Grammaticality in Generative Grammar, Radical Construction Grammar and Functional Discourse Grammar

(1)

Master’s Thesis MA General Linguistics Graduate School for Humanities

Universiteit van Amsterdam

Grammaticality in Generative Grammar,

Radical Construction Grammar and

Functional Discourse Grammar

Mario Laurens van der Velden student number: 5740827

(2)

Acknowledgements

I am grateful to Hella Olbertz and Petra Sleeman for their valuable comments, for their patience, and for their time. I am also very happy that Hella Olbertz was willing to continue her supervision from Brazil. Furthermore, I feel fortunate that I was supported in pursuing the somewhat abstract research question of this thesis.

(3)

List of figures

Figure 1 Examples of types of data used in linguistics ... 14 Figure 2 Grammar and grammaticality ... 17

(5)

1 Introduction

Series of well-formed and ill-formed example sentences are often found in linguistic argumentation. Usually, the ill-formed sentences are preceded by an asterisk. This is also the case in the following example.

a. Stephen punished Alan. (1)

b. *Punished Stephen Alan. c. *Stephen Alan punished. d. (*)Alan punished Stephen. e. *Alan Stephen punished f. *Punished Alan Stephen.

These sentences all use the same words. The only thing that changes is the order of the words. As there are 3 words in total, there are 6 possible configurations in which the words can occur. All these combinations are represented in the example. The fact that only two of these sentences are well-formed, (1a) and (1d), shows that in English there are only 2 ways in which these particular words can combine. A number of conclusions can be drawn from example (1) about the behavior of Stephen and Alan on the one hand, and the behavior of

punished on the other. Some conclusions can also drawn about the order in which these

words need to occur.

However, example (1) does not only say something about the specific words occurring in it, as it is possible to substitute Stephen and Alan with other names.

a. Sandra punished Josephine. (2)

b. Bill punished David.

Furthermore, it is possible to substitute punished with other inflected verbs. a. Alan impressed Stephen.

(3)

b. Bill amazed Stephen.

Example (1), then, can also be taken to say something about the word types to which

(6)

Another thing that is interesting to note is that if sentences (1b – 1f) are taken to mean the same as example (1a), only the first sentence is well-formed. The otherwise well-formed (1d) means the exact opposite of (1a), and would therefore be ill-formed on ‘the intended reading’ (cf. Keizer 2009: 1201). That is why (1d) is preceded by an asterisk in brackets.

In the discussion so far, well-formedness has been taken for granted. However, part of the following thesis is dedicated to examining what well-formedness is. It can mean at least two things. Firstly, a linguistic expression can be well-formed in the sense that it is acceptable in the ears of a native speaker. A term often encountered in this context is acceptability

judgment. Secondly, a linguistic expression can be well-formed in the sense that it is in

accordance with the grammar of the language the expression belongs to. Whether or not this is the case is a question of grammaticality.

A problem with the latter notion is that grammar can be understood in many ways. In fact, regarding this issue there are several, sometimes extremely diverging standpoints taken in linguistics. The present study examines three such standpoints, associated with Minimalist Generative Grammar (henceforth GG), Functional Discourse Grammar (henceforth FDG), and Radical Construction Grammar (henceforth RCG) respectively.

The main research question of the thesis is the following.

(A) What are the similarities and differences between GG, FDG, and RCG with respect to the notion of grammaticality?

In order to make this comparison, this thesis critically analyzes some of the relevant literature, both regarding grammaticality in general and regarding GG, FDG, and RCG in particular. In chapter 2 the notions of grammaticality and acceptability will be dealt with. Some complications that are associated with these notions will also be discussed there. In chapter 3 the results of the previous chapter will be put to use in a small scope

characterization and comparison of the GG, FDG, and RCG regarding their viewpoints on grammaticality. First, however, the remainder of this chapter will introduce GG, FDG, and FDG and put them in a historical context.

1.1 Preliminary complications

Before starting the comparison, it might be useful to examine what is actually being compared: what are GG, FDG, and RCG? There seems to be no straightforward answer to

(7)

this. One of the main problems is that we are dealing with highly abstract concepts that can refer to a multitude of things: theories of language, theories about grammar, theories about specific languages (particular grammars), models of grammar using a formal metalanguage, research programs, groups of linguists (possibly organized in research groups and

universities); to name just a few. In fact, the names or labels GG, RCG, FDG can be rightfully used to refer to any of these things.

Given these complications the following thesis must severely restrict itself. One way of doing this is to focus on specific aspects of GG, RCG, and FDG. The present study will be limited to the notion of grammaticality and to notions closely associated with it.

GG, FDG, and RCG all provide a general way of analyzing linguistic expressions in whatever language, and they all have their own distinct way of doing this. In this sense, they can all be called general theories of linguistic structure (the term ‘general theory of linguistic structure’ is taken from Chomsky 1957/2002: 50). GG, FDG, and FDG, in fact, have

produced a body of analyses of linguistic expressions in various languages that together may form theories about these languages. Even though GG, FDG, and RCG are much more than general theories of linguistic structure, in what follows they will sometimes be referred to as the three (general) theories (of linguistic structure). Another term that might be used in this context is that of ‘theory about grammar’ (FDG is called that in Hengeveld & Mackenzie 2008: 2), but the term grammar itself can be ambiguous, as will be seen in chapter 2, and it is therefore less suited for the present purposes.

There is something problematic about the term general theory of linguistic structure: it contains the word theory. In Minimalist GG (Chomsky 1995) the term ‘conjecture’ (Lasnik 2002: 436) is explicitly used instead of the word theory on many occasions, reflecting the fact that Minimalist GG is actually a (research) program. The following quote from Hengeveld suggests that the something similar might be the case for FDG: ‘FDG is a research

programme rather than a fully-fledged theory’ (2004: 365). Still, GG, FDG, and RCG provide very clear views—preliminary or not—on how linguistic expressions are structured. Given this fact, the term theory, in spite of the complications associated with it, will be used throughout this text.

A further complication is that GG, FDG, and RCG are elusive in a number of ways. If the names are seen as referring to groups of linguists for example, there is the problem that a group can change: new members may join, while others may leave the group. Also, the ideas of individual members of the group may change over time, possibly leading to schisms (i.e. between Generative Semantics and mainstream Generative Grammar (cf. Taylor 2007: 571))

(8)

or leading to proposals to amend some of the fundamental aspects of the general theory (i.e. the Minimalist Program for GG (Chomsky 1995), and Functional Discourse Grammar for Functional Grammar (Hengeveld & Mackenzie 2008)). The names (or labels) GG, FDG, and RCG are used by linguists who associate themselves with them, that follow the theory somehow and more importantly, that amend it, propose new grammatical analyses (that are followed or discarded), in sum, that keep the theories alive and evolving. While this social aspect of GG, FDG, and RCG is not of interest here, it is it worthwhile to be conscious of it, and of the type of changeability that follows from it.

From what has been said thus far, it can easily be seen that GG, FDG, and RCG are actually very complex. Even though they all carry the word grammar in their name, there is a risk of comparing things that cannot be compared. In what follows it is assumed that there are enough similarities to warrant a comparison, even though this is still a very challenging task. The fact that GG, FDG, and RCG differ from each in a number of crucial respects and represent some of the very different standpoints linguists can take regarding grammar, at the same time makes comparing them all the more interesting.

It should be clear that this thesis cannot pretend to give a full, detailed description of GG, FDG, and RCG. All three are highly developed and involve complex and very abstract notions. Two of the three also offer a ‘technical language’, which, according to Croft, can be so difficult to learn that it is ‘like learning a human language’ (Croft 2001: 3).

Lastly, a number of comparisons have been published over the last decade. A notable example is Gonzálvez-García & Butler (2006), who compare FDG and RCG amongst others; Taylor (2007) compares Cognitive Linguistics (of which RCG is part) and Autonomous Linguistics (to which GG belongs); Nuyts (2007), in turn, compares Cognitive Linguistics with Functional Linguistics explicitly mentioning Functional Grammar, the predecessor of Functional Discourse Grammar, as part of the latter approach. All these publications strengthen the idea that a comparison between several theoretical approaches is worthwhile.

1.2 Broader context

In this section GG, FDG, and RCG will be positioned within the field of linguistics relative to other approaches. After that, they will be put in their historical contexts and some basic tenets will be named.

(9)

In linguistics there are quite a few very general approaches to the study of language, encompassing several specific formulations and theories. It is possible to distinguish between various schools of linguistics such as ‘Cognitive Linguistics’, ‘Functional Linguistics’ (Nuyts 2007), and ‘Generative Linguistics’ (Taylor 2007), although it is very difficult—if not

impossible—to give precise criteria for demarcating these (Nuyts 2007). Rather the theories and frameworks that fall under the heading of such a school may share with each other, inter alia, some basic assumptions, methods, and orientations.

The three general theories of linguistic structure on which this thesis focuses can be said to be part of more general approaches to the study of language. GG is considered part of Generative Linguistics (to which it lends its name). FDG and its predecessor Functional Grammar (Dik 1997) are sometimes seen as belonging to Functional Linguistics (Nuyts 2007: 546). RCG is explicitly part of Cognitive Linguistics (cf. Croft & Cruse 2004 and Croft 2007b).

In the present study, the emphasis lies on notions such as grammar and grammatical analysis—the analysis of the structure of linguistic expressions. Therefore, rather than the schools in their entirety the specific ‘approaches to grammatical analysis’ (Hengeveld & Mackenzie 2008: 26) that follow from them are of interest here. Regarding what can be called approaches to grammatical analysis, sometimes two ‘major schools of linguistic thought’ are distinguished, namely ‘formalism’ and ‘functionalism’ (Hengeveld & Mackenzie 2008: 25). The most basic dividing line between the two is that formalism is interested solely in the form of a linguistic expression, independently of its function, while functionalism claims that linguistic structure cannot be studied independently of its function (i.e. meaning and communication). Note that functionalism is used differently here than in the term Functional Linguistics: functionalism is seen as an approach to linguistic analysis with regards to the distinction between form and function, while Functional Linguistics refers to a general approach to linguistics encompassing several distinct theories. Even though FDG can be said to belong to Functional Linguistics and RCG is part of Cognitive Linguistics, FDG is arguably less functionalist than RCG, as FDG is positioned halfway between formalism and functionalism construed in the former way, while RCG is positioned more towards the functionalist end of this spectrum. GG, on the other hand, is radically formalist, maybe even to the extent that it can be viewed as the prototypical instance of radical formalism.

(10)

1.2.1 Minimalist Generative Grammar

GG has a long tradition. It developed in the 1950s. One of the oldest, seminal

formulations is Chomsky’s 1957 publication Syntactic Structures (Chomsky 1957/2002), which really propelled the generative linguistic approach. Several versions of GG have since been proposed. In this study, GG is understood as the general theory of linguistic structure that builds on older versions and that is part of the research program known as the Minimalist Program (Chomsky 1995). The Minimalist Program itself—as a research program—is not of concern here, as this would complicate matters enormously. Rather those parts that were part of the older versions of the theory and are retained and built upon in present theorizing are construed as a general theory of linguistic structure that is referred to here as GG.

The name Generative Grammar refers to the goal within GG of providing an explicit formal model that can generate the expressions of a language. In recent versions (such as the Minimalist Program) cognition plays a very important, if not defining, role. GG is concerned with the cognitive aspects of language, in particular with the human language faculty, also called ‘I-language’ (Chomsky 1990: 513). This language faculty is seen as a system that is internal to human cognition, locatable in the mind/brain. As such it is assumed to be a real entity in the mind (cf. Chomsky 1990: 514). The idea is that while at the moment this entity is still an abstract characterization—pertaining to the notion of mind— it will become more concrete as the ‘biological sciences’ (Chomsky 1995: 1) progress—ultimately becoming locatable in the brain. The language faculty, then, is seen as a real entity.

Furthermore, in GG, the language faculty is divided into a ‘performance systems’ and a ‘cognitive system’ (Chomsky 1995: 2). As such, there is a fundamental difference between performance (e.g. the actual production of an utterance by the speaker) and competence, the knowledge a speaker has of (his/her) language. GG is primarily interested in competence as opposed to performance, and consequently GG examines the cognitive system and not so much the performance systems. That is, GG examines how the knowledge a native speaker has of language is represented in his/her mind (or brain) and does not try to model linguistic performance (i.e. language production and comprehension).

One of the distinguishing characteristics of GG is that it sees cognition as something that is subdivided into independent, but interacting modules. So for example the aforementioned performance systems and cognitive system function independently from one another. More specifically the cognitive system is said to interface with two other performance modules: a module of the mind that that takes care of articulating and perceiving actual utterances and a

(11)

postulated autonomy is one of the points of contrast between GG and other linguistic theories (cf. Taylor 2007).

The cognitive system, which takes care of knowledge of the structure of language,

functions independently within the mind. Therefore, it can also be studied independently. The structures of language produced by this computational system (morphosyntactic structures) can be analyzed independently of their meaning and independently of their use in

communication. It therefore straightforwardly follows that, in GG, syntax is autonomous.

1.2.2 Functional Discourse Grammar

FDG is the successor of Simon Dik’s Functional Grammar (as presented in inter alia Dik 1997) developed in the 1970s. One of the major publications that comprehensively set out the theory is Hengeveld & Mackenzie (2008). The name of the theory reflects the fact that the function of language (i.e. communication) plays an important role and that language is seen in terms of discourse acts. FDG takes a ‘structural-functional’ approach to the study of language (cf. Hengeveld & Mackenzie 2010: 372 and Gonzálvez-García & Butler 2006): it is neither ‘radically formalist’ (as can be said of GG), nor ‘radically functionalist’ (Hengeveld & Mackenzie 2008: 26).

FDG is not a cognitive theory in the sense of GG, as it does not seek to describe how knowledge of language is represented in the mind of the speaker, but rather seeks to describe patterns found in language: it is a ‘pattern model’ (Hengeveld 2004: 366; cf. also Butler 2008: 17). The notion of language is thus more external in FDG than in GG, where the knowledge of the individual plays a more important role (cf. the distinction between

‘I-language’ and ‘E-‘I-language’, Chomsky 1990: 510, 513). However, to some extent FDG is also interested in the ‘individual-psychological […] dimension of language’ (Hengeveld &

Mackenzie 2008: 29), and therefore with individual grammatical ‘communicative competence’ (Gonzálvez-García & Butler 2006: 82), although the addition of ‘communicative’ is crucial here.

FDG does, however, try to ‘reflect psycholinguistic evidence in its basic architecture’ (Hengeveld & Mackenzie 2008: 2). This is an important aspect of what is called

‘psychological’ (Dik 1997: 13) or ‘cognitive adequacy’ (Mackenzie 2012: 422), for which FDG strives. This means that the model looks in some ways like a speaker model. However, the theory does not posit an innately specified grammar, as does GG, or a type of mental organization of linguistic knowledge, as does RCG; rather it specifies how languages are structured.

(12)

1.2.3 Radical Construction Grammar

Radical Construction Grammar (henceforth RCG), presented in Croft (2001) is a theory about syntactic structure. It is a construction grammar (cf. Croft & Cruse 2004). That is, it does not make a distinction between a lexicon and a grammar with rules that make the combination of lexical items into linguistic structures possible. Instead, in construction grammar (and in RCG) there is a ‘syntax-lexicon continuum’ (Croft 2001: 58). RCG differs from other construction grammars—and is therefore radical—in that it does not define constructions in terms of what parts can occur in it, but it defines the parts that can occur within a construction in terms of the construction.

In RCG, like in GG, cognition plays a very important role. In fact, it is part of what is called Cognitive Linguistics (see Croft & Cruse 2004). Just as GG, RCG is concerned with how grammatical knowledge is represented in the mind (Croft 2001: 3). It seeks to model this knowledge, describe how it is represented. However, RCG sharply differs from GG in the way it tries to realize its cognitive aims, much of which has to do with the overarching

approach to linguistics in which RCG is situated. In Cognitive Linguistics, contrary to what is the case in Generative Linguistics, there is no separate language module in the mind (cf. Croft & Cruse 2004: 1–2). In this line of thought, there is no such thing as a computational system with its own principles of operation. Instead, language is viewed in terms of ‘more general cognitive abilities’ that are not exclusively linguistic (Taylor 2007: 566).

Furthermore Cognitive Linguistics in a broad sense advocates, much more so than Generative Linguistics, the incorporation of external data (Kertész et al. 2012: 652), that is, data from adjacent disciplines (mainly those that are part of the cognitive sciences, such as psychology). These types of data are even used in the establishment of the basic architecture of the model, contrary to what is the case in GG (but similarly to what is the case in FDG). Cognitive Linguistics tries to formulate theories that are compatible with ‘psychological and neurophysiological results’ (Kertész et al. 2012: 652). That is not to say, of course, that there is no support from other sciences for the claims made by Generative Linguistics.

RCG does not only describe how knowledge of language is represented in the mind, it also seeks to provide a means to describe how knowledge of a particular language is represented in the mind. In other words, it tries to offer ‘language-particular cognitively realistic description’ (in the words of Haspelmath 2004: 555). Grammatical analyses of linguistic expressions within RCG are in fact hypotheses about how these linguistic expressions are represented and structured in the mind.

(13)

1.3 Data in linguistics

A comparison of GG, FDG, and RCG with respect to the notion grammaticality is situated in the discussion about data in linguistics. Grammaticality is ultimately something that is supposed to reflect reality. In order to determine whether or not a certain conception of grammaticality is correct, one has to rely on the manifestations of that reality, in other words, the data that are available or furnished by research.

The topic of data in linguistics has always been very important and a divide between linguists of different persuasions. According to Kertész & Rákosi, in the second half of the twentieth century there was a ‘dichotomy of introspective and corpus data’ that divided linguists (2012: 12-15). This led to various methodological debates, starting from the middle of the 1970s, between adherents of corpus linguistics and those defending introspective data. The origin of data, i.e. a corpus or introspection, was the main criterion for adequacy; both parties claiming that their source was the adequate one and the other source was flawed.

Today, however, attention has shifted more towards the function of data (Kertész, Schwarz-Friesel & Consten 2012: 651). Moreover, in linguistic practice, there no longer seems to be an exclusive reliance on one data source only. Kertész, Schwarz-Friesel & Consten argue that the contemporary discussion of data in linguistics concentrates on the question of what different types of data should be used and how (ibid.).

The topic remains very complicated: there are several conflicting approaches to the study of language and even within these approaches there seems to be no general consensus on the matter, as is shown by some recent publications within, for example, Generative Linguistics1 or Cognitive Linguistics.2

The issue of data in linguistics is of current significance as illustrated by a number of recent debates that were ‘sometimes extremely heated’ (Kertész & Rákosi (2012: 1). A notable example is a special journal issue of Studies in Language (volume 28: 3, 2004) entitled ‘What counts as evidence in linguistics’, in which linguists from very different persuasions published. Another example of a debate that definitely deserves to be called

1_{An example is the discussion within generative grammar sparked by Schütze (1996) that}

continues to this day with special journal issues like Theoretical Linguistics 33–3 (2007), entitled ‘Data in generative grammar’, in which Featherston (2007) argues for the use of experimentally gathered data in generative grammar.

2_{Cf. the special issue of Language Sciences, 43 (2012) titled ‘Converging data sources in}

cognitive linguistics’ for which Kertész, Schwarz-Friesel & Consten (2012) provides the introduction.

(14)

heated was ignited by an article by Evans & Levinson (2009) entitled ‘The myth of language universals’, in which ‘linguistic diversity’ is presented as a ‘crucial datum for cognitive science’ (ibid.: 429).

The present thesis aims at contributing to the discussion by descriptively comparing three important approaches in linguistics (GG, FDG, and RCG) that propose different relations between theory and data. There is no attempt of taking a stance in the matter; the three approaches all have their own merits and are highly sophisticated. The purpose of what follows is to give a small, neutral comparison that highlights some important differences. First, however, the remainder of this chapter will discuss several types of data and present a number of classifications that have been proposed in the literature.

1.3.1 Types of data in linguistics

At first sight, linguistic theories seem to be supported by a wealth of different types of data: acceptability judgments, ‘intuitions about derivational morphological relationships among words’, ‘intuitions about correspondences among different utterance types’,

‘identification of structural versus lexical ambiguity’, ‘discriminations of the syntactic status of superficially similar word strings’ (Schütze 1996: 1–2); but also data from ‘interviews or questionnaires’, data from ‘psycholinguistic experiments’, ‘neurolinguistic experiments’, and from ‘computer modeling’; even ‘errors performed’ can be used. ‘Cross-linguistic data’ can be used, as can ‘corpus data’ from various types of corpora (Penke & Rosenbach 2004: 485– 508).

It is clear that linguists can use a lot of different data types. The types of data mentioned in the previous paragraph can roughly be grouped in the following categories: data from disciplines other than linguistics proper (e.g. neurolinguistic results), data that are available in advance (that naturally occur, e.g. corpus data) and data that are produced by the researcher (e.g. questionnaires). Without going into all the aforementioned data types, the following sections will deal with these broad categories. In the last section of this chapter, a special kind of data produced by the researcher will be discussed, namely that of intuitions that can be used in linguistics (all the examples of Schütze 1996 mentioned in the previous paragraph fall into this category).

(15)

1.3.2 Data from adjacent disciplines

First of all, data from other disciplines than linguistics such as psychology, neuroscience, computer science, and philosophy3 can be brought forward in support of linguistic theories, and, more specifically, in support of a general theory of linguistic structure (or a theory about grammar). These types of data are called ‘external evidence’ by Haspelmath (2004: 574).

Results from other disciplines are of great importance to all the three theories. GG makes claims about how the language faculty should be embedded in the cognition and ultimately the brain (Chomsky 1995: 1) and thereby raises important questions for ‘the biological sciences’. Such claims ultimately need to be backed up by neuroscientific data. RCG is part of the school of Cognitive Linguistics, and as such it must to some extent be in line with the ‘cognitive commitment’ (Lakoff 1991: 54), ‘the commitment to make one’s account of human language accord with what is generally known about the mind and the brain from disciplines other than linguistics’. In FDG results from psycholinguistics are reflected in the theory (Hengeveld & Mackenzie 2008: 2).

Data from other disciplines, often stemming from interdisciplinary fields such as psycholinguistics, sociolinguistics, and neurolinguistics will not be taken into account here; that would lead far beyond the scope of this thesis. This means that results from

psychological experiments, research into language acquisition using ‘computer models’ (Penke & Rosenbach 2004: 491), or research using ‘fMRI’ (Garnham 2013: 30.4) scans, to name a few examples, will not be dealt with here. It is, however, important to be aware of the fact that these types of data can be used in linguistic argumentation, and can be of great importance in determining the adequacy of a general theory of linguistic structure that is concerned with cognitive aspects of language.

1.3.3 Naturally occurring data and data produced by the researcher

Borsley makes a distinction between ‘naturally occurring data’ and ‘data […] produced by the researcher’ (2004: 1475). Naturally occurring data are real utterances uttered by native speakers, for example, or written text. These naturally occurring data can be taken from a corpus. Corpora can come in a variety of types depending on the source of the data (e.g. spoken or written language). In addition, data taken from a corpus can also come in a number

3_{These are among the ones considered by Miller to be part of cognitive science and to have}

been involved in the ‘cognitive revolution’ (2003: 143). All are in some way related to linguistics.

(16)

of types. There are ‘qualitative’ and ‘quantitative’ corpus data (Penke & Rosenbach 2004: 486). Mackenzie, for example, gives three tables that show how often the word fail and its derivations occur in combination with the word to in the British National Corpus and the Time Corpus of American English and concludes that they are ‘markedly more common in newspaper, academic and non-fiction macroregisters than in fiction or spoken English’ (2009: 3). The naturally occurring data used in this case are quantitative.Qualitative corpus data will be discussed in the following chapter (section 2.2.2).

Data that are produced by the researcher also come in a number of types. The results of various types of experiments (where native speakers have to produce or judge certain

utterances for example) are data in this sense. Data that are obtained through introspection are also data produced by the researcher.

The preference for either naturally occurring or data produced through introspection of the researcher led to what Kertész & Rákosi (2012: 14) call a ‘dichotomy of introspective and corpus data’. Linguists working with corpus data regarded linguists working with

introspective data as working in a wrong way, and the other way around. Borsley’s (2004: 1475) distinction, however, in itself does not take a stance in the matter. The distinction is used in classifying data, and it is important to emphasize that the distinctions made in this thesis are not value judgments in any way.

Concluding, it should be said that every utterance might in principle be used as a datum. The sentence Stephen punished Alan may be used, even though it was made up and not taken from actual discourse. The sentence might have occurred in actual discourse, however, (it probably did multiple times) and even if this were not the case it is easily imaginable that it would be used in actual discourse. Still, it might be a good thing to keep in mind the

somewhat artificial nature of this type of examples.

1.3.4 Intuition-based data

As seen in the previous section, naturally occurring data can be obtained through introspection and through experiment. There is, however, a very important type of data widely used in linguistics that can be obtained both through introspection and experiment: this type of data can be called ‘intuitive data’ Penke & Rosenbach (2004: 491). Intuitive data are intuitions native speakers of a language have about specific linguistic expressions in their language, in particular about the ‘acceptability and interpretation’ of these expressions (Newmeyer 1983: 48). In other words, this type of data is about what speakers consider to be

(17)

the meaning of particular linguistic expressions and whether or not they feel that the expressions are well-formed.

As already said, intuitive data can be obtained both through introspection and through experiment. In the former case the data are based on the intuitions of the linguist providing the data. This method has been criticized by a number of linguists, for example by Labov (1972), Featherston (2007) and Schütze (1996). Featherston and Schütze prefer experimental methods of gathering these intuitions. However, the use of introspective data is also defended by others (e.g. Newmeyer (1983: 48)) and continues to be an important source of data.

An experimental way of gathering acceptability judgments is letting informants fill in a ‘forced-choice questionnaire’ (Newmeyer 1983: 61)); hereby the linguist can obtain, in a controlled manner, the intuitions of a group of (native) speakers of a language. The data gathered in this way are experimental, intuition-based data. However, as was pointed out by Newmeyer (1983: 48), it should be kept in mind that even experimentally gathered intuitive data are introspective in the sense that they are about the introspection of the informants.

The method used in obtaining intuition-based data (i.e. introspection or experiment) matters in determining the validity of the data. However, experimentally and introspectively gathered data can be used in roughly the same way, that is, the have can have the same function in linguistic argumentation.

There are various types of intuition-based data. The examples of Schütze (1996) in section 1.3.1 contain a few. The fact that example (1d) *Alan Stephen punished is ill-formed on an indented reading is an example of an intuition regarding the meaning of this sentence, namely that it cannot be taken to mean Stephen punished Alan. This thesis, however, mainly focuses on what are called acceptability judgments, judgments of native speakers regarding the well-formedness of particular linguistic expressions. This type of intuitive data will be examined in more detail in the following chapter.

(18)

NATURALLY OCCURING DATA

DATA PRODUCED BY THE RESEARCHER

UTTERANCES INTUITIVE DATA

EXPERIMENTAL DATA

INTROSPECTIVE DATA

(19)

2 Grammaticality

Before moving to the notion of grammaticality it is first necessary to establish what is meant by grammar. The term grammar is somewhat ambiguous. It can refer both to the ‘linguist’s theory’ and to the ‘subject-matter of that theory’ (Chomsky 1990: 513). As the subject matter of a theory, grammar is something that exists in reality. However, what it is that exists in reality is not something that is generally agreed upon in the field of linguistics. Is grammar, for example, something that exists independently in the minds of speakers (as is claimed by GG)? 4 Or is it a social system of some kind? GG, FDG, and RCG have different opinions regarding the matter, although in one way or another they all see language as a system that exists in reality. Because of this complication the term grammar will not be used in this second sense (thereby incidentally following Chomsky (1990)).

A grammar in the sense of a linguist’s theory usually means a theory of a particular language. Such a theory can be proposed within the framework of a general theory of

linguistic structure (like the one proposed in Chomsky (1957/2002)). Because GG, FDG, and RCG propose very different general theories of linguistic structure, their grammars are dissimilar in at least a number of crucial ways. Chapter 3 will go into these matters in more detail.

The notion of grammaticality is tightly connected to that of grammar: the grammaticality of a linguistic expression is something that follows from a grammar. Consequently, the same ambiguity applies: grammaticality can either follow from the theory or from the grammar as it is assumed to exist in reality. Here, the term will be used in this first sense, in which case it can be understood as a ‘theoretical construct’ (Newmeyer 2007: 398).

In some general theories, grammars can assign certain values to linguistic expressions. The sentences in example (1) in the introduction (Stephen punished Alan etc.) can be

assigned the values grammatical and ungrammatical, for example. In which case, the asterisk or its absence would indicate (un)grammaticality. According to Newmeyer this is the right way of using asterisks (1983: 51).

The assignment of the values grammatical and ungrammatical is sometimes called a grammaticality judgment. Newmeyer argues against the use of the term grammaticality

4_{Cf. for example, Chomsky (1990: 514): ‘language is a system that is internalized in the}

(20)

judgment (1983: 51), because the word judgment implies that the intuitions of native speakers

are involved. However, according to Newmeyer, native speakers do not have direct access to their grammars (in the subject-matter of the theory sense) and are therefore not able to make judgments about grammaticality. Grammaticality follows from the grammatical theory and is not a judgment.

Native speakers do have intuitions about the acceptability of a linguistic expression, however. These are often called acceptability judgments. Many factors may be involved in these judgments and particular judgments may be based on other things than grammar. These acceptability judgments are often used as data. Newmeyer argues that the terms

grammaticality and acceptability should be sharply distinguished from one another (2007). However, in practice the term grammaticality carries in itself an own ambiguity (on top of the one inherited from the term grammar). It can refer to the being in accordance of a

linguistic expression with a grammar (theory), or to whether or not a native speaker finds the linguistic expression acceptable. That is, the term grammaticality is used for what can also be called acceptability judgments. In the present study, the term grammaticality will be used in the narrow sense advocated by Newmeyer.

Contrary to what according to Newmeyer should be the case, example sentences and their asterisks (like the ones in (1)) are often referred to as data and used as such in linguistic argumentation. If grammaticality is understood as a theoretical construct, the assignment of the asterisks would be something that follows from the theory. It might make more sense to treat the assignment of asterisks as acceptability judgments instead of a reflection of

grammaticality in those cases that examples and their asterisks are referred to as data, because this way the data are independent of the theory. This is what will be done in the present study.

The following figure shows the ambiguity of both grammar and grammaticality. What will be meant by them in the rest of the thesis has been made bold in the figure.

(21)

GRAMMAR

LINGUIST’S THEORY SUBJECT-MATTER OF

THAT THEORY GRAMMATICALITY IN ACCORDANCE WITH LINGUIST’S THEORY ACCEPTABILITY JUDGMENT OF NATIVE SPEAKER

Figure 2 Grammar and grammaticality

The remainder of this chapter will look at some notions closely tied to the notion of grammaticality. First the concept grammatical analysis will be examined and some

grammatical analyses in the spirit of GG, FDG, and RCG respectively will be given. Then, a specific method of linguistic argumentation will be introduced (namely the distributional method of argumentation). The terms acceptability and qualitative corpus data will be examined more closely in that section because they are closely related to this method of argumentation. The final section of this chapter will examine the relation between grammaticality and data (such as acceptability judgments).

2.1 Grammatical analysis

Grammatical analysis will be regarded here as the analysis of a linguistic expression in terms of a general theory of linguistic structure. Grammatical analysis describes the structure of a linguistic expression. A grammatical analysis is not just a theoretical artifact, something the linguist creates. It is supposed to reflect something in reality. Grammar can refer to both theory and subject matter of that theory (grammar as part of reality). Consequently, a grammatical analysis reflects what is assumed in the analysis to be the structure of the

(22)

linguistic expression as it exists in reality. For instance, according to Taylor many linguists ‘subscribe to the view that a linguistic analysis is a hypothesis about the mental

representation of language in the minds of its speakers’ (2007: 570). In the case of GG ‘structural descriptions’ are assumed to be produced by and in the minds of speakers

(Chomsky 1995: 167). The grammatical analyses of specific linguistic expressions proposed in GG are supposed to reflect these structural descriptions that are presumably to be found in the mind and ultimately perhaps even physically locatable in the ‘brain’ (Chomsky 1980: 5).

2.1.1 Three grammatical analyses

The general theories of linguistic structure associated with GG, FDG, and RCG are very different in nature. To illustrate some of the differences, the following paragraphs provide three analyses in the spirit of the three general theories. The FDG, and RCG analyses are mainly based on Hengeveld & Mackenzie (2008) and Croft (2001). The GG analysis is based on a number of texts. In the following analyses the example sentence (1a) of the introduction will be analyzed from all three perspectives. For convenience the sentence is repeated below.

Stephen punished Alan. (4)

In GG, a sentence like this one is analyzed as a phrase structure with some movement. The following analysis is adopted from Radford (1997: 389) and Hornstein, Nunes & Grohmann (2005: 99–102). The analysis stops at the TP level and the only features that are included here are two strong features that trigger movement, the strong [–V] feature on the light verb and the strong [–N] feature on T.

(23)

TP (5) DP Stephen T’ T [Past] [–N] vP DP Stephen [N] v’ v VP V punished v Ø [–V] V punished [V] DP Alan

Some of the crucial aspects of this representation are that Stephen is higher up in the structure and that Alan is part of the VP. Also, there is movement from V to v and from spec,vP to spec,TP. This movement, however, is not visible in this case, due to the fact that there are no constituents between where the movement starts and where it stops.

In FDG this example is analyzed in a very different way. Apart from the morphosyntactic structure, the semantic and pragmatic structure of the expression are also taken into account. It can also be analyzed in terms of its phonological structure, but this has been left out.

(24)

(MI: [(AI: [(FI: DECL (FI)) (PI)S (PJ)A (CI: [(TI) (RI: Stephen (RI)) (RJ: Alan

(6)

(RJ))] (CI))] (AI)] (MI))

The first example (6) shows the pragmatic or interpersonal representation of the

expression. It consists of a Move (MI) with one Discourse Act (AI), a declarative illocution, a

Speaker (PI)S, an addressee (PJ)A and the Communicated Content, in which there are one

Ascriptive Subact—the punishing— and two Referential Subacts—the two persons involved in the punishing, Stephen and Alan. Because these persons are specific and known to speaker and presented as known to the addressee they are included in the Interpersonal Level instead of the Representational Level.

The structure at the Representational Level is as follows:

(pi: [(past epi: [(ei: [(fi: [(fj: punish (fj)) (xi) (xj)] (fi))] (ei))] (epi)] (pi))

(7)

There is a past operator specifying the Episode (placing it in an absolute time frame). Furthermore the individuals and the activity in which they take part are stated. Both the representational (7) and interpersonal structure (6) are reflected in the morphosyntactic structure.

(Lei: [(Cli: [

(8)

Pm-1 (Npi: [(Nwi: Stephen (Nwi))] (Npi))

Pm (Vpi: [(Vwi: [(Vsi: punish (Vsi)) (Affi: ed (Affi))] (Vwi))] (Vpi))

Pm+1 (Npj: [(Nwj: Alan (Nwj))] (Npj)

] (Cli))] (Lei))

The Linguistic Expression takes the form of a Clause with an NP a VP and another NP. In contrast to the GG analysis, the second NP is not part of the VP. There is no TP between the three constituents and the clause level. Also, there is no movement of the verb. Inflection of the verb by means of an affix takes place within the Verbal Word layer (Vwi) and is a

reflection of the Past Operator that was present in the Representational Level. That is, the Past Operator is codified in the Morphosyntactic Level by means of the affix –ed.

In the FDG analysis there are language-specific mapping rules between the levels of representation. In this case the lexemes Stephen and Alan (introduced at the Interpersonal Level) and punish (introduced at the Representational Level), as well as the past operator (introduced at the Representational Level) are mapped onto the Morphosyntactic Level, using

(25)

the levels of representation (morphosyntactic alignment). First, the absolute medial position of the clause (Pm) is filled in with punish and the past operator is reflected by the affix –ed. Then Stephen is put in the position immediately in front of punished (Pm-1) and finally, Alan is put in the position immediately after punished (Pm+1).

An RCG analysis of the linguistic expression (4) would be given in terms of its place in a taxonomy of constructions. However, in RCG the question of whether or not a particular construction (abstract or substantive) is taken to be part of the taxonomy is an empirical question (Croft 2001: 57). A very simplified illustration of a relation between the

construction and the abstract Transitive Construction (or Monotransitive Construction) could be the following.

[TrSbj TrVerb TrObj] (9)

[Punish1 punish Punish2]

[Stephen punishes Alan]

There are three constructions represented here in a ‘fragment of a constructional hierarchy’ (Croft 2004: 643). The one at the bottom is the most substantive one, the

construction immediately above abstract away from his construction, and could be called the ‘Punish’ Construction, in which Punish1 punishes Punish 2. An even further abstraction is the Transitive Construction that sits at the top of this schematic representation, in which a Transitive Subject does something to a Transitive Object. The three constructions are related to each other as a whole, and the parts of the individual constructions are related individually to one another. The past tense of punish would be taken care of by another construction in the taxonomy. The actual construction would have multiple parents in this view. In other words, it would be an instantiation of multiple other constructions (not included in (9)).

What is important to note here is that there are no syntactic relations between the elements of the construction (as is very clearly the case in GG). Only relations between different constructions, between the elements and the construction as a whole, and between the elements of different related constructions are posited.

2.1.2 Issues concerning grammatical analysis

In all three analyses (the one in GG, FDG, and the one in RCG), it is important to keep in mind that grammatical analysis depends heavily on the analytical tools provided by the

(26)

general theory. According to Aissen & Bresnan even such basic notions as ‘direct-object’ and ‘transitivity’ are ‘theory-laden’ (2004: 110).

However, even if grammatical analysis is theory-laden, a grammatical analysis of a linguistic expression or set of expressions is not something that directly follows from a general theory. Rather, different analyses within the same general theory are often proposed by linguists working in that particular general theory. Hengeveld & Mackenzie (2008: 45) are very specific about the fact that in FDG different analyses can be proposed: ‘our analyses of particular phenomena […] should be taken as indicative rather than representing any claim to a definitive FDG statement (if such were even possible) about the phenomena in question’. This shows that—within certain boundaries—various analyses can be proposed. Moreover, the general theory does not impose a priori a particular analysis.

In the case of different theories of linguistic structure with an own metalanguage, differences between their grammatical analyses are highly visible, due to the fact that the analyses are formulated in different metalanguages. This does, however, not mean that these analyses cannot be translated into other metalanguages (see also Dryer 2006: 229). In fact, advocates of a given theory often adopt analyses from other theories. In Hengeveld & Mackenzie (2008: 298), for example, ‘an analysis of Warlpiri’ put forward by someone working ‘in a formalist framework’ (as opposed to FDG’s structural-functional framework) is seen as support for their own analysis of comparable languages.5 Another example is from Croft where an analysis by Nunberg, Sag & Wasow, 6 in which ‘idiomatically combining expressions’ (1994: 504) are analyzed as ‘semantically compositional’ (Croft 2001: 181), is used as an argument for doing away with syntactic relations in RCG. In each case, however, the analyses are translated into the structure of the new theory (see also Croft 2004: 648).

The fact that different grammatical analyses can be proposed within a single general theory also shows that general theories of linguistic structure are actually ‘moving targets’in the words of Mackenzie (2012: 423). This fact becomes even more problematic when one considers that the grammatical analyses are part of the link between general theory and distributional data, and as such provide a means of establishing the adequacy of the general theory. The changing grammatical analyses, and the fact that no grammatical analysis can be

5_{Note that this analysis is highly abstract: it is about a language in its entirety, not about an}

individual linguistic expression or a group of related linguistic expressions.

(27)

Head-considered a definitive analysis of the general theory—something which can also be said about GG and RCG—make for an (at least to some extent) problematic link between the general theory and distributional data.

2.1.3 Descriptive adequacy

Whether or not a structure of a linguistic expression proposed by a grammatical analysis is the right one is a question of what is sometimes called descriptive adequacy. This notion originated in Chomsky (1964),7 in which three ‘levels of success’ are sketched one of them being descriptive adequacy (Chomsky 1964: 29). Another notion introduced there is that of ‘observational adequacy’ (ibid.). Descriptive adequacy is a higher level of success than observational adequacy, amongst other things because it presupposes observational adequacy (Radford 1988: 28). Consider the following definition of descriptive adequacy.

A grammar of a language is descriptively adequate if it correctly specifies which sentences are (and are not) syntactically, semantically, morphologically, and phonologically well-formed in the language, and also properly describes the syntactic, semantic, morphological, and phonological structure of the sentences in the language in such a way as to provide a principled account of the native speaker’s intuition about this structure. (Radford 1988: 28)

The definition of observational adequacy given by Radford is almost an exact copy of the first part of the previous quote.

A grammar of a language is observationally adequate if it correctly specifies which sentences are (and are not) syntactically, semantically, morphologically, and phonologically well-formed in the language. (Radford 1988: 28)

From these quotes the conclusion can be drawn that descriptive adequacy is observational adequacy—the correct specification of well-formedness of a linguistic expression—coupled with a correct description of the structure of that linguistic expression.

According to Haspelmath the notion of descriptive adequacy ‘corresponds’ to that of ‘cognitive realistic description’ (2004: 575). This, however, makes it a very troubling concept if taken to apply to all general theories. There is a large amount of general theories (of which GG, RCG, and FDG form just a small part). All these theories describe linguistic structure in a different way (sometimes wildly different). If there is one right way of describing linguistic structure, there seems to be no way of firmly establishing what it is at the moment (cf. also Seuren 1998: 257). This is illustrated by the fact that even simple sentences receive such a different analysis under different general theories— as is apparent from the three example

(28)

analyses in 2.1.1. Furthermore not all linguists subscribe to the idea that a right description of the structure of a linguistic expression is one that resembles how it is cognitively represented. Descriptive adequacy remains an important and often-encountered concept, however, and sometimes receives other interpretations. It is crucial for grammatical analysis. The emphasis of this thesis, however, lies on grammaticality, for which only the concept of observational adequacy is of direct relevance.

2.2 The distributional method of argumentation

The distributional method was crucial in American structural linguistics (cf. Harris 1951: 5) and still is widely used in linguistics (Croft 2001: 11). It can be used to support linguistic claims, but also to criticize. As such it is an important ‘method of argumentation’ (ibid.). This method is characterized by analyzing linguistic expressions with respect to the behavior of the elements of which the expressions are made up. These elements are placed in various linguistic contexts, that is, in a number of linguistic expressions structured in a different way. In some of the linguistic expressions the elements are structured in such a way that the resulting linguistic expression is well-formed, while in others the elements are structured in such a way that the resulting expressions are ill-formed. Carefully analyzing sequences of these well-formed and ill-formed expressions reveals something about the behavior of the parts that make up linguistic expressions in a particular language. Such an analysis is called a ‘distributional analysis’ (Croft 2001: 11; cf. also Chomsky 1964: 57 for an example).

A classic example is the use of distributional analysis in distinguishing different parts of speech in theories where the ‘morphosyntactic behavior’ of individual words is used to determine to what parts-of-speech category they belong (cf. Croft 2001: 63–64). The morphosyntactic behavior of parts-of-speech can be determined by distributional analysis, e.g. by placing a word in different linguistic contexts and determining in which contexts the linguistic expression is well-formed and in which contexts it is ill-formed. The following examples are used in Croft (2001: 12) to illustrate how distributional analysis works.

(29)

a. Jack is cold. (10) b. *Jack colds. a. Jack is happy. (11) b. *Jack happies. a. *Jack is dance. (12) b. Jack dances. a. *Jack is sing. (13) b. Jack sings.

In these examples, the morphosyntactic behavior of cold is the same as that of happy;

dance and sing, however, behave differently. The (a) examples are well-formed for cold and happy and ill-formed for dance and sing, while the exact opposite is the case in the (b)

examples where dance and sing are well-formed and cold and happy are ill-formed. The difference in distribution that is apparent in examples (10) to (13) can be used as criteria for distinguishing the categories verb and adjective.8

Another example is from Mackenzie (2009: 18). Here, instead of defining parts-of-speech categories, the distributional analysis reveals the behavior of a particular word.

a. *The clown failed to never amuse the audience. (14)

b. The clown never failed to amuse the audience.

The first sentence is ill-formed, as is indicated by the asterisk preceding (14a); (14b) on the other hand is considered well-formed and consequently is not preceded by an asterisk. Mackenzie uses this pair of sentences to show how the word fail behaves, i.e. what

complement it can take, and, more importantly, what complement it cannot take. In this case the complement to amuse the audience cannot be negated, doing so results in an ill-formed sentence (12a). In other words, fail cannot take this kind of negated complement.

Mackenzie draws a conclusion from a series of well-formed and ill-formed linguistic expressions about the behavior of words or groups of words within these expressions (in this case the behavior of the verb fail, in the sense of what complement it can take). The

8_{It is important to note that Croft (2001) actually argues against the use of such criteria in}

(30)

examination of a set of linguistic expressions of a particular language (well-formed and ill-formed ones) in terms of where within the structure of these expressions particular linguistic elements of this language can occur, can be called distributional analysis because this type of analysis looks at the distribution or behavior of these linguistic elements, that is, where within the structure of a linguistic expression they can occur and where they cannot occur. If the location of such a linguistic element within a linguistic expression results in a well-formed expression, it follows that the element can occur in that particular place. If the place of an element leads to an ill-formed expression, the element cannot occur in that particular place.

2.2.1 Acceptability judgments

Data used in distributional analysis come essentially in two forms. On the one hand, one can use a particular type of intuitive data called acceptability judgments. This type of data can be gathered both through introspection and through experiment (as is the case with other types of intuitive data). The acceptability of a linguistic expression is the being or not being acceptable of this particular linguistic expression for a native speaker. An acceptability judgment is an attitude of a native speaker towards a linguistic expression and therefore a behavioral datum.

As has already been mentioned, acceptability judgments are sometimes also called grammaticality judgments. However, following Newmeyer (1983: 48), the two are

distinguished from each other here. Grammaticality, on the one hand, is a theoretical notion referring to the being in line or not being in line of a linguistic expression with a particular grammar. Acceptability, on the other hand, has to do with the intuitions native speakers have regarding the well-formedness of a linguistic expression within the context of their language.

2.2.2 Qualitative corpus data

The previous section was concerned with data produced by the researcher. There are naturally occurring data that can also be used in distributional analysis: qualitative corpus data. The occurrence of a linguistic expression, that is, whether it occurs or does not occur in a given corpus, can be used in much the same way as the acceptability of a linguistic

expression: it can show the distribution (or behavior) of the elements of that linguistic expression.

The occurring of a linguistic expression in a corpus is called ‘positive evidence’ while the lacking of linguistic expressions from a corpus is called ‘negative evidence’ (Penke &

(31)

Rosenbach 2004: 486–487). The problem with negative evidence, however, is that it has a ‘meager epistemological status’ (ibid.). So for those working with corpus data, acceptability judgments (of expressions that are regarded as ill-formed) can be a useful addition to the kind of data they can find in a corpus. The following example from Keizer (2010: 52) illustrates how acceptability judgments can be used alongside positive evidence from a corpus.

a. *I couldn’t put up with his paranoia and up with his (15)

megalomania.

b. *He went back on his word and along with our proposal. a. Turner went back to skating and back to school at Northern (16)

Michigan University. (COCA)

b. Every night after my mother falls asleep, I climb out of the fire escape and up to the roof. (COCA)

The examples (16a) and (16b) are taken from the Corpus of Contemporary American English (COCA), while (15a) and (15b) are intuitive data. The distributional analysis in this case reveals that, what Keizer calls ‘composite predicate constructions’ (ibid.: 50), examples (15a) and (15b), differ from she calls ‘appositional constructions’ (ibid.: 51), examples (16a) and (16b), in the way they can be structured. In the first case coordination is not possible, while in the second it is.

Examples (15) and (16) show that both intuitive data and qualitative corpus data can be used in distributional analysis. By using both types of data, some of the problems associated with intuitive data can be overcome; while at the same time, some of the shortcomings of negative evidence can be overcome by using acceptability judgments.

The term distributional data could be used to refer to the function of the data instead of their origin. That is, while the terms acceptability judgment and qualitative corpus data signal the origin of the data, distributional data would refer to the fact that the data are used in distributional analysis.

The examples (15) and (16) from Keizer (2010) illustrate that there does not need to be a dichotomy between introspective (or intuitive) and corpus data in linguistic practice. When one focuses on the function of the data instead of on the source, both kinds of data seem remarkably similar when used in distributional analysis.

In the examples offered in this thesis, only FDG uses both acceptability and occurrence data at the same time. In fact, only FDG uses occurrence data at all in the examples provided;

(32)

there are no examples of GG and RCG using occurrence data in this thesis. This does by no means mean that qualitative corpus data cannot be used in GG and RCG.9

Distributional data, that is, acceptability judgments and qualitative corpus data used in distributional analysis, accompany specific linguistic expression. In the case of acceptability this is straightforward, as the judgment itself is just that, an assignment of a value to a particular linguistic expression. Also in the case of occurrence, the status assigned to a linguistic expression (whether or not it occurs) is essentially different from the linguistic expression itself. Given the fact that distributional data accompany specific linguistic expressions, and given the fact that these expressions belong to a specific language, distributional data are language-specific.

A final remark that should be made regarding distributional data is that the linguistic expressions that are accompanied by the distributional data are not grammatically analyzed, which is something a theory does. The linguistic expressions, as far as distributional analysis is concerned, just show linear order and are not (hierarchically) structured in a

theory-dependent way. Not even segmentation (also a form of analysis) has to be presupposed, although in practice it often is.

2.3 The relation between grammaticality and data

As seen in the previous sections, a grammar can be understood as a theory modeling a language system that is assumed to exist in reality (e.g. in the mind or in society). Similarly, a grammatical analysis of a linguistic expression can be seen as a reflection of the way a

linguistic expression is assumed to be structured in reality (e.g. as a result of how the expression came into being computationally in the mind, or as a result of the purpose it fulfills in communication). Grammaticality, then, can also be seen as a theoretical construct that reflects reality: the language system as it exists in reality can be understood as something that produces certain linguistic expressions and cannot produce others. Acceptability

judgments and occurrence are manifestations of this ability and inability of producing certain linguistic expressions in a language. Consequently, the model of grammaticality provided by

(33)

a grammar and the associated grammatical analyses can be constructed using these manifestations and (perhaps more importantly) be tested by these manifestations.

Following the definition Radford (1988: 28) gives, the term observational adequacy can be used to designate the criterion that such a model of grammaticality should be in

accordance with the distributional data (i.e. data regarding acceptability and occurrence). These are the manifestations of the well-formedness of linguistic expressions.

The accordance of general theories of linguistic structure, grammars, and grammatical analyses with reality can be determined in many ways. External, psycholinguistic data can be used for example to see how a cognitive model of language (which is also a general theory of linguistic structure) reflects cognition as it is manifested in the psychological experiments and observations. An especially linguistic way of dealing with this relation, and the one that is of interest here, is the relation between grammaticality and distributional data. However, grammaticality is a very complex notion and it is construed very differently in various linguistic general theories. It is this difference between GG, RCG, and FDG that will be examined in chapter 3. In the remaining paragraphs of this chapter some general remarks about grammaticality will be made.

Grammaticality can be explicit (this is what is meant by generative) to certain degrees, and depends on what (formal) metalanguage is used. Usually a certain grammatical analysis is presupposed, as grammaticality often needs to be determined on the basis of a grammatical analysis, such as what parts of speech the words in a linguistic expression belong to and how the linguistic expression is structured. On the basis of a structural analysis of a linguistic expression, and on the basis of what language-specific rules the grammar proposes, the grammar can come to a grammaticality verdict regarding the expression in question. However, the structural description of a linguistic expression provided by grammatical analysis matters only to the extent that it facilitates grammaticality. It does not matter, for the present purposes, whether or not it is the right structural description (which would be a matter of descriptive adequacy). In this thesis the focus is not on descriptive adequacy but rather on observational adequacy or the ‘external conditions of adequacy’ Chomsky 1957/2002: 49) of a grammar. Grammaticality, in other words, must be in accordance with what is well-formed in a language and what is well-formed in a language can be established by looking at

acceptability judgments and data regarding the actual occurrence of linguistic expressions. If the grammaticality predicted by the grammar overlaps to a substantial extent with the acceptability judgments or the occurrence/nonoccurrence of actual linguistic expressions, the grammar is observationally adequate. The correspondence between grammaticality and

(34)

acceptability/occurrence is thus related to the notion of observational adequacy. There is no direct relation between distributional data and to the general theory of linguistic structure (although the general theory can be evaluated on the basis of the individual grammars it produces). Distributional data are language-specific and to account for it a language-specific notion of grammaticality is needed.

Distributional data (i.e. acceptability judgments in the sense of Newmeyer (1983, 2007) and occurrence) can be freely used across different general theories, because, as has been said, they are not structured in a theory-dependent way. In this way they can really be seen as data, as theory-externally given facts, or in Chomsky’s words as ‘external conditions of adequacy’ (1957/2002: 49).

However, the notion of grammaticality is very theory-dependent. Given the fact that GG and FDG have their own formal metalanguages for structural description (i.e. grammatical analysis) and GG, RCG, and FDG have their own notion of grammar, as will be elaborated in the next chapter, one might expect that their notions of grammaticality are different.

A general theory of linguistic structure is about linguistic structure in general, and not about the structure of any one particular language. Grammars, in the sense of theories of particular languages that are proposed within the framework of a general theory, can give an account of cross-linguistic variation for that general theory. Whether or not a general theory is capable of accounting for variation and similarities across languages is a question of ‘typological adequacy’ (Dik 1997: 14): the tools provided by the general theory for the individual grammars need to be versatile enough to allow the grammars proposed within the general theory to be in accordance with the facts of the languages they describe. The general theory itself must be able to account for the similarities between languages, and reflect the generalizations that can be made about language.

The three general theories that are of concern here (GG, RCG, and FDG), all give their own account of cross-linguistic variation and similarity—although they differ in where they put emphasis. On the one hand, the three theories all have to give an account of what is determined language-specifically (that is, where cross-linguistic variation can occur) and what needs to be formulated in terms of particular grammars of particular languages. On the other hand, they need to capture within their general theory where languages are similar. In other words, they need to specify how language in general isstructured.How the division of labor between general theory and language particular grammar is construed differs between GG, RCG, and FDG. These differences will be discussed in chapter 3.

Grammaticality in Generative Grammar, Radical Construction Grammar and Functional Discourse Grammar