3.1. Japanese

Japanese PHs have been extensively examined (Uemura 1996; Kitano 1999; Hamaguchi 2001; Hosoda 2002; Hayashi 2003; Hayashi & Yoon 2006; Suga 2018; Seraku 2022a; 2022b), but what has received little attention is that a target form may be a syntactically negated predicate. Consider (9) (see Seraku et al. 2021: 83–84 for further data).

The NPI anmari occurs without a negator.³ The acceptability of (9) suggests that the adjectival PH aredesu substitutes for a negated form such as tokuija-nai-desu ‘good.at-neg-hon’ (= ‘not good at’). If aredesu is replaced with the non-negated predicate tokuidesu ‘good at’, the NPI is no longer licensed, as can be seen in (10).

Unlike the adjectival PH aredesu, the verbal PH aresuru cannot license NPIs without a negator. This is shown in (11), where areshi- is the infinitive form of aresuru.

Example (11) is grammatical if areshi-ta-yo ‘ph-pst-fp’ is replaced with its negated analogue: areshi-nakat-ta-yo ‘ph-neg-pst-fp’.

As mentioned in Section 2, previous studies generally assume that a PH is meta-linguistic in that it substitutes for a linguistic expression. This assumption is corroborated by (9); in this example, what the PH substitutes for is a negated form such as ‘not good at’, where the negator grammatically licenses the NPI anmari.

3.2. French

Let us turn to French PHs. The focus of our description lies in what we may call a “gender mismatch”. We say a gender mismatch occurs when a PH-involving sentence S is grammatical but the replacement of the PH with its target form causes inconsistency in gender agreement and renders S ungrammatical. We will illustrate this with French truc and machin.

French PHs are briefly noted in a few studies (Palacios Martínez & Núñez Pertejo 2015: 428, Vogel 2020), but they offer no concrete data. We thus start with basic examples.

Emma fails to utter the target word détergent ‘detergent’ on the spot and fills in its slot with truc or machin. When a speaker refers to tangible objects, as in (12), truc and machin are both permitted. These PHs may also be deployed to refer to intangible objects, as in (13).

When a speaker refers to persons, machin is preferred to truc and used without an article.

In (12)–(13), the use of du, a contraction of de ‘of’ and the masculine le ‘the’, indicates that the gender of truc and machin is masculine. This is consistent with the gender of their target words: détergent in (12) and COVID-19 in (13).⁴ In (15), a gender mismatch occurs.

The target form here is vis ‘screw’. This is particularly evident when Emma suddenly recalls the word vis and utters it later, as shown in parentheses. The gender of vis is feminine, and it is incompatible with the gender of truc/machin. Thus, the replacement of truc/machin with vis results in ungrammaticality, as can be seen in (16)a, which sharply contrasts with (16)b.

Emma could produce (15) appropriately even when she does not recall the target form itself but remembers its gender (i.e. feminine) at the time of uttering masculine d’un and gros.

This subsection has described the French nominal PHs turc and machin with special reference to the gender-mismatch phenomena. As will be pointed out in the next subsection, another type of gender mismatch is observed in German.

3.3. German

German PHs have been surveyed from a morphological point of view (Vogel 2020) and from the perspective of language education (Cutting 2015; 2019), but their grammatical properties have been largely untouched. An exception is Sailer & Dörne (2021), who describe and model morphosyntactic issues in German PHs; still, they do not discuss gender-mismatch phenomena. In what follows, we consider the nominal PHs Dings, Dingens, Dingsda, and Dingsbums, all of which derived from Ding ‘thing’ (Vogel 2020: 373). For the sake of presentation, Dingens is chosen as a representative form in the examples to be presented below.

In German, when a noun combines with an article and adjectives, they agree in terms of case, number, and gender (i.e. masculine, feminine, neuter). In the case of Dingens, however, its form remains the same irrespective of the case, number, and gender of its target form. In (17), Dingens substitutes for Schraube ‘screw’, a feminine noun.

From the hearer’s point of view, the feminine forms eine ‘a’ and große ‘big’ perform a filtering function: the hearer, who hears the speaker uttering these feminine forms, only searches the context for entities that can be denoted by feminine nouns.

In the context of (17), the speaker cannot use the masculine forms for ‘a’ and ‘big’ since she is aware at the time of speech that the target form is feminine.

Note that (18) itself is grammatical and may be uttered appropriately in different contexts, e.g., when the speaker is referring to an entity denoted by a masculine noun such as Schraubendreher ‘screwdriver’. This is why (18) is marked with ♯, rather than *.

Then, what form would a speaker use for an article and adjectives when she cannot recall the gender of a target form? In such cases, a neuter form may be used, as shown in (19).

Unlike (17), the speaker in (19) cannot select the feminine forms because she cannot recall the gender of a target form. As shown in parentheses in (19), the use of the neuter forms ein and großes is appropriate even if the non-neuter target form Schraube is uttered subsequently. This is where a gender mismatch occurs. If Dingens in (19) is replaced with its target form Schraube, the sentence becomes ungrammatical, as shown in (20)a, which is in stark contrast with (20)b.

The mismatch data in German differ from those in French. In French, truc and machin are masculine, and a co-occurring article and adjectives must also be in the masculine form. In German, Dingens (and the related forms such as Dings and Dingsda) do not encode information about gender, and it is therefore compatible with an article and adjectives of any gender. When a speaker cannot recall the gender of a target form, they may exploit the neuter form of an article or adjectives.

3.4. Related work

We have described Japanese, French, and German PHs, and have revealed the gender-mismatch phenomena in the latter two languages with some cross-language differences. We now make use of these empirical findings to review previous studies.

Whilst there is a wealth of descriptive and functional studies on PHs (see Section 2), few serious attempts have been made to formally analyse them. Cheung (2015) analyses Mandarin wh-derived PHs based on the standard syntactic/semantic machinery (Heim & Kratzer 1998) with additional meta-linguistic mechanisms (e.g. the type-shifting function shift, which takes a linguistic object and returns its semantic representation). Sudo (2008; 2013), who analyses Japanese wh-derived PHs, also augments the standard framework with meta-linguistic tools. Sailer & Dörne (2021), who investigate several German PHs, develop another type of meta-linguistic account in Head-driven Phrase Structure Grammar (e.g. Sag et al. 2003); their account is seen as meta-linguistic in that a target form is listed as a morphological daughter of the sort placeholder.expression-complex.⁵

A challenge for these meta-linguistic accounts is posed by examples like (21) (= (7)).

A possible interpretation of (21) is that at the time of uttering whatchamacallit, a speaker only entertains some vague concept such as ‘a tool for opening a box’ and does not have in mind any specific target form.

In addition to the above general issue, we specifically discuss the previous accounts in mainstream generative grammar: Cheung (2015) and Sudo (2008; 2013). We focus on Cheung (2015) because Cheung presents his account in detail and illustrates it with concrete examples. Cheung analyses wh-derived PHs in Mandarin. It is cross-linguistically observed that wh-words are recruited as PHs (Seraku et al. 2022). In Mandarin, various wh-words, together with the distal demonstrative na ‘that’ and the general classifier ge, serve as a PH. In (22), the speaker utters the PH na ge shei to substitute for Zhangsan, a personal name that the speaker is unable or unwilling to produce.

The structure for (22) is presented in (23), where a linguistic expression is notated in italics (e.g. Zhangsan), and its denotation in italics with a prime symbol (e.g. Zhangsan′).

Cheung enriches the set of semantic types with the type for linguistic expressions, u. Then, the wh-word shei is treated as a predicate of type <u, t>, λx: ⟦x⟧ is a human under discussion. This function takes a linguistic expression x of type u and returns the statement that ⟦x⟧, namely the denotation of x, is a human under discussion. Na-ge is treated as a variant of the Russellian definite operator, taking a predicate of type <u, t> and returning the unique x such that x satisfies the description. The relevant semantic composition is shown in (24).

In (22), ⟦na-ge⟧(⟦shei⟧) outputs the linguistic expression Zhangsan, and the null operator shift turns Zhangsan into its denotation Zhangsan′. Finally, Zhangsan′ combines with the denotation of VP, giving rise to the semantic representation of the whole clause: left′(Zhangsan′).

A potential problem with this type of analysis is how it would handle our data in Japanese, French, and German. One could argue that NPI-licensing may be ensured by postulating a target form involving a negator (or a neg feature) in a syntactic structure so that it c-commands the NPI before the semantic compositions begin. This analysis, however, is not straightforwardly applicable to the gender-mismatch data since a stipulation would have to be made that when a target form is introduced to a syntactic structure, its gender information (e.g. a gender feature) is not posited (or somehow removed after the target form is introduced). This, then, leads to the following ambivalent situation: a negator (or a neg feature) must be present to license an NPI, whilst a gender feature must be absent to prevent gender inconsistency.

We do not claim that this issue cannot be overcome within the standard syntactic/semantic framework. Given the vast amount of extant syntactic/semantic work in mainstream generative grammar, one could make use of some previous mechanism to account for the data. For instance, it might be possible to complement Cheung’s analysis with a device for gender mismatches in other phenomena such as ellipsis (Merchant 2014). Our claim, then, is modest; we will show that if we adopt the Dynamic Syntax view of grammar, we can readily and naturally account for a wide range of cases, including examples like (21), the NPI data, and the gender-mismatch phenomena, and as a bonus of this approach, we can extend the analysis to dialogic properties of PHs.

It should be emphasised that we will not propose an entirely new approach; our account rests on the conventional machinery of Dynamic Syntax. What we newly propose is the general parse-mechanism for PHs (see the end of Section 5.1), and even this new addition is based on the established apparatus of the framework: (i) the postulation of a meta-variable and its subsequent update, (ii) structural underspecification, and (iii) LINK relations. Therefore, our account, to be developed in the rest of this article, should not be taken as an analysis specifically tailored for PHs that relies on new theoretical tools.

4.1. Basic architecture

In DS, the dynamicity, i.e. constant change, of structure building is explicitly modelled in the time-linearity of parsing. For example, the parse of (25) builds up a conceptual representation incrementally and monotonically, as displayed in (26).

The structures in (26) are conceptual; each node, once fully developed, accommodates semantic content (e.g. run′) and its semantic type (e.g. e→t).⁶ Each tree in (26), except for the final state, forms a partial structure in that it contains “requirements” such as ?t, which indicates that this node should be eventually annotated (“decorated”) with semantic content of type t. The tree transitions model monotonic information growth in that once a structure is built, it will not be destroyed and the information on tree nodes will be preserved.

A tree is gradually updated by lexical, computational, and pragmatic actions. We illustrate lexical and computational actions by explicating the tree transitions in (26). (Pragmatic actions will be illustrated later; see (36) below.) The initial state in (26)a is repeated here as (27) with a symbol ♢, called a pointer. In each state, the pointer ♢ marks the node under development, indicating where the next step in the parse process will take place.

The first word to be parsed is Ken, whose lexical entry is formulated as in (28).

The IF-line sets an input condition. In (28), it states that the current node (marked with ♢) should be decorated with ?e. If this condition is met, the parser looks at the THEN-line, which instructs the parser to decorate the node with the annotation Ken′ : e, an abbreviation standing here for an individual (indicated by the type annotation e) represented by the constant Ken′, which picks out the contextually accessible individual named Ken. Otherwise, the parser looks at the ELSE-line, which aborts the parse. Notice that the tree state in (27) does not contain a node decorated with ?e. The parser may predict such a node by creating it through the computational set of actions (a so-called “macro”) of Local *Adjunction.

This macro of actions introduces an “unfixed” node, a node whose structural position in the tree is not yet determined and will be resolved later. This structural underspecification is visually shown by a dashed line in (29). The unfixed node introduced by Local *Adjunction must be resolved within a local propositional structure, but at the stage of (29), the parser is not sure about the exact position where the unfixed node will be resolved.⁷ Now that the tree in (29) has a node decorated with ?e, the action in the THEN-line of (28) is licensed to be executed with the result shown in (30).

The lexical macro encoded in the nominative particle -ga then specifies that the mother of the unfixed node is the root node (Cann et al. 2005: 254). As a result, the unfixed node is identified as the subject node, as in (31).

What comes next is the intransitive verb hashit- ‘run’, which constructs a predicate node and decorates it with the content–type pair run′ : e→t.

The current tree licenses functional application and type deduction. This is formulated as the computational macro of Elimination.

Setting aside tense and aspect (see Cann 2011), the tree in (33) is in the final state, where the root node represents the truth-conditional content of (25): ‘Ken is running.’

In the above illustration, one may have had the impression that hashit- ‘run’ only creates a predicate node, but this is technically not true. To clarify this point, consider (34).

Japanese is a pro-drop language, and argument NPs may drop as long as they are contextually recoverable. To model this pro-drop phenomenon, DS assumes that verbs and adjectives in Japanese encode the actions to construct a schematic propositional structure. Thus, the parse of hashit- in (34) updates the initial state in (27) into (35).

The subject node in (35) is decorated with a meta-variable U. Meta-variables in DS are values that do not belong to the object language in which the semantics of natural languages is expressed. Instead, they serve as a temporary label until an appropriate value from the object language becomes available. Such values can either be provided by processing further linguistic input (for example, in cases of cataphora) or through a pragmatic process of context search for accessible semantic values. Here, the meta-variable U is pragmatically assigned a value from the accessible context (the “common ground” (Clark 1996)), such as Ken′.⁸ This is an instance of the pragmatic macro called Substitution.

In (31) earlier, the subject node has been introduced before hashit- ‘run’ is parsed. This does not create a duplication problem due to the “model-theoretic” formulation of DS as a grammar formalism (as opposed to a “generative-enumerative” formulation; see, e.g., Pullum & Scholz 2001). As long as the annotations on a node are compatible, the node can be updated with more informative values. Here the subject node created by hashit-, which is decorated with a meta-variable, harmlessly collapses with the pre-existing subject node as they are both defined as occupying the same position as the left daughter of the root node of a binary tree. The decorations on the nodes then collapse too, since the term Ken′ can update the underspecified meta-variable U with a fully specified value. The entry for hashit- is explicated in (37).

In (37), make/go(↓₀) creates an argument-daughter node and moves the pointer ♢ to this node; go(↑₀) moves the pointer ♢ back to the mother node; make/go(↓₁) creates a functor-daughter node and moves the pointer ♢ to this node.

Finally, we introduce the mechanism of LINK. So far, a single tree has been built up at a time, but DS allows a pair of trees to be constructed in tandem via a formal relation called LINK. The LINK device has been applied to various phenomena such as relatives, left dislocations, co-ordinations, and topic constructions (Cann et al. 2005: chs. 3–6). For example, the parse of the topicalised element Ken-wa in (38) is modelled as in (39).

A LINK relation starts from a tree that has been constructed to another tree that will be built up with the first tree as context. In (39), the LINK relation starts from the type-e node to the type-t-requiring node, and the latter will be developed by the parse of the rest of the sentence: yasashii ‘kind’. The parse of yasashii builds a schematic propositional structure with a subject meta-variable; with the initial tree (i.e. the node decorated with Ken′) as context, this meta-variable pragmatically receives the value Ken′ (Substitution). After functional application and type deduction are performed (Elimination), the tree is updated to (40), where the paired trees share the term Ken′.⁹

In (40), the LINK relation associates a type-e node to a type-t node, but LINK relations are free from type restrictions. In Section 5.1, we will discuss a LINK relation that relates a type-e node to another type-e node.

This subsection has outlined a DS grammar for Japanese. For more details, the reader is referred to Cann et al. (2005: ch. 6), Kempson & Kiaer (2010), and Seraku (2013).

4.2. Crosslinguistic universals and variations

In DS, crosslinguistic universals are mostly formulated as the set of computational macros (e.g. Elimination), whilst crosslinguistic variations are formulated as the set of lexical macros. The reason why we added the proviso mostly with regard to crosslinguistic universals is that not all computational macros are available across languages. As mentioned in Section 4.1, in Japanese, a propositional structure is lexically built up. This is true of other pro-drop languages such as Greek (Chatzikyriakidis 2012: 650) and Spanish (Bouzouita 2008: 249). By contrast, in French, a non-pro-drop language, a propositional structure is constructed by the computational macros called Introduction/Prediction (Cann et al. 2005: 44). For example, before initiating the processing of each word in (41), the parser performs Introduction/Prediction to construct a tree structure appropriately prepared to accommodate a subject–predicate structure in the semantic object language, as shown in (42).

Now that a skeletal structure is established, Ken decorates the current node with Ken′ : e, and court annotates the predicate node with the decoration run′ : e→t. The tree in (43) will reach the final state with the application of Elimination.

Since a propositional structure is lexically constructed in pro-drop languages, the computational macros of Introduction/Prediction are not relevant to such languages.

Another crosslinguistic variation pertinent to the present study is concerned with gender agreement. Consider (44).

After the Introduction/Prediction macros are executed, Leo decorates the subject node. Since utilise ‘use’ is a transitive verb, it introduces a more complex structure than an intransitive on the binary tree: the “?e→t”-node is expanded with two daughters whose combination will derive the value needed to combine with the subject. So, an object and another predicate node are introduced by the actions of the verb with the pointer ♢ left at the object node anticipating the actions of the lexical entries that will introduce a semantic value on this node. The next word to be processed is the masculine article un, which introduces the requirement ?MAS at the object node, standing for a constraint that the semantic value to occupy this node needs to be contributed by a lexical entry that can satisfy the morphological feature ‘masculine’ (MAS):¹⁰

The current node is then decorated by the lexical entry for détergent, which introduces not only the usual content–type pair detergent′ : e but also contributes the feature MAS, which allows the requirement ?MAS to be satisfied and removed.

If the feminine article une were parsed in lieu of un in (44), the object node would be decorated with ?FEM(inine), but this requirement cannot be met by the masculine noun détergent. This is tantamount to what is descriptively called an agreement failure.

We have only sketched a DS grammar of French, but this suffices for our analysis of PHs. It is left for future research to develop a full-fledged grammar of French.¹¹

5.1. Proposal

As has been illustrated in Sections 2–3, PHs can be taken as referring expressions, albeit possibly with meta-linguistic reference (for further such cases and their treatment in DS, see, e.g., Gregoromichelaki 2017). For this reason, we hold that PHs also introduce a meta-variable in the developing conceptual representation. More specifically, we assign whatchamacallit the entry in (47), where PH_{whatchamacallit} is a meta-variable introduced by whatchamacallit.¹²

This lexical entry is illustrated with (48) (= (1)).

The parse of I need a whatchamacallit in (48) constructs the tree in (49). (The meta-variable of the first-person pronoun I has been contextually updated as Naomi′.)

Given the DS machinery outlined in Section 4.1, there are two ways in which PH_{whatchamacallit} may be resolved: parsing the target form a pair of scissors on an unfixed node or on a LINKed node.

The first strategy is to introduce an unfixed node and parse a pair of scissors there. This process is illustrated in (50). In English, an unfixed node is introduced by *Adjunction (Cann et al. 2005: 61; see also footnote 7), but the unfixed node in (50) is different from an unfixed node introduced by *Adjunction.¹³

The parser may resolve the unfixed node by running Merge. This computational macro merges an unfixed node of semantic type α with a fixed node of the same semantic type α, to the effect that their node descriptions are unified (Cann et al. 2005: 65). In (51), the unfixed node has been resolved as the object node through the application of Merge, and the meta-variable PH_{whatchamacallit} has received the value scissors′.

This tree leads to the final state in (52) with the applications of Elimination.

The second strategy is to initiate a LINK relation from the object node in (49) to a type-e-requiring node and parse a pair of scissors on this LINKed node. This is illustrated in (53).

The LINK relation is then “evaluated” (see, e.g., Cann et al. 2005: 113): the content scissors′ is reflected in the object node, and PH_{whatchamacallit} is updated to scissors′. After Elimination is executed, the final state in (54) emerges.

Though this is not evident in (52) and (54), the two parsing strategies may output distinct node descriptions. In the strategy involving an unfixed node, all node descriptions of an unfixed node (including a negation feature or a gender feature, if any) are reflected in a fixed node. In the strategy involving a LINKed node, only the semantic content on a LINKed node is passed on to the node from which this LINK relation is initiated. This difference in node descriptions is the key to explaining the NPI data in Japanese and the gender-mismatch data in French and German, as will be shown in Sections 5.2–5.4.

In (48), a PH is immediately followed by its target form, but our account also deals with cases where a PH and its target form occur across sentences. In such cases, the parser may put on hold the tree update for the PH-involving string and re-initiate it later once it encounters the target form. To this end, the parser needs to keep track of previous tree-updates; this is modelled in terms of a Directed Acyclic Graph (Sato 2011), as demonstrated in, e.g., Hough (2015: Sections 6.4–6.5).

Further, our account handles cases where a target form is absent on the surface. Recall that a DS structure is updated by both linguistic (i.e. lexical or computational macros) and non-linguistic actions (i.e. pragmatic macros). This means that the DS system is not “encapsulated” in the sense that the parser may employ non-linguistic information at any stage during structure building. Consider (55) (= (7)).

If the hearer presumes that Naomi refers to a pair of scissors, the parser may execute the actions encoded in the expression a pair of scissors by introducing an unfixed node or a LINKed node. In either case, the meta-variable provided by whatchamacallit is resolved. It is also possible that the hearer only has in mind a vague concept such as ‘a tool for opening a box’ without thinking of a specific linguistic item. In this case, the parser can retrieve and introduce this “ad hoc” concept (Carston 2002: ch. 5) since such concepts are, in fact, the eventual inhabitants of tree nodes in DS.¹⁴ This process updates (49) into (56), where the ad hoc concept ‘a tool for opening a box’ is notated as tool′.

The tree in (56) is mapped onto the final state in (57) with the applications of Elimination.

The proposal made in this subsection is summarised in the bullet points below. These will be applied to various examples in Japanese, French, German, and some other languages in the ensuing subsections.

• A PH introduces a meta-variable PH with appropriate semantic restrictions.¹⁵

• When a target form is present in the string, the parser may resolve a PH meta-variable by parsing the target form on an unfixed node or a LINKed node.

• When a target form is absent in the string, the parser may (i) run the actions encoded in a contextually available target form on an unfixed node or a LINKed node or (ii) construct an ad hoc concept and replace a PH meta-variable with it.

5.2. Japanese

The entry for the nominal PH are is specified in (58).

We do not formalise the semantic restrictions on the use of are; hence, the informal notation PH_are. The lexical entry in (58) is illustrated with (59) (= (4)).

Setting aside desukara ‘so’ and bunpootekina ‘grammatical’,¹⁶ the parse of are-ga yields the tree in (60) through a step of Local *Adjunction.

The target form kinoo ‘function’ is then parsed on an unfixed node or a LINKed node. In either case, PH_are is resolved as function′, as shown in (61).¹⁷

The rest of the process is as outlined in Section 4.1. The parse of chigau- ‘differ’ constructs a propositional skeleton, and Elimination performs functional application and type deduction.

The treatment of predicative PHs is more complex because Japanese verbs and adjectives are assumed to build a propositional structure. Consider (62)–(63).

For instance, suppose that aredesu is parsed sentence-initially. (Recall that Japanese is a pro-drop language.) The lexical actions specified in (63) construct the tree in (64).

As has been explained with regard to (37), make/go(↓₀) creates an argument-daughter node and moves the pointer ♢ to this node; put(U : e) decorates this node with U : e; go(↑₀) moves the pointer ♢ to the mother node; make/go(↓₁) creates a functor-daughter node and moves the pointer ♢ to this node; put(PH_aredesu : e→t) decorates this node with PH_aredesu : e→t.

The entry for aredesu is more concretely illustrated with (65).

First, Local *Adjunction creates an unfixed node, which is decorated by Naomi and resolved as the subject node by the nominative -ga. Second, aredesu constructs a propositional template and posits PH_aredesu : e→t on the predicate node. The subject meta-variable in this propositional template harmlessly collapses with the content of the pre-existing subject node, Naomi′ (see the discussion about (36)).

The parser then processes the target form kireidesu ‘beautiful’ on an unfixed node or a LINKed node, and PH_aredesu is updated to the fully specified content beautiful′.

The entry for aredesu in (63) accounts for the NPI data in (68) (= (9)).

Cann et al. (2005: 216) assign the NPI any an entry whose input condition is that the root node should be decorated with the feature NEG. For Japanese anmari, we assume that the root node is decorated with the requirement ?NEG because in Japanese, a negator comes towards the end of a clause. Thus, the parse of anmari in (68) outputs (69).

Then, after Local *Adjunction induces a type-e-requiring unfixed node, shaberu-no ‘talking’ decorates this unfixed node and -ga resolves it as the object node, as shown in (70).¹⁸ In this partial structure, the object node is designated as the one whose mother is the functor-daughter of the root node, as decorated with ?e→t.

Now, the adjectival PH aredesu constructs a propositional skeleton and decorates the predicate node with PH_aredesu : e→t.

In (68), a target form is absent on the surface, but given that a DS grammar is non-encapsulated, the hearer may be able to recover the form tokuija-nai-desu ‘good.at-neg-hon’ as a potential target form. The parser may then execute the associated actions in the lexical entry, creating a propositional template. If the parser runs these actions on an unfixed node, the tree in (72) arises. In (72), a type-t-requiring unfixed node has been introduced because tokuija-nai-desu is a predicate and can be parsed only on a type-t-requiring node. The parse of tokuija-nai-desu has created a schematic propositional structure, where the top node, marked with the pointer ♢, is decorated with NEG due to the negator -nai.

In the current propositional structure, the meta-variables V and W are contextually updated to TVP′ (denoting the TV personality in question) and talking′, respectively (Substitution), and functional application and type deduction are then performed (Elimination).

The tree in (73) licenses Merge, which unifies the information on the type-t unfixed node with that on the type-t-requiring fixed node. Consequently, U is updated to TVP′, and PH_aredesu to good_at′(talking′). And crucially, the requirement ?NEG is satisfied by NEG.

The NPI data are thus properly modelled based on the unfixed-node approach. On the other hand, if a LINK relation is launched, its evaluation process reflects only the semantic content (but not the feature NEG) in the main tree, and the requirement ?NEG remains unfulfilled. Therefore, the NPI data are accounted for only through the introduction of an unfixed node.

In closing, let us consider why the verbal PH aresuru cannot substitute for a negated expression, as illustrated in (75) (= (11)).

In Japanese, when an adjective is negated, the negator -nai comes before -desu, as in kireija-nai-desu ‘beautiful-neg-hon’. Note that -desu is a constituting part of the PH aredesu. That is, the negator -nai is included in the range of sequence to be substituted for by the PH aredesu. By contrast, when a verb is negated, -nai comes after the verb, as in ne-nai ‘sleep-neg’. Since the verbal part of aresuru is -suru, the negator -nai is not included in the range of sequence to be substituted for by the PH aresuru. This is why aresuru cannot substitute for a negated verb. In fact, if -nai is put after the verbal part -suru in aresuru, the extended PH form areshi-nai ‘ph-neg’ substitutes for a negated verb, as shown in (76), where areshi-nai is realised as areshi-nakat before the past tense marker -ta.

5.3. French

French truc and machin are assigned the entries in (77)–(78).

As pointed out in Section 3.2, truc and machin exhibit semantic idiosyncrasies. For instance, truc is usually not used for humans (see (14)). Again, these idiosyncrasies are not addressed here. Of particular note in (77)–(78) is the THEN-line, where the feature MAS is introduced, reflecting the masculine gender of truc and machin.

Consider (79) (= (12)).

After the parse of J’ai besoin du truc, the tree in (80) emerges.

The current node is decorated with MAS, which satisfies the requirement ?MAS, posited by du. If the parser runs the actions encoded in the target form détergent on an unfixed node, the tree in (80) is updated into (81).

This tree licenses Merge, unifying the information on the unfixed node with that on the object node. PH_truc is then updated to detergent′, and MAS is reflected in the object node. Since the object node has already been decorated with MAS, this additional introduction of MAS does not alter the node description.

Alternatively, at the stage of (80), one could induce a LINKed node to run the actions encoded in détergent. The evaluation of this LINK relation reflects the content detergent′ (but not the feature MAS) in the object node, and PH_truc is resolved as detergent′. In either case, example (79) is appropriately modelled.

The present account naturally deals with the mismatch data in (83) (= (15)).

The parse of J′ai besoin d′un truc (excluding plus gros ‘bigger’) outputs the tree in (84).

Suppose that the parser runs the actions encoded in the target form vis ‘screw’ on a LINKed node. Since the evaluation of the LINK relation reflects the content screw′, but not the feature FEM, in the object node, PH_truc is resolved as screw′ without inconsistency.

By contrast, if the parser runs the actions encoded in vis ‘screw’ on an unfixed node, the tree in (86) is constructed.

After the application of Merge, the object node is decorated with not only screw′ but also FEM. As a result, feature inconsistency is detected, as highlighted in grey in (87). The current parse process, therefore, fails to derive a well-formed tree.

The upshot is that the gender-mismatch data are handled only through the introduction of a LINK relation.

As will be argued in the next subsection, the analysis of the French mismatch phenomena straightforwardly carries over to the German examples, but with a twist due to the genderless PH forms.

5.4. German

The lexical entry for German Dingens is formulated in (88).

Unlike French truc and machin, Dingens is genderless. This is why the action in the THEN-line only introduces the usual content–type pair.

Let us illustrate this entry with (89) (= (17)).

The parse of (89), excluding große ‘big’, engenders the tree in (90).

The presence of ?FEM is due to the feminine article eine ‘a’. Unlike French truc and machin, Dingens encodes no gender information; this is why ?FEM remains unfulfilled at this stage. If the parser executes the actions encoded in the target form Schraube ‘screw’ on an unfixed node and this unfixed node unifies with the object node (Merge), PH_Dingens is resolved as screw′ and FEM is introduced. The latter satisfies ?FEM.

Alternatively, the parser could run the actions encoded in Schraube on a LINKed node. In this case, however, the evaluation of the LINK relation only reflects the content screw′ (but not the feature FEM) in the object node, and ?FEM remains unfulfilled.

We turn to our crucial example (92) (= (19)).

We assume that the neuter forms such as ein ‘a’ and großes ‘big’ introduce the NEUT feature.¹⁹ With this assumption, the parse of (92), excluding großes ‘big’, builds the tree in (93).

If the parser runs the actions encoded in the target form Schraube ‘screw’ on a LINKed node, the evaluation of the LINK relation reflects screw′ (but not FEM) in the object node. Thus, no inconsistency is detected in the node description. By contrast, if the parser executes the actions encoded in Schraube on an unfixed node and this unfixed node unifies with the object node, the object node is newly decorated with screw′ and FEM. This leads to inconsistency because the object node ends up accommodating NEUT and FEM.

In sum, only the LINK-based parsing derives a well-formed tree for the gender-mismatch data in French and German. Although the analysis of French PHs differs from that of German PHs in terms of gender specifications of PH forms, they both rely on the central DS machinery: the introduction of a meta-variable and its subsequent update through a LINK relation.²⁰

5.5. Further crosslinguistic applications

Up until now, our discussion has centred around English, Japanese, French, and German, but PHs are widely observed in the languages of the world, such as the nominal PH ceke (< ‘that’) in Korean (Hayashi & Yoon 2006: 492), the verbal PH cosare (< cosa ‘thing’) in Italian (Podlesskaya 2010: 15), and the adjectival PH ini (< ‘this’) in Indonesian (Wouk 2005: 246). In our account, these PHs encode a meta-variable of appropriate semantic type, such as PH_ini : e→t for the adjectival PH ini in Indonesian.

A noteworthy case is kua in Ilocano (Austronesian), which substitutes for a wide array of elements such as a noun, a verb, an adjective, an adverb, and a clause (Rubino 1996: 657–659). In (94), kua substitutes for a clause.

A possible solution to this versatile PH is to assign it an entry with type underspecification.

When kua stands for a clause, as in (94), the decoration PH_kua : t is selected. When kua stands for an adverb, we simply follow Marten’s (2002) type-e analysis. (Alternatively, we could adopt a situation-term analysis of adverbs; see Gregoromichelaki 2006.)

Another challenge to our account is posed by (96).

The speaker cannot fully recall the proper name Ao-lang-de, a Mandarin translation of Hollande, and replaces its intermediate syllable with the PH shenme (< ‘what’). Since DS tree transitions are triggered by morphemes, not by phonemes, (96) is problematic (see Sailer & Dörne 2021: 143 for similar data in German). In order to accommodate such examples, we need to establish the phonology–grammar interface in DS, which goes far beyond the scope of the present work (see Cann 2021 for preliminary remarks).

6.1. PHs in dialogue

Since PHs typically occur in talk-in-interaction, they are often used as a grammatical resource for achieving a speaker–hearer joint action: a speaker may produce a PH and invite a hearer to search for a target word collaboratively. Such dialogic functions are associated with the uses of PHs in various languages, such as Japanese are (Kitano 1999: 388), Korean kuke (Kim & Suh 2002: 193–197), Romani kako (Matras 1998: 422), and Estonian see (Keevallik 2010: 157). A case pertaining to Japanese are is presented in (97).

In line 1, speaker N utters are presumably because the title of a novel temporarily eludes him. This instance of are is followed by two interjective fillers nanka ‘like’ and ano ‘um’, and then speaker A provides a candidate for the referent of are. Though A’s guess turns out to be wrong (see line 4), what is important here is that N’s use of are may be seen as an invitation to A for joining N’s word-search. Also, line 3 reveals a further characteristic property of dialogue: a participant’s contribution may be nonsentential.

As another example, consider Estonian see in (98).

This example uncovers another feature of dialogue. In line 1, speaker M produces the PH see, and speaker L (who was a hearer in line 1) offers a candidate for the referent of see in line 3, which is immediately acknowledged in line 4 by speaker M (who was a hearer in line 3). That is, a speaker–hearer switch takes place seamlessly and fluently during an ongoing interaction. Further, as shown in lines 1–3, a speaker–hearer switch may occur mid-sententially.

These dialogic phenomena (i.e. the use of nonsententials, speaker–hearer joint actions, mid-sentential speaker–hearer switches) have been generally seen as something in the realm of language performance, except for some cases of nonsententials (Barton 1990; Merchant 2004; Progovac et al. 2006). On the other hand, these dialogic features fall in the scope of DS, which analyses a string of words uttered and interpreted time-linearly in context. In the next subsection, we will extend our account of PHs by applying the DS treatment of dialogue issues to the PH–dialogue phenomena (see Purver et al. 2014 for speaker–hearer joint actions and speaker–hearer switches and Gregoromichelaki et al. 2020 for nonsententials).

6.2. Analysis

Our presentation of DS has so far concentrated on parsing, which licenses the incremental construal of word-sequences in comprehension. However, DS models both parsing (comprehension) and generation (production) simultaneously and this provides the basis for accounting for dialogue phenomena like “split utterances” (see, e.g., Gregoromichelaki & Kempson 2013). In DS, parsing and generation use the same predictive and incremental mechanisms which is the reason for arguing that the grammar itself has the role of establishing and sustaining coordination between speaker and hearer in dialogue. The only difference between comprehension and production is that in the latter, the speaker is manipulating a “goal-tree” in addition to the common parse tree. The goal tree is possibly a partial tree (modelling the incrementality of production) but at least one processing step ahead of the parse tree. The goal tree represents the conceptual content the speaker plans to linearise verbally and it serves as a self-monitoring device in that the speaker constantly checks each tree state to ascertain that it “subsumes” the goal tree (“subsumption check”; see, e.g., Purver et al. 2006), with implicit or explicit self-repair ensuing otherwise. A tree is said to subsume a goal tree if it does not contain any nodes or node decorations absent in the goal tree. For illustration, Ken runs is incrementally parsed/generated as depicted in (99) (see also Purver & Otsuka 2003: 81).

As can be seen in (99), the same set of mechanisms underlies parsing and generation, with the sole difference being that in the case of generation, each tree state undergoes a subsumption check against a goal tree and unsuccessful candidates are removed and not pursued further.

Let us now analyse (97). Prior to line 3, N and A have constructed the tree in (100), where PolQ is a feature introduced by the interrogative form of polar questions.

At this stage, the tree update is incomplete in that PH_are is not resolved and that the root node is decorated with ?t. The goal-tree, therefore, is the one identical to (100) except that PH_are has been assigned a value and Elimination (i.e. functional application and type deduction) has been performed. In line 3, A updates (100) by resolving PH_are through the parse of Tokuzawa keeji shochoo ‘PD Chief Tokuzawa’ on an unfixed node or a LINKed node, as shown in (101).

Elimination then performs functional application and type deduction, yielding (102).

Since the tree in (102) subsumes the goal-tree mentioned above, this generation step (i.e. the production of Tokuzawa keeji shochoo) is licensed. Note that participants N and A share the tree in (100) and that either N or A could have extended it by producing words. This is why the change in speakership is seamless in the course of talk-in-interaction.

As stated above, a goal-tree is an extension of a current tree, and a current tree may be a partial one. In the Estonian example (98), a speaker–hearer switch takes place in line 3, but at this point, M’s utterance in line 1 is incomplete, as indicated by sound lengthening. In DS terms, M and L set out a goal-tree relative to the partial tree for M’s incomplete utterance. In line 3, L employs this partial tree as a basis for the next generation step (i.e. the production of ajalugu ‘history’).

The upshot is that the PH-involving dialogue phenomena follow from a unified treatment of incremental comprehension and production. One might object that the successful modelling of dialogue data does not favour our account over previous ones, given that it is generally held that dialogue phenomena fall outside the scope of standard competence grammars. Still, a growing body of work has been carried out to develop grammatical frameworks that can handle dialogue data (Lücking et al. 2021). Furthermore, it is in principle possible even for standard competence frameworks to incorporate parsing incrementality (Phillips 2003), thus blurring the lines between competence and performance. From this point of view, it seems justifiable to argue that the modelling of the dialogic properties of PHs is a welcome and confirming consequence of our analysis.