to Fradin (2009: 430 f.), this property may concern a physical dimension (shape, length, weight) (8), an intrinsic capacity (slowness, quickness, strength, duration) (9) or a function (10) Even though Fradin recognizes that the morphological status of these compounds is open to debate (cf. Section 4), he claims that the regular interpretative patterns found in these subordinate compounds are similar to those of some derived lexemes, such as French adjectives derived with the suffix -able (Fradin 2003). In the same way as productive suffixes, the N 2 of subordinate [N 1 N 2 ] formations can be combined with a broad range of stems and forms a productive constructional pattern with a regular interpretation. This similarity with derivation is taken as an argument in favor of their morphological status. Whereas coordinate and subordinate [N 1 N 2 ] sequences follow a constrained pattern and have a regular semantic relationship between the constituents, this is not the case with two-slot nominal constructs (Fradin 2009: 432 f.) and identificational [N 1 N 2 ] sequences. The examples in (11) all denote the referent expressed by N 1 , but they completely differ from subordinate compounds because N 2 does not refer to an intrinsic and salient property of N 1 . Moreover, the sequence usually corresponds to a syntactic phrase in which N 2 forms part of a prepositional phrase (12), which suggests a syntactic origin.  (13) la catégorie adjectif 'the adjective category' l'institution Opéra 'the Opera institution' N 2 identifies N 1 ('N 2 is an N 1 ') and from this point of view, these sequences are equivalent to syntactic (appositional) [N 1 N 2 ] constructs in which N 2 is a proper noun and N 1 expresses a socially recognized category (e. g. le président Mandela 'President Mandela', la région Bourgogne 'the region of Burgundy').

Discussion: morphological and syntactic approaches to [N 1 N 2 ] subordinatives
We agree with Fradin that [N 1 N 2 ] coordinates are true compounds and cannot be the result of syntactic formation. We also subscribe to his view on two-slot nominal and identificational [N 1 N 2 ] constructs: both sequences can be shown to correspond to syntactic phrases. However, subordinate [N 1 N 2 ] formations are more problematic than acknowledged by Fradin (2009) and it can be demonstrated that the examples mentioned for this class are not all of the same kind. At first glance, it can, for instance, be observed that some of them permit degree modification of N 2 while others do not (discours vraiment fleuve 'really lengthy discourse' (lit. discourse really river) vs. *requin vraiment marteau 'really hammerhead shark' (lit. shark really hammer)), and some but not all N 2 s form productive series (e. g. discours-fleuve 'lengthy discourse' (lit. discourse-river), roman-fleuve 'novel cycle', film-fleuve 'lengthy movie', débat-fleuve 'lengthy debate', etc.), while no series formation is possible for [N-marteau], for instance. We will discuss these differences more extensively in Section 4. As already mentioned, these formations have been the subject of some debate. Amiot/Van Goethem (2012: 350 ff.) and Van Goethem (2012: 77-81) provide an overview of the different accounts, which range from purely syntactic analyses (cf. Noailly 1990 andGoes 1999) to strictly morphological accounts, like the one by Fradin (2009).
With regard to the syntactic approaches, a distinction can be made between analyses where the second component of the phrase is still considered a noun in spite of some adjectival properties (cf. Noailly 1990, who labels N 2 as 'substantif épithète' and Arnaud/Renner 2014, who detect adjective-like syntactic behavior to some extent), and others like Lehmann/Martin-Berthet (2008: 206), who claim that N 2 is converted into an adjective if it complies with a set of criteria typical of adjectives (such as degree modification and predicative use).
With regard to the morphological approaches, we can contrast Fradin's classification of French compounding with the general typology of compounds by Scalise/Bisetto (2009) (applied to French by Villoing 2012), according to whom these 'problematic' compounds are not subordinatives but belong to the ATAP (attributives-appositives) class, and more particularly to the subclass of appositives: Attributive compounds can actually be defined as formations whose head is modified by a non-head expressing a 'property' of the head, be it an adjective or a verb: actually, the role of the non-head categorial element should be that of expressing a 'quality' of the head constituent. Appositives, to the contrary, are compounds in which the non-head element expresses a property of the head constituent by means of a noun, an apposition, acting as an attribute. (Scalise/Bisetto 2009: 51) As these definitions show, attributives (e. g. high school) and appositives (e. g. snailmail, swordfish) belong to the same ATAP class because they have similar functions. The metaphorical value of the modifier is argued to be an important distinctive criterion between [N 1 N 2 ] subordinatives (e. g. mushroom soup), on the one hand, and [N 1 N 2 ] appositives (e. g. mushroom cloud), on the other: In appositives that, together with attributives, make up the ATAP class, the noun plays an attributive role and is often to be interpreted metaphorically. Metaphoricity is the factor that enables us to make a distinction between, e. g. mushroom soup (a subordinate ground compound) and mushroom cloud, where mushroom is not interpreted in its literal sense but is rather construed as a 'representation of the mushroom entity' (...) whose relevant feature in the compound under observation is shape. (ibid.: 52) In the next section, we will take a closer look at this specific type of formation and will argue that we need to distinguish between two different subclasses: classifying and qualifying [N 1 N 2 ] subordinatives, of which only the former undoubtedly belong to morphology.

Classifying vs. qualifying [N 1 N 2 ] subordinatives
In this section we will argue that two types of [N 1 N 2 ] subordinatives should be distinguished: classifying (e. g. requin-marteau 'hammerhead shark' (lit. shark-hammer)) and qualifying (e. g. guerre éclair 'blitzkrieg' (lit. war lightning)). The difference can essentially be found in the different role of N 2 with respect to N 1 . 12 Despite their similarities (in all these subordinate compounds, N 2 denotes a salient, metaphor-based property of N 1 ), N 2 has a classifying role in some [N 1 N 2 ] formations (e. g. requin-marteau) but a qualifying role in others (e. g. guerreéclair). We will present the distinguishing properties of both types of [N 1 N 2 ] subordinatives in 4.1 and 4.2, respectively.

Classifying [N 1 N 2 ] subordinatives
Classifying [N 1 N 2 ] subordinatives are characterized by a number of particular semantic and syntactic properties: (i) Semantically, they behave like designations ('names'): they refer to stable concepts (Kleiber 1984), but their reference is established in a motivated way: N 1 , the semantic head, is the hyperonym and N 2 , which does not have a referential meaning, refers to a salient property of N 1 that allows the [N 1 N 2 ] sequence to be distinguished from other N 1 s. Hence, N 2 expresses a classifying property of N 1 . 13 This is why, at least when they denote biological species, classifying [N 1 N 2 ] sequences are often the vernacular denominations corresponding to scientific taxonomies: for instance, serpent-tigre corresponds to Notechis Scutatus, pin-par asol to Pina Pinea and oiseau-lyre to Menura Superba, etc. (cf. Ureña/Faber 2010 for English compounding). When [N 1 N 2 ] is not a vernacular denomination corresponding to a scientific taxonomy, it can at least integrate a hierarchical folk categorization (Wierzbicka 1996): for example, a fauteuil-crapaud 'squat armchair' (lit. armchair-toad) is a kind of armchair (fauteuil), in the same way as a club chair or a rocking chair. And, in turn, an armchair is a piece of furniture, etc. This signals the relationship of inclusion [X is a Y], typical of the hierarchy between a hyponym and its hyperonym.
(ii) N 1 often denotes a biological species, especially animals (14a), vegetables (14b) or sometimes human beings (14c). More exceptionally, compounds denoting artefacts can also be found (14d): (14a) poisson-scie 'sawfish' (lit. fish-saw) oiseau-lyre 'lyrebird' (lit. bird-lyre) serpent-tigre 'tiger snake' (lit. snake tigre) (14b) saule têtard 'silver willow' (lit. willow tadpole) pin-parasol 'umbrella pine' (lit. pine-umbrella) tomate-cerise 'cherry tomato' (lit. tomato-cherry) (14c) homme-grenouille 'frogman ' (lit. man-frog) femme-objet 'woman as object' (lit. woman-object) enfant-roi 'spoilt child' (lit. child-king) (14d) voiture-bélier 'ram-raid ' (lit. car-ram) fauteuil-crapaud 'squat armchair' (lit. armchair-toad) noeud-papillon 'bow tie' (lit. bow-butterfly) (iii) In these cases, and as opposed to coordinate compounds, the two nouns denote concrete entities that do not belong to the same semantic class and the metaphor that underpins the relation between N 1 and N 2 is often based on physical resemblance: the nose of a poisson-scie is shaped like a saw (scie) and a saule têtard has roughly the shape of a tadpole (têtard): a big head like the upper part (the foliage) of the willow, and a short tiny bottom part (like the trunk). In our examples, the only sequences that do not instantiate this relation are enfant-roi, femme-objet and voiture-bélier, in which the metaphor is based on behavioral resemblance. For example, an enfant-roi is a child (enfant) who is treated like a king (roi) and who often becomes a 'domestic tyrant'. (iv) Syntactically, all the linguistic tests usually used to measure the lexical integrity of a sequence (cf. Sections 2.2 and 2.3) show that these classifying [N 1 N 2 ] formations are words, insofar as they do not accept any of these manipulations, unlike the qualifying [N 1 N 2 ] subordinatives that we will study in Section 4.2.
(v) The last property to be mentioned is the fact that, unlike the qualifying [N 1 N 2 ] formations, these classifying subordinatives do not give rise to productive series.
We can conclude from this survey that the subordinate [N 1 N 2 ] formations like those exemplified under (14) are binominal words and true compounds in which N 2 metaphorically denotes a classifying property of N 1 .
(iii) It is precisely the qualifying role of N 2 that could, in our view, explain the specific behavior of these [N 1 N 2 ] formations, and particularly their lack of lexical integrity (cf. 'with a lightning punch (lit. punch-of-fist), she deflects the ball' In these examples, the N 1 s resemble phrases: they result from the association of a noun and an adjective (16a-b) or of a noun and a prepositional phrase (16c-d).
The N 2 slot can also be filled by a complex item, but this is more exceptional: Interestingly, a lexicalized multi-word expression such as coup de poing can fill, in its literal meaning ('punch'), the N 1 slot or, in its metaphorical meaning ('lightning'), the N 2 slot. It should be noticed that, since the complex sequences that may fill the N 1 or N 2 slots are lexicalized phrases, this is less problematic for the LIH than if they were free, compositional phrases (cf. Booij's (2010) use of 'lexical phrasal constructions' in 2.4).
(b) Most N 2 s can be modified by an adverb of degree, as shown in (18): (18a) on avait le sentiment d'assister à un moment vraiment charnière (www) 'we had the feeling of witnessing a truly pivotal (lit. hinge) moment' (18b) un conseil vraiment éclair (www) 'a really whirlwind (lit. lightning) council meeting' (18c) la multiplicité des voix de ce roman vraiment fleuve (www) 'the multiplicity of voices in this really lengthy (lit. river) novel' This second property is more challenging for the LIH: the lexical integrity of the [N 1 N 2 ] sequences is undoubtedly called into question by the insertion of an adverb of degree between the two Ns. This is why some authors put forward a weakened version of the hypothesis, including Ackema/Neeleman (2004), Booij (2005) and Lieber/Scalise (2007). 15 Our previously conducted corpus research (Amiot/Van Goethem 2012;Van Goethem 2012, 2015 indicate that the most frequently inserted adverb is vraiment 'really, truly', as in (18), but other degree adverbs can be found too: absolument 'absolutely ' (19), réellement 'really' (20), extrêmement 'extremely' (21) and even, but more rarely, très 'very' (22) The presence of such adverbs conflicts not only with the lexical integrity of the [N 1 N 2 ] sequence, but also with the nominal status of N 2 : usually an adverb of degree modifies a gradable adjective, not a noun. However, in the context of the qualifying [N 1 N 2 ] sequences, N 2 seems to switch to adjectival status. Syntactically, evidence for this adjectival status is not only provided by the possibility of modification by an adverb, but, like a qualifying adjective, N 2 can also be inserted into a comparative construction: This demonstrates the qualifying value of N 2 vis-à-vis N 1 . We will return to this in Section 5, but it is worth noting for the time being that this behavior distinguishes the qualifying subordinative [N 1 N 2 ] from the classifying subordinative (Section 4.1).
(c) Some N 2 s can be used predicatively. Predicative use is the most prototypical use of qualifying adjectives. In some cases, 'N 2 ' can fill the slot of an adjective in a predicative construction (25) with or without degree marking: (25a) La période est charnière également sur le plan économique (www) 'The period is also pivotal in economic terms' (25b) Leur rôle est ainsi plus clé que jamais (www) 'Their role is thus more key than ever' (25c) c'est déjà arrivé quand l'interview est vraiment fleuve (www) 'It has already happened when the interview is really lengthy' In this use, the [N 1 N 2 ] construction (période charnière in (25a), rôle clé in (25b) and interview fleuve in (25c)) is broken up, and N 2 acquires autonomous adjectival behavior. This separation of compound-like sequences has been labeled 'debonding' by Norde (2009)  The first step is the emergence of a productive constructional idiom -via so-called 'constructionalization' (Traugott/Trousdale 2013; Hüning/Booij 2014)in which N 2 develops a specific (metaphoric) qualifying meaning when combined with an N 1 in a compound(-like) sequence (e. g. question-clé 'key question', moment charnière 'pivotal moment', réunion marathon 'marathon meeting', cas limite 'borderline case ', etc.). This qualifying meaning may be seen as the result of 'coercion' (cf. Audring/Booij 2016) in which the metaphoric meaning sometimes already available for the noun outside the compound-like pattern (e. g. la clé du succès 'the key of success') is selected ('coercion by selection') and/or in which N 2 develops adjective-like (semantic and formal) properties within the [N 1 N 2 ] pattern ('coercion by override'). This semi-schematic construction, applied to the example of [N charnière] formations, can be represented as follows: (26)  However, the constructionalization of N 2 goes beyond this morphological stage, since it may occur in innovative syntactic constructions with the same semantics (cf. Section 4.2). As already suggested by Amiot/Van Goethem (2012) and Van Goethem (2015), the adjective-like uses of N 2 can be seen as the result of an interaction between the closely related morphological [N 1 N 2 ] N and syntactic [N A] NP constructions. 16 The fact that N 2 s such as charnière, clé, fleuve, limite and so on 16 The schematic representations are a bit simplified since, as we have seen in 4.2, N 1 and N 2 can include a multi-word sequence, and the A can be instantiated by a phrase in the case of degree modification (e. g. une période vraiment charnière). developed a qualifying meaning in the former construction, typical of adjectives, may have favored this constructional ambiguity. In constructional terms, this interaction can be translated as an instance of 'multiple inheritance'. Schematically, this multiple inheritance can be represented as in (27) (vraiment) crucial 'a (really) crucial moment'). As a consequence, and as shown in Section 4.2, it is a hybrid between a morphological and a syntactic construction and N 2 can, in some cases, gradually develop more adjective-like syntactic uses, such as the predicative use. This approach indicates that French [N 1 N 2 ] subordinatives, and especially the subclass of formations with a qualifying N 2 , are in reality closely related to [N A] or [A N] formations. As we have seen in Section 3.2, Scalise/Bisetto (2009) merge [N 1 N 2 ] appositives and [N A]/[A N] attributives within the class of ATAP compounds because the modifier in both cases expresses a qualifying property of the head noun. We can therefore conclude that their classification for these types of formations is highly insightful. However, what is still missing in this approach is the fact that this ATAP class contains not only pure (morphological) compounds, but also hybrid constructs with both morphological and syntactic properties.

Conclusion
Compared with Germanic languages, it turns out to be very difficult to delineate French compounds from syntactic multi-word units. In the first part of this contribution, we outlined three different approaches dealing with compounding in the French tradition: non-restrictive, scalar and restrictive (lexicalist). Although we believe morphological formations should be distinguished from syntactic formations, it is insightful to highlight their shared potential for expressing the same denominative functions. We therefore added a fourth approach: we believe a constructionist, non-modular approach to the language system provides a more appropriate account. From this perspective, both compounds and phrasal structures with a naming function can act as conventionalized form-meaning pairings or 'constructions' and we should accept the existence of what Booij (2010: 190) calls 'lexical phrasal constructions', namely phrasal constructions that are stored in the (mental) lexicon.
Another advantage of this constructionist approach is that it can deal with structurally ambiguous formations, such as [N 1 N 2 ] structures with a qualifying N 2 . As shown throughout this paper, these sequences are particularly difficult to deal with in a modular approach because, on the one hand, they formally and semantically resemble [N 1 N 2 ] (subordinative) compounds, but, on the other hand, they allow syntactic operations to a greater or lesser extent. In a conception of language as a constructionist network, these hybrid formations can be fruitfully accounted for by the mechanism of 'multiple inheritance'. Following this process, we have argued that the hybrid properties of French qualifying [N 1 N 2 ] sequences result from the inheritance of properties from both a morphological and a syntactic parent construction.