2.1 Basic assumptions

The dynamic syntax (DS) framework (Kempson, Meyer-viol & Gabbay Reference Kempson, Meyer-viol and Gabbay2001, Cann et al. Reference Cann, Kempson and Marten2005) is a processing-oriented framework. One of the basic assumptions behind DS is that natural language syntax can be seen as the progressive accumulation of transparent semantic representations with the upper goal being the construction of a logical propositional formula (a formula of type t). Such a process is driven by means of monotonic tree growth, representing the attempt to model the way information is processed in a time-linear, incremental, word-to-word manner. Dynamic syntax is a goal driven framework generally driven by means of requirements (requirements are denoted by the question mark ?). A well-formed sentence is one where all nodes have a formula (semantic content) and a type (semantic type) value and no outstanding requirements exist. In order to reach this well-formed structure, a sequence of partial-tree updating takes place. This allows the transition from a sole requirement for a proposition to a complete tree. The example below shows the starting point and the end result of parsing o Gianis agapai ti Maria ‘John loves Mary’:

In DS, a tree structure is well-formed iff no outstanding requirements remain. A string is then said to be grammatical iff there exists a tree update that leads to a well-formed structure. As can be seen in the above example, all nodes have formula (Fo) and type (Ty) values. Formula and type values combine via functional application and modus ponens respectively. The sign, called the pointer, shows the place in the tree where the parsing process is at a given point.Footnote ^[11] The whole system is underpinned by the logic of finite trees (LOFT) (Blackburn & Meyer-Viol Reference Blackburn and Meyer-Viol1994), an expressive language to talk about trees. The LOFT uses two basic arrow relations $\downarrow$ and $\uparrow$ , corresponding to the daughter and mother relation respectively. Left nodes are addressed as 0 nodes, whereas right nodes as 1 nodes. Conventionally, nodes on the left correspond to the argument nodes, i.e. the nodes in which the arguments will be represented, whereas the 1 nodes correspond to the functor nodes, i.e. the nodes in which all the various types of predicates will be represented. The rootnode, defined as the sole node that does not have a mother node, is given the treenode address 0. The example below illustrates the flexibility of LOFT by showing a binary tree where different nodes are addressed from the perspective of other nodes using treenode relations:

In the above tree, all nodes have a treenode address and a further statement identifying another node in the tree. For example, the statement $\langle \uparrow _{0}\rangle \langle \downarrow _{1}\rangle Tn(011)$ found in the 010 node reads as ‘you will find treenode 011 if you take a step across the 0 mother relation followed by a step across the 1 daughter relation’. Furthermore, the two Kleene operators * and + are used in combination with the basic tree relations, denoting the reflexive transitive and the transitive closure of the tree relation in each case. Thus, $\langle \downarrow ^{\ast }\rangle Tn(x)$ reads as ‘ $Tn(x)$ holds at the current node or at a node below the current one (of arbitrary depth)’, whereas $\langle \downarrow ^{+}\rangle Tn(x)$ reads as ‘ $Tn(x)$ holds on a node below the current one’.

The procedural part of the framework consists of two basic types of actions, lexical and computational. The first are language specific rules, roughly the lexical entries for individual words. An example of lexical rule is shown below:

The above example reads as follows: if you are at a node that has a type e requirement, then decorate this node with a type e value and a formula value representing the concept ‘Bill’. In any other case abort. In that sense, a proper noun like Bill in English will be able to get parsed as soon as a node has a requirement for a type e. This will allow a word like Bill to be parsed either as a subject or as an object in English.

As has been said, syntax is seen as the growth of transparent semantic representations. According to this view, morphosyntactic phenomena are also seen in this way. There are different ways in which morphosyntax operates in DS, ranging from providing purely procedural, declarative information as regards the parsing process and restrictions on tree updating to providing underspecified semantic representations or, in other cases, fully fledged semantic representations. For example, for case marking languages, the above lexical entry given for English would involve further information besides positing a formula value and a type $e$ . It will further involve a tree requirement that the node immediately dominating the current node is the predicate; it will impose in this way the requirement that the present node is the accusative node. Let us assume a language with constructive accusative. Then, the English entry would be transformed to something like the following:Footnote ^[12]

The second basic set of actions are computational rules. Computational rules are general computational devices, comprising the basic tree construction mechanism. They are assumed to be a closed set of rules universally available to every language. This closed set of rules are basically rules that help the parsing process to unfold. These involve pointer movement rules, rules that perform functional applications and modus ponens (for formulae and types respectively) or rules that get rid of requirements as soon as these are satisfied. A characteristic case of a computational rule is the rule of ELIMINATION. This rule performs functional application on formulae and modus ponens on types, in case both the argument and the function nodes bear complete formula and type values. Therefore, in the example below, the bottom nodes are combined by functional application and modus ponens in order to produce the result at the node above in the second tree (the node where the pointer, , is):Footnote ^[13]

2.2 Structural underspecification – unfixed nodes

Another basic assumption of DS is that languages are to a large extent underspecified regarding both content and structure. While content underspecification has been largely employed within the formal semantics literature of the past 30 years, no attempts to move underspecification into the area of syntax have been made.Footnote ^[14] Dynamic syntax uses structural underspecification to deal with a range of phenomena, i.e. scrambling, clitic doubling and person restrictions, to name a few (see Kempson & Kiaer Reference Kempson and Kiaer2010, Chatzikyriakidis Reference Chatzikyriakidis2010 and Chatzikyriakidis & Kempson Reference Chatzikyriakids and Kempson2011 respectively). The basic mechanism used in order to express structural underspecification is unfixed nodes. Unfixed nodes are nodes that have not yet found their position in the tree structure. As such, they are structurally underspecified at the time they are introduced. However, they need to be updated later on in order for the parse to be successful; otherwise, outstanding requirements remain and the tree update fails. The rule of *ADJUNCTION introduces an unfixed node:

The node is structurally underspecified since it does not carry a fixed treenode address. The only thing the node ‘knows’ with regard to its treenode position is that somewhere up above or at the current node, Tn(n) must be found.Footnote ^[15] The *ADJUNCTION rule works neatly for OV focus cases in Greek. In such structures, the preposed object is parsed on an unfixed node. Then the verb is parsed, projecting the rest of the propositional structure, a type $e$ and a formula metavariable at the subject node (for the subject drop properties) and leaving a requirement for a type value in the object node:Footnote ^[16] Footnote ^[17] Footnote ^[18]

In the above structure, there is an unfixed node with a type and a formula value and an open slot decorated with a type e requirement (the 010 node). It is at that point that a process of unification between the unfixed node and the fixed object node (010) can take place using MERGE. MERGE is a computational rule which unifies two nodes just in case one of the two updates the treenode address of the other. The notion of update is defined by treenode address entailment. If a treenode address entails another treenode address, then the former can be seen as an update of the latter.Footnote ^[19] The treenode address of the direct object node (010) can be a proper update of the underspecified address the unfixed node carries. Furthermore, the fact that the 010 node has a proper treenode address will eliminate the requirement of the unfixed node that a fixed treenode address should be found ( $?\exists \text{}\text{x}Tn(\text{}\text{x})$ ). The trees below display the tree structure before and after MERGE has been applied:

2.3 Parsing in context – LINK structures

Besides the tree structures in which each sentence involves a single tree (regardless of tree embedding), DS also makes use of pairs of trees which are linked to each other via a relation called LINK. LINK structures involve two separate tree structures. The node from which the LINK starts can be seen as setting the context in which the LINKed tree is going to be parsed. Examples of LINK relations include relative clauses, in which case the relative clause is parsed within the context of the head noun or hanging topic left dislocation (HTLD) construction, in which case the HTLD sentence is parsed within the context of having parsed the left-dislocated element first. LINK structures have a variety of uses in DS. Let us illustrate LINK structures in more detail by looking at an HTLD example. In order to analyse HTLD constructions, Cann et al. (Reference Cann, Kempson and Marten2005) define two rules which link a type e node, where the dislocated element is parsed, to a type t requiring node, where the rest of the HTLD structure is parsed. The first rule introduces a LINK transition from a type e requiring node to a type t requiring node, while leaving the pointer in the first of the two. The second rule introduces a requirement for a shared term as soon as the dislocated element is parsed. In parsing an HTLD sentence like the one shown in (23), we apply the first rule (TOPIC STRUCTURE INTRODUCTION) which introduces the type e requiring node. The NP is parsed on that node and then the second rule takes effect, moving the pointer to the type t requiring node and positing (on the same node) a requirement that a copy of the formula found in the node where the LINK begins must be found somewhere in the LINKed tree or in a tree LINKed to the LINKed tree. ( $?\langle D\rangle \mathit{Fo}(\mathit{Giorgos}^{\prime })$ ). In order to be able to talk about LINK relations in our tree language, we introduce the $\langle L\rangle$ relation and its inverse $\langle L^{-1}\rangle$ , which are used to refer to LINK relations ( $L$ referring to a LINKed node and $\langle L^{-1}\rangle$ to a node that a LINK starts). $D$ is a generalization of the unfixed node modality $\langle \downarrow ^{\ast }\rangle$ that further includes the LINK relation:Footnote ^[20]

In simple terms, in the above example the result of the two rules is to project a LINK relation from the HTLDed topic o Giorgos. Then, in this new LINKed tree, the rest of the sentence is going to be parsed. For our specific example, the requirement that a copy of the HTLDed NP must be found in the LINKed tree will be satisfied as soon as the object clitic is parsed.

For RD, we find a similar treatment, the difference being that the LINK structure is now initiated from a type $t$ complete node, i.e. a complete proposition. This idea has been used in the DS literature for BRD, with particular emphasis on pronoun doubling, and CRD in clitic languages like, e.g., Greek (Cann et al. Reference Cann, Kempson, Marten, Swinburne, Tsoulas and de Cat2004, Reference Cann, Kempson and Marten2005; Chatzikyriakidis Reference Chatzikyriakidis2010, inter alia). Dynamic syntax has a natural mechanism to account for these constructions, via the right periphery analogue of the rules used for HTLD constructions, RECAPITULATION. The difference between the two is that RECAPITULATION builds a LINK transition from a type t complete node, i.e. a complete proposition, to a type e requiring node by further positing that the formula value of the LINKed tree must be shared with some subterm of the main tree. This captures the intuition that BRDs are in fact topic re-establishers and are used to somehow re-establish the context. The rule is shown below:Footnote ^[21]

To put this in perspective, let us say we want to parse (30):

This construction contains a right dislocated subject, i.e. o Gianis ‘the John’. The first step in parsing such a sentence is parsing of the verb. The verb projects the whole propositional template as well as a type value and a formula metavariable in the subject node:

The NP to Giorgo comes into parse and decorates the object node with a type and formula value.Footnote ^[22] Then, after a value is provided from the context for the subject node ( $\mathit{Fo}(\mathit{Gianis}^{\prime })$ ) and the application of functional application for formula values and modus ponens for types we arrive at a complete parse:

Notice that parse of the sentence could stop here, since a well-formed parse has been established, given that a proposition and no outstanding requirements exist. However, in the case of BRD, the rule of RECAPITULATION can further apply, which will create a LINK transition from the type complete t node to a type e requiring node. It will further posit that a formula value of type e found somewhere in the main tree must be the formula value of the node of the LINKed tree. The structure after RECAPITULATION has been applied is shown below:

At that point the right dislocated subject can be parsed in the LINKed tree. Now, the same reasoning can also be used to account for right dislocated structures in which the right dislocated element is coreferential with a preceding clitic, i.e. CRD.Footnote ^[23] Such an account is based on the idea that right dislocated elements are optional elements that are somehow re-used to further establish known information. This neatly predicts the backgrounding topic effect, since it basically re-introduces an argument that has already been provided.Footnote ^[24]

Clitic right dislocation in Greek shows case connectivity, as the examples below illustrate:

As already mentioned, case in DS is seen as providing tree requirements that act as a filter on output. Simplifying a little bit, nominative case can be seen as imposing a requirement that its mother node is the top node ( $Ty(t)$ ) while for accusative case, this is the predicate node ( $Ty(e\rightarrow t)$ ). The two tree relation statements are shown below:

Chatzikyriakidis (Reference Chatzikyriakidis2010) uses this conception of case in order to get the connectivity effects exhibited in CRD in Greek by modifying the above statements as follows:

Basically, the above reads as follows: you will find a predicate node (for (33)) or a type $t$ node (34) if you traverse a LINK relation $\ast$ times (potentially empty) and then go up the argument node. If the $L$ is empty then we have cases of regular structural case (within the same tree structure). If it is not, this captures peripheral connectivity effects like RD discussed here.

The account in Cann et al. (Reference Cann, Kempson and Marten2005) and Chatzikyriakidis (Reference Chatzikyriakidis2010) just presented, as already said, is not an account of ATs but rather BRDs in Greek. However, ATs are clarifications and as such are used in order to help in the construction of a complete proposition. This means that by definition ATs do not operate on a full propositional structure, on a complete tree. Afterthoughts provide more information in order for an element previously introduced to receive reference. If ATs are indeed clarifications, one can look at them as responses to implicit clarification requests. In effect, the speaker in these constructions is trying to avoid a clarification question and does so by providing the answer to this hypothetical question. This answer is the AT. If such an account is to be pursued, one should at least have a way to deal with phenomena like fragment answers in a dialogue setting. In what follows, I take up the challenge of providing an account of ATs on that basis, showing that the gender mismatches associated with ATs in Greek are a direct consequence of such an approach.

3.1 Some introductory remarks: afterthoughts versus backgrounded right dislocation

The first thing that seems to be following out of recent work with respect to RDs is the rather uncontroversial claim that RDs involve two main different constructions: (a) backgrounded right dislocation (BRD) and (b) afterthoughts (ATs) (see Averintseva-Klisch Reference Averintseva-Klisch2006, Reference Averintseva-Klisch, Fabricius-Hansen and Ramm2008, Reference Averintseva-Klisch, Benz and Kahnlein2010; de Vries Reference de Vries2007; Ott and de Vries Reference Ott and de Vries2012a, Reference Ott and de Vries2014, among others). All of these researchers distinguish between the two on both prosodic and syntactic grounds. As such, there is consensus that the two structures involve different intonational patterns, with BRDs prosodically integrating into the host clause but ATs forming a prosodic unit of their own (Averintseva-Klisch Reference Averintseva-Klisch2006 and de Vries Reference de Vries2007, among others). There is also consensus that the two structures differ structurally. For example, Averintseva-Klisch (Reference Averintseva-Klisch2006, Reference Averintseva-Klisch, Fabricius-Hansen and Ramm2008) offers a number of pieces of evidence pointing towards this direction. The first of these, which is also relevant for the needs of this paper, is that morphological agreement is strict in the case of BRDs but not in the case of ATs. This has been shown by Averintseva-Klisch for German, and it is certainly true for Greek, at least as regards gender:Footnote ^[25] Footnote ^[26]

Optional additions between the clause internal pronominal (in our case the clitic) and the RD element are also possible in ATs but not in BRDs. In the example below, the optional addition enoo ‘I mean’ is used:Footnote ^[27]

Ott & de Vries (Reference Ott and de Vries2012a) propose an account for RDs based on the idea that RDs are instances of clausal ellipsis. In effect, the idea put forth by Ott & de Vries (Reference Ott and de Vries2012a) is that RDs involve the biclausal structure depicted below ( $\mathit{dXP}$ stands for dislocated XP):

In this paper, I take the view that indeed ATs should be analysed on a par with elliptical phenomena. In particular, I argue that ATs can best be seen as fragment answers to implicit questions. In order to formulate our claim, we base our analysis on some recent advances with regard to dialogue modelling using DS. This is what we turn to now.

3.2 Dialogue modelling in dynamic syntax

Dynamic syntax, as already mentioned, is a parsing-oriented, incremental framework where semantics and syntax work in tandem to provide structured representations of content. Dynamic syntax takes syntax to be procedures for context-dependent interpretation (Gregoromichelaki, Kempson & Cann Reference Gregoromichelaki, Kempson and Cann2012). Under such an approach, as Gregoromichelaki et al. (Reference Gregoromichelaki, Kempson and Cann2012) argue, ‘We have to see “semantics” as structured representations of content, “syntax” as the set of actions for constructing these representations, and “context” as a store of content, i.e. structures, plus the actions involved, hence the semantic representations and the syntactic process combined.’ (Gregoromichelaki et al. Reference Gregoromichelaki, Kempson and Cann2012: 574). This stance, alongside the incremental nature of the DS framework, has been able to provide a natural account of a number of problematic dialogue phenomena, like fragment answers or split utterances (Purver et al. Reference Purver, Gregoromichelaki, Meyer-Viol and Cann2010, Kempson et al. Reference Kempson, Gregoromichelaki, Meyer-Viol, Purver, White, Cann, Alain and Troncon2011, Kempson, Gregoromichelaki & Chatzikyriakidis Reference Kempson, Gregoromichelaki and Chatzikyriakidis2012, Gregoromichelaki et al. Reference Gregoromichelaki, Kempson and Cann2012, among others). To give an idea of how these assumptions work, let us see two different cases of dialogue phenomena, a split utterance and a fragment answer:

Example (40) is problematic given that the full sentence arising out of the conjunction of the two utterances is ungrammatical. Thus, did you burn myself as a single utterance is ungrammatical, but is however fine in a split utterance setting. Dynamic syntax offers a solution to this. This is based on the assumption that production and parsing work tightly together, and the same mechanisms are used in both.Footnote ^[28] Each of the interlocutors is seen as building representations incrementally and relative to some context. This has the welcome consequence that at any point in the parsing process, the interlocutors can switch the roles of parser and producer at any time. This assumption, along with the semantics/syntax of the reflexive, will give us a straightforward account of (40). The reflexive’s lexical entry is as shown below:

Myself is seen as copying a formula from a local co-argument node onto the current node. This formula must satisfy the conditions set by the person and number of the uttered reflexive; in our case it has to be naming the speaker. In the following example, we see two trees. The first one stands for the partial tree having been constructed by parsing of the first utterance, i.e. did you burn?. Note that a value has been provided for the metavariable projected by the second person pronoun you, in effect a value that identifies the hearer. Then, at that point the second participant can take on and provide the rest of the sentence, building on the previous partial tree. Then, myself can be used with no problem, since the speaker in this turn was the hearer when the first utterance was parsed. As such, and given the lexical entry for myself, the value provided by you by the first participant to identify the hearer can now be copied to the object node, given that it now coincides with the speaker:

Dynamic syntax is in this respect a framework that can not only deal with incremental syntactic and semantic processing, capturing the need for allowing the extension of partial semantic structures as well as syntactic dependencies, but can furthermore deal with the speaker/hearer changes at any step in the dialogue. Indeed, other grammatical frameworks can be modified in order to account for incremental semantic/syntactic parsing, and indeed such attempts to modify existing syntactic frameworks to handle incremental parsing have been made using categorial grammar, minimalist grammars or tree adjoining grammar (see Hefny, Hassan & Bahgat Reference Hefny, Hassan and Bahgat2011, Stabler Reference Stabler2013 and Demberg, Keller & Koller Reference Demberg, Keller and Koller2013 respectively). It is, however, the speaker/hearer interchanges, an example of which is shown in (40), and the associated problems that this brings about that make DS a better suited and motivated framework to handle incrementality as well as dialogue phenomena compared with other frameworks. For fragment answers like A: Who did Mary upset? B: John, John is parsed within the context of the Wh question. It is parsed on a structure LINKed to that structure with a type $e$ requirement (remember that the structure where the LINK starts can be seen as setting the context):

Now, we can parse the fragment answer John. Then the LINK rule evaluation substitutes John for the Wh metavariable:

In what follows, and for ease of exposition, we will not show the LINK transition in cases of fragment answers but we will rather substitute the Wh metavariable directly. The reader should have in mind that a LINK relation is involved in these cases, which is omitted just for illustration purposes and to make the trees a little more readable.

3.3 Afterthoughts

Afterthoughts, as already discussed, involve a distinctive period intonation and have been claimed to act as a kind of clarification rather than a means to re-establish an already introduced referent. Averintseva-Klisch defines the function of ATs as follows:

One way to model ATs, as already said, is to look at them as being clarification answers to implicit questions. The idea is that the speaker in these cases provides a clarification in order to ‘avoid’ the question that the AT is an answer to, in effect preventing the hearer from asking a clarification question. For example, in (47), the AT xtes ‘yesterday’ further clarifies the past event that the sentence before the AT expresses by restricting the past time occurrence of this event to the day before the time of utterance:

The same goes for argumental ATs. For example, in (48), it seems that the speaker tries to avoid an explicit clarification request by the hearer, by providing the answer to this never asked question:

Assuming that ATs are fragment answers to implicit clarification questions, then what is first needed is a way to analyse clarification answers. Work in dialogue modelling in DS and specifically fragment anwers, among other dialogue phenomena, can provide the basis of such an analysis of ATs. Let us see where such an assumption can lead us to. First of all, we have to look in more detail at how fragment answers are dealt with in DS. Let us assume that the AT in (48) is an answer to a question like pion htipise o Giorgos ‘who did George hit’:

Then, given the DS perspective as sketched, e.g., in Kempson et al. (Reference Kempson, Gregoromichelaki, Meyer-Viol, Purver, White, Cann, Alain and Troncon2011) or Gregoromichelaki et al. (Reference Gregoromichelaki, Kempson and Cann2012), among others, the fragment answer will be parsed within the context of the question. In effect, the fragment will provide a value for the Wh metavariable posited by the Wh element pion ‘who.acc’. The Wh question will act in this respect as the context in which the fragment is parsed. The idea is that structure provided by the immediate context can be used by the next utterance. In our case, the content provided by parsing the Wh question can serve as the point of departure for the fragment answer (Gregoromichelaki et al. Reference Gregoromichelaki, Kempson and Cann2012: 575). The AT ton Giani provides the value for the Wh metavariable ( $\mathit{Giani}^{\prime }$ ):Footnote ^[29]

The same process can be argued to occur in cases of adverbial ATs. In the case of (47), the adverb xtes ‘yesterday’, functioning as the AT, can be also seen as a fragment answer to the implicit clarification question pote ides ton Giorgo ‘When did you see George’. Thus, the adverb xtes is parsed within the context of this question. In effect, the AT in this case further specifies the past time metavariable projected by the verb to be those past times spanning the whole day before the day of utterance. The context set by the implicit question is shown below. Note that in this case we are also making use of a situation/event node. This is the node where tense/aspectual information is represented in DS. We will represent this node only in case this is needed (note that the subject metavariable has been updated to the value $\mathit{Fo}(Maria^{\prime })$ ):

Then the AT comes into parse, updating the $R$ time metavariable into yesterday’s past time:

However, we are not over yet. We have not yet explained how the structure shown above is linked to the sentence prior to parsing the right dislocated element. We start from the parse of the main sentence without the right dislocated element, i.e. (53):

Note that there is an uninstantiated time metavariable, i.e. $R$ . This metavariable along with Wh metavariables are the only metavariables that are allowed to be used in functional application prior to being updated into proper values. All other metavariables (like metavariables standing for argument NPs) have to be substituted before functional application takes effect. In the case above, the $R$ metavariable is turned into a Wh metavariable. In effect, this is the only difference between the declarative sentence just parsed and the implicit question that the right dislocated element functions as a fragment answer to in terms of the end parse result.Footnote ^[30] The trigger for this might be considered to be the specific AT intonation. However, the exact trigger and whether this assumption has any basis at all have to be further elucidated. Assuming, however, that something triggers an AT interpretation (like, e.g., as already said, period intonation), we can assume that as soon as the AT is parsed, the context in which it is parsed includes the implicit question. In the example of interest, the difference between the already parsed sentence and the implicit question is the $R$ metavariable which is substituted by a Wh time metavariable in the implicit question. Thus, we get the following two steps in parsing the AT: (a) we parse the sentence before the AT, (b) the AT is parsed within the context of an implicit question potentially triggered by AT intonation:

The next step is to see whether such an account can account for gender mismatches found in Greek ATs. This is what I turn to now.

3.4 Gender mismatches

The repeated examples below exemplify cases of gender mismatch in MG ATs. In both examples, a neuter clitic is doubled by an NP AT which is marked for female gender:

How can we make sense of these data? Well, if what we have proposed has any basis at all, we should approach cases like those above as cases involving the same line of reasoning: ATs are treated as fragment answers. Let us see how far this analysis can take us with respect to gender mismatches like the ones in (55)–(57). We first parse the sentence before the AT. Let us exemplify this with (56), i.e. to ida xtes. The structure we get is the following, where the metavariable projected by the clitic has not managed to be substituted. The structure is shown below (note that we assume that the subject metavariable has received a fixed value, $\mathit{Fo}(Stergios^{\prime })$ ):

Note that the tree cannot be compiled given that the object node does not have a proper formula value. This is the structure the producer has in mind. It is not necessarily the one that the parser has, since the addressee might well have resolved the anaphora and provided a value for the pronoun’s metavariable. This is in effect the assumption the producer makes with regard to the sentence before the AT, and this is why he uses a clarification strategy, i.e. the AT. As already said, the AT functions as a fragment answer to a clarification question. The clarification question in the above example provides a Wh metavariable in the object node, and a complete parse can be obtained given that Wh metavariables are allowed to participate in funtional application. This is the key part that will give us the explanation for the attested gender mismatches. The relevant Wh question is the one shown below:

The Wh element used in this case, ti, is underspecified for gender. It can be used to refer to neuter and masculine referents as well:

This is the key to the attested mismatches. Given that the Wh metavariable will be underspecified for gender, the AT tin ekpobi, ‘the show’, even though marked for feminine gender will be able to get parsed given that no gender restriction is posed by the Wh element. However, there is something missing here. Even though ti is underspecified for gender, the clitic correlate is not. In this respect, it is impossible for the metavariable projected by the clitic ( $\mathit{Fo}(U_{neut})$ ) to be updated to the Wh metavariable with no gender restriction. This will violate the requirement for monotonicity of the system. However, this fact is also the key to why gender mismatches are attested only with the neuter clitic and not with the male or feminine clitic. This has to do with the special status of the neuter clitic to. Let us explain. The clitic to can function as both a type $e$ argument as well as a type $t$ propositional argument. In the latter case, $to$ refers to a whole clause, usually a sentential complement. The examples below are illustrative of the phenomenon. In both cases, the referent of $to$ is a type $t$ propositional complement:

Given this behaviour of to, its lexical entry must ensure that both cases ( $e$ and $t$ ) can be captured. In order to ensure this, the metavariable cannot have a neuter value given that this will not be the case in the case that $to$ is of type $t$ . On the other hand, it must ensure that in the case that it is a type $e$ value, the metavariable will have a neuter value. In order for this to be captured, the following entry for $to$ is proposed. This entry has both an underspecified type and an underspecified Fo metavariable with no gender restriction. However, it posits that if a type e has been established and a metavariable is still there, then the metavariable must be gender restricted:

Now, the Wh element ti, besides being underspecified for gender, is also underspecified for type, given that similarly to to it can refer to a full proposition. The relevant data are shown below:

Now let us see how the account works, by looking at the parse of example (56). We first parse the structure before the AT, i.e. to ida xtes:Footnote ^[31]

This is now the context for the implicit question. Parsing the implicit question ti ides xtes ‘what did you see yesterday’ will return the following tree, where ti updates the metavariable $\mathit{Fo}(U)$ to $\mathit{Fo}(Wh)$ , keeps the type metavariable and decoration $Ty(X),?\exists x.Ty(x)$ given that ti is also underspecified for type, as we have said, and also keeps the restriction $?\exists X.((\mathit{Fo}(X)\wedge Ty(e)\rightarrow (\mathit{Fo}(X_{neut}))))$ :

As it stands, the restriction is satisfied. This is because the first part of the implication is false given that no type $e$ is present ( $\exists X.\mathit{Fo}(X)\wedge Ty(e)$ ). Hence, the whole formula is true.

Now, within the context of this implicit question, the AT can be parsed, substituting a proper formula value for the Wh metavariable and also updating the type metavariable ( $Ty(X)$ ) into a proper type value ( $Ty(e)$ ):Footnote ^[32]

Again, the restriction $(\exists X.\mathit{Fo}(X)\wedge Ty(e))\rightarrow (\mathit{Fo}(X_{neut}))$ is satisfied since again the first part of the implication is false (given that no metavariables exist anymore).

In case of ungrammatical examples where the clitic correlate is either male or female and there is gender mismatch with the AT like the example below, the account makes the correct predictions:

Given non-monotonicity of the system and given that the masculine/feminine clitics will contribute a gender restricted metavariable, the metavariable provided by ti will have to be gender restricted as well. Thus, the implicit question in (69) will involve the following:

Thus, the account proposed makes the correct prediction with respect to gender mismatches. It predicts the licit mismatches while it further explains the non-attested ones. There is, however, an added complication to the data. In the case that animate referents are involved, no gender mismatches are allowed, as the data below show:

Now, if the same Wh metavariable was used, then both of the above examples should have been grammatical. However, things are a little different in the case of questions that expect an animate referent as an answer. The relevant question to the AT in the above examples will not involve ti but rather pio, which is used for neuter animate referents. In this respect, if we want to ask a clarification question given to pantreftike ‘it married him’, the question will use pio and not ti:

Thus, in cases like (74) and (75) the context in which the AT is going to be parsed will involve a Wh metavariable specified for gender, and thus will not allow substitution in cases where a gender mismatch is observed:

Given this context, the DP i Maria cannot be parsed, since update of the metaviariable to the value $\mathit{Fo}(Maria^{\prime })$ will be impossible due to incompatible gender values. Thus, the approach proposed here provides a natural explanation for the grammaticality of (55)–(57) on one hand, and the ungrammaticality of (72) and (73) on the other.

To recap, let us first summarize the data. The following generalization explains the full range of the pattern:

Thus, the only licit combinations are between to ‘it’ and an inanimate feminine or masculine NP AT. Now, the reason why gender mismatches are possible in Greek ATs is due to the fact that in the cases where these are manifested, the implicit question involved can always involve the Wh element ti, which is underspecified for gender. In cases of gender mismatch where a clitic and its double are involved in the mismatch, there are two options. (a) If the Wh is the neuter clitic to and its referent (or at least the referent the speaker has in mind and that surfaces as the AT) is inanimate, then the implicit question in this case involves the underspecified for gender Wh. Thus, in this case, the construction is grammatical, as in (55)–(57). (b) If the referent is animate, then the marked for gender Wh pio should be used, in which case gender mismatched ATs are not possible, as in (72) and (73). The account proposed here thus explains ATs as well as gender mismatches as a by-product of tree growth. This follows the standard DS assumption that discourse effects can be seen as such.

3.5 Multiple ATs

Constructions with more than one AT can be found, as witnessed by the examples below:

These data are naturally explained within the account proposed in this paper. Let us take a look at (78) in order to see how this is done. We first parse the correlate ton ida:Footnote ^[33]

This is now for the context for the implicit question, and after the implicit question what we get is the following:

In this context, the AT will update the Wh metavariable, leading to the following tree:

Now, we have reached a tree exactly the same (modulo the different Fo values for the arguments) as the one in (53). From this point on the same steps can be used to parse the second AT, as is the case in (53).

4.1 Some further properties of ATs

4.2 A further remark on a future cross-linguistic comparison

We have discussed gender mismatches in Greek, in a range of specific constructions involving a clitic correlate and a gender mismatched AT. One of the interesting questions to be asked (also posed by an anonymous reviewer) is what are the predictions that the present account makes cross-linguistically with respect to these structures. The idea is that if you have a language with elements comparable to neuter clitic to in Greek, comparable underspecified Wh elements like ti and also gender marking, one would expect similar phenomena to arise in these languages as well. A cross-linguistic comparison is out of the scope of this paper. However, there is preliminary evidence that a similar situation might be operative in German. My German consultants have judged sentences containing gender mismatches with the correlate being es to be much better than sentences where the correlate is either the masculine or feminine pronoun. The relevant data are shown below:

In principle then, a similar idea can be pursued for the above cases as well. In a more general perspective, it would be good to investigate cross-linguistically and empirically verify or falsify the predictions the current account makes. This cannot, however, be done in this paper. Even for German, the data are all very tentative and more speakers have to be consulted. For the moment, we leave this as a subject of further research.