Language Learners' Extrapolation of Word Order in Micro-Artificial Languages
Vivian Cook
Draft of paper in Language Learning, 38, 4, 497-529, 1988
Abstract
This article is concerned with the ability of language learners to 'extrapolate' from the word order in one type of phrase to that in another, namely from the order in the Verb Phrase to the orders in the Prepositional and Noun Phrases, and from the Verb Phrase order plus either the Prepositional Phrase order or the Noun Phrase order to the other phrase type. Such extrapolation relates both to the implicational universals of possible word orders in language described by Hawkins and to the head parameter of Government/Binding Theory. An experiment with eight Micro Artificial Languages (MALs) is described that tested the ability of 409 secondary school children to extrapolate in this way. The results showed that most learners (340) were highly consistent in ascribing word order, and most succeeded in learning the MALs (321). Four extrapolation strategies followed by the learners are described that are not covered by the syntactic analyses mentioned: preferences for: (1) postpositions, whatever the VP order, (2) adjectives in the same position as Objects in the VP, (3) prepositions the same side as adjectives in the NP and Objects in the VP, (4) adjectives before nouns if the VP order and the NP order are consistent. Further research is needed to see whether such extrapolation strategies occur in real learning or are an artefact of the MAL approach.
This paper investigates the belief that learning the word order for one type of phrase has a knock-on effect on learning the word order for other types of phrase; this is here called 'extrapolation' to avoid the presuppositions of such alternative terms as 'generalisation' and 'transfer'. Languages vary in the position of various items in phrases; put in terms that are fairly neutral between syntactic theories, Verb Phrases either have Object Noun Phrases after Verbs, as in English, or before them, as in Japanese and Hungarian; Noun Phrases either have Adjectives before Nouns, as in English, or after them, as in Arabic and Malay; Prepositional Phrases either have Prepositions before Noun Phrases, as in English, or after them, as in Finnish and Hindi (in which case they are known as Postpositions). If learners know the order of elements in the Verb Phrase (VP), do they extrapolate to the order in the Preposition Phrase (PP), and to that in the Noun Phrase (NP)? Presented with a language that has Subject Object Verb (SOV) order, or Verb Subject Object (VSO), are they likely to put adjectives before or after nouns in the Noun Phrase, Adjective Noun versus Noun Adjective, and prepositions before or after nouns in the Preposition Phrase, Preposition Noun or Noun Postposition? Furthermore, is extrapolation easier if learners know the order in two types of phrase rather than one alone: given the order in both the Verb Phrase and Noun Phrase, will they extrapolate to the Prepositional Phrase? Given the order in both the Verb Phrase and the Prepositional Phrase, will they extrapolate to the order in the Noun Phrase? The four main questions we are concerned with are then: do learners extrapolate:
i) from the order of elements in the Verb Phrase to that in the Prepositional Phrase (Prep N versus N Post)
ii) from the order of elements in the Verb Phrase to that in the Noun Phrase (Adj N versus N Adj)
iii) from the Verb Phrase order plus the Prepositional Phrase order to that in the Noun Phrase (Adj N versus N Adj)
iv) from the Verb Phrase order plus the Noun Phrase order to that in the Prepositional Phrase ((Prep N versus N Post)
The interpretation of these questions is complex because of the different ways in which this area of syntax is analysed in terms of the implicational universals of Hawkins (1983) and the Universal Grammar of Chomsky (1986a). Nor has the existing literature on the acquisition of word order provided an overall answer to the question of extrapolation in these phrases, most of it having concentrated on one or other of the phrase types rather than the range of the above questions, e.g. Meisel, (1986), Clahsen and Muysken (1986), and Lujan, Minaya, and Sankoff (1984).This paper starts from the neutral ground of these four questions; it interprets them in terms of two contemporary models of language and language acquisition; it reports an experiment to investigate extrapolation; and it discusses the implications of the results.
Implicational universals of word order and language acquisition
The approach to be considered first is that proposed by John Hawkins, most extensively in Hawkins (1983). The essential claim, based on a sample of some 336 languages, is that, if a language has a particular word order in two types of phrase, it must have a particular word order in a third. Such 'implicational universals' express generalisations that are true of all languages. The four universals relevant to the present argument concern the relative order of Object NP and Verb, of Adjective and Noun (Adj N), of Genitive and Noun (Gen N), and of Preposition and Noun (Prep/Post N); these are couched without reference to the Subject of the sentence.
I. 'If a language has OV order, then if the adjective precedes the
noun, the genitive precedes the noun' (Hawkins, 1983, p.64), i.e.
OV > (Adj N > Gen N)
An example of this is that languages with SOV order and Adj N order have Gen N order, as does Japanese. One reason for expressing this universal without reference to the Subject is to include the recently discovered but rare Object-initial languages.
II. 'If a language has verb first order, then if the adjective follows
the noun, the genitive follows the noun' (Hawkins, 1983, p.66), i.e.
V1 > (N Adj > N Gen)
So languages with Verb Object order and Noun Adjective order have Noun Genitive order, as in Arabic. Again, as the Subject is not referred to, this universal covers both VOS and VSO.
III. 'If a language has Prep word order, then if the adjective follows the noun, the genitive follows the noun' (Hawkins,1983, p.66), i.e.
Prep > (N Adj > N Gen)
As in Greek.
IV. 'If a language has Postp word order, and if the adjectiveprecedes the noun, then the genitive precedes the noun' (Hawkins,
1983, p.67), i.e. Post > (Adj N > Gen Adj)
As in Chinese.
These four universals come in pairs, the even-numbered being the corollary of the odd. While the relationship between the Object position in the VP and the Noun Phrase position in the PP is not itself captured in these four, these are related through Hawkins' broader concept of Cross Category Harmony (Hawkins, 1982): the more a language is consistent, the more it is preferred; 'there is a quantifiable preference for the ratio of preposed to postposed operators within one phrasal category (i.e. NP, VP/S, AdjP, AdpP) to generalise to the others' (Hawkins, 1983, p.134).
Hawkins' statement of his implicational universals usually takes the form:
If P and Q then O
A language that has order P and order Q must have order O. This does not rule out a language that has P without Q, or Q without P, or ones that have Q and O without P, or indeed one that has both P and O; what it precludesis P and Q without O. That is to say, the following possibilities occur, using the asterisk to denote languages that should not be found: PQO, QO, P, Q, O, PO, *PQ.
Hawkins (1987) suggests that children always speak possible human languages; they never have a stage that breaches the universals, i.e. *PQ. The child's language at all times must be one of the permissible combinations; 'At each stage in acquisition, PreAdult Languages and Interlanguages remain consistent with implicational universals derived from current synchronic evidence.' (Hawkins, 1987, p.457). Thus the child may start with Q, go on to QO and end up with PQO, but the child cannot go from P to *PQ to PQO. They must, so to speak, go from right to left; they cannot know the left two items without the third .
Hawkins (1987) argues from acquisition data that children do indeed respect implicational universals: they never have impossible languages. His data do not however concern the three-term word universals ("If P and Q then O") universals but are based on two-term universals ("If P then Q") of fricative consonants and conditional clauses. He claims that, if a language has conditional sentences of the form "If X then Y", it also has conditional sentences of the form "X and Y". English children progress for instance from conditionals expressed by "and" ("You do that again and I won't like you") to those using "if" ("If you do that again I won't like you."); children do not go through an illegal stage when they have "if" sentences but not "and" conditionals; they do not violate the implicational universal. This is supported by evidence presented by Lujan et al (1984, p.369) 'that innovations are constrained by observance of linguistic universals, thereby disqualifying the notion that the change process must be initiated by a violation of a synchronic universal.'
If correct, there are nevertheless several other possible interpretations. One is that it is always a property of the data that the children encounter; they all hear actual human languages that have precisely this property; they have all heard examples of most of the relevant evidence for all the phrase types for the language since natural language input is not restricted, at least for the core linguistic phenomena that we are concerned with here. The sequence of acquisition may also be confused by the various performance restrictions on the first language learner. If P is always a higher syntactic category or a more complex construction, there may be short term memory reasons why children never produce it before the lower or simpler category. Though this may not concern the phonological arguments, it clearly applies to Hawkins' syntactic example; producing a conditional sentence with "and" - "You do that and I'll knock your block off" - allows the child to use two connected clauses rather than one more complex clause, a sequence well attested from the literature (Sinclair, 1969). To take the present examples, producing SVO involves constructing three phrases, minimally three words; producing Adjective Noun requires one phrase, minimally two words. It would not be too surprising if children produced Adjective Noun or Preposition Noun before SVO, as LARSP (Crystal, Fletcher, and Garman, 1976) and the Bristol study (Wells, 1985) seem to suggest.
The Hawkins argument has been elaborated at length because in many ways it differs from the links between language universals and acquisition usually made in the literature. The type of universal that has received most attention in L2 learning has been the Accessibility Hierarchy of Keenan and Comrie (1977) which claims that there is a series of positions for relativisation:
Subject > Object > Indirect Object > Object of Preposition > Genitive > Object of Comparison
All languages start at the left and have subject relative clauses; some go one point along and have object clauses; others go all the way along. Learning research has mostly interpreted this as a sequence of acquisition; learners should acquire relative clauses from left to right in the series, given or take one or two oddities and L1 differences (Gass, 1979; R. Hawkins, 1987). The Hawkins interpretation precludes the learner at any time from speaking a language that has gaps in the hierarchy, say, having Genitive relative clauses but no Indirect Object relative clauses; his argument concerns possible synchronic states of language, not sequences of acquisition. Only in Eckman et al (1988) is something approaching Hawkins' logic applied to the Accessibility Hierarchy.
Far from addressing our four questions, Hawkins' universals are neutral about extrapolation; the claim is instead that the learner at no stage knows an impossible language. Our four extrapolations involve the following stages: (i) V -> P, (ii) V -> A, (iii) V & P -> A, (iv) V & A -> P, none of which are disallowed. If learners extrapolate in this way, it is not particularly a consequence of his theory.
Nevertheless there are still predictions from Cross Category Harmony; learners should tend to prefer Cross Category Harmony if it has the importance that Hawkins attaches to it; Lujan et al (1984, p.363) indeed found 'the CCH principle appears to be the single strongest influence on the children's language, moving it in the direction of those language types that exhibit a more harmonious balance in their operand/operator basic orderings'; it would be odd if learners went against what it is allegedly one of the main constraints on language types and on the development of languages. If there is a tendency towards 'harmony', learners should extrapolate from Verb Object order in the VP to Prep Noun in the PP and N Adj in the NP; from Object Verb order in the VP to N Post in the PP and Adj N in the NP.
An alternative, but not entirely exclusive, approach to word order is found in the head parameter of X-bar syntax within Government/Binding (GB) theory (Chomsky, 1986a; 1986b; 1988). Using X to stand for any of N, A, P, or V, the maximal phrase (X'') consists of a possible specifier and a head of a particular type, either X' or X''; this head in turn consists of a head that is a lexical category (X), and possible complements projected from the lexical entry. Thus a Noun Phrase in English (N'') such as "the belief that the earth is flat", consists of a specifier, "the", and a head (N'), "belief that the earth is flat"; this head consists of a lexical category (N), "belief", and a complement, "that the earth is flat". All the phrases in English - Verb Phrases (V''), Adjective Phrases (A''), Prepositional Phrases (P''), Inflection Phrases (I''), and Complement Phrases (C'') - conform to this structure (Chomsky, 1986b), as do the phrases of other languages. Word order variation between languages comes down to whether the complements within the phrase occur before or after their heads, i.e. complement X versus X complement; this is the head parameter. A major claim of the theory of X-bar syntax is that the position of the complements relative to the head can be stated once and for all in the phrase structure of any language: languages are 'consistent'.
The Universal Grammar (UG) theory of acquisition claims that the expectation that phrases conform to X-bar syntax and the constraints on variation are built-in to the mind. To learn English, the child needs to discover that complements occur on the right of heads; to learn Japanese, that they occur on the left; everything else about word order follows from this discovery. To do this, they need comparatively minor 'triggering' evidence; a phrase or two of a language will establish which setting for the head parameter is appropriate;
'These facts can be determined from very simple sentences, for example the sentence "John ate an apple" (in English) or "John an apple ate" (in Japanese). To acquire a language, the child's mind must determine how the switches are set, and simple data must suffice to determine the switch settings, as in this case.' (Chomsky, 1987)
In our terms the learner should immediately extrapolate from one phrase type to the others.
The theory is neutral whether all the stages of language acquisition are equally subject to UG; one possibility is that all the principles of UG are simultaneously present from the beginning, called by Chomsky (1988) the 'no-growth' position, by others the 'continuity hypothesis' (Borer & Wexler, 1987; Pinker, 1984). A related question is whether the child ever possesses 'wild' grammars that do not conform to UG at particular developmental stages (Goodluck, 1986); in other words whether all the principles of UG are simultaneously present from birth. The 'wild grammars' position maintains that the various principles appear in the mind one after the other; earlier stages of the child's language development may breach those principles that are not yet manifest in the child's mind; this position holds that the language faculty itself matures, as for example permanent teeth appear after milk teeth. A crucial difference is the existence of stages at which the child's grammar does not conform to UG - the presence of 'wild' grammars; further discussion can be found in Cook (1988).
To come back to our four specific extrapolations, the head parameter analysis suggests that the learner should extrapolate the Verb Phrase order to the Prepositional Phrase (i and iv); it has no particular predictions for the adjective/noun order in the Noun Phrase (ii and iii), since the adjective in English is outside the X-bar specification above in being neither a specifier nor a complement. The X-bar analysis differs also from the implicational universals in that the former implies a two level structure to the phrase - specifier and head, and head and complements - while the latter uses a single level of structure. This creates a difficulty with statements involving the Subject of the sentence - SOV, VSO, and so on. An analysis such as SVO implicitly divides the sentence into three constituents, S, V, and O. GB syntax however currently maintains the traditional two-way division between the Subject NP and the VP, and a second division of the VP into Verb and Object NP; 'there is a subject-object asymmetry in that the verb and the object form a single phrase, whereas the subject is a distinct phrase' (Chomsky, 1988, p.70); technically the Subject NP is the compulsory specifier of I'', the VP the complement (Chomsky, 1986b). X-bar syntax therefore needs a special way of handling VSO and OSV languages since the S appears in a position where it interrupts the O and V; such word orders are treated as Verb-movement from an underlying structure in which the S occurs outside the VP.
Methods of investigation
The relationship between these questions and empirical evidence from actual learners is hard to establish. On the one hand the languages that learners encounter, i.e. their linguistic input, are neither 'pure' nor altogether consistent in their word order, e.g. English has a handful of apparent Adjectives that come after the Noun ("books galore"). The evidence available in real-world learning of an actual language is difficult to assess since the frequency and nature of the examples the learner has encountered is not known. On the other hand the development of other aspects of the child's mind may obscure linguistic progress; data collected from observation may represent performance rather than competence, and so on. Furthermore, because children hear a range of structures, they are never in the position of extrapolating from one phrase to another, totally unfamiliar, type of phrase; rather they are carrying over a word order to a phrase they have already heard to some extent. Nor has any substantial body of research tackled the overall question of word order in acquisition. Barbara Lust (1983) has shown children's sensitivity to one aspect of word order, namely Principal Branching Direction. General observational studies such as LARSP (Crystal et al, 1976) or the Bristol study (Wells, 1985) can be raided to find incidental observations. But for the most part it has been covered by such pronouncements as Brown (1973) 'the child's first sentences preserve normal word order.'
The argument has sometimes been advanced that second language learning can provide a better testing ground for such ideas than first language development (Gass & Ard, 1980). The comparative maturity of the L2 learner permits a greater variety of investigative techniques and eliminates effects caused by the children's lack of cognitive capacity. But here too we have imponderables in that L2 learners are exposed to a variety of data not available in the first language - explicit explanation, dictionaries, correction, all feature for example in the classroom that many of them face. The precise nature of extrapolation has not however been studied in L2 learning except through a study by Alison Henry (1986), which provided the inspiration for the present research; she found that English-speaking learners of Chinese extrapolated from the Object Verb order in the VP to the prenominal position of relative clauses, which they had never actually encountered.
The alternative to data from either L1 learning or normal L2 learning that is explored here is the learning of a Micro-Artificial Language (MAL). Recently interest in this paradigm has revived and it has been used to investigate aspects of language acquisition as diverse as 'universal operating principles' (MacWhinney, 1983), the contribution of bracketted input to acquisition (Morgan, 1986), and the importance of syntactic 'markers' to acquisition (Valian & Coulson, 1988). Arguments for using MALs will be found for instance in Smith and Braine (n.d.), McLaughlin (1980), Morgan and Newport (1981), and Morgan (1986), and will not be presented at length here. Basically in the present context a MAL gives greater precision. The sample of the language that the learners encounter is known since they have no other source of evidence; in the case of word order, the language they hear can be fully consistent rather than having the exceptions that are found in natural languages such as English. The language can also be 'pure' in the sense that it can focus on a particular point of syntax rather than being the mixture of different types of sentence and phrase found in natural language input; word order can be isolated from everything else in the language, excluding not only other aspects of syntax but also phonology and vocabulary. The evidence that the learner has encountered is also limited and totally controlled; research into extrapolation by learners of natural languages cannot overcome the disadvantage that they are quite likely to have heard examples of the new phrase type; in a MAL this is prevented. It might be that children do not entertain impossible languages, not because they cannot do so, but because the language evidence they receive precludes them, being for the most part possible sentences.
There are of course objections to the use of MALs as being artificial and arbitrary. Even if it is assumed that language is acquired through a specific language faculty, this does not imply that language-like knowledge cannot be acquired by other faculties of the mind; after all grammatical explanation in the second language teaching classroom rightly or wrongly has aimed at creating language knowledge through non-language means. The very 'purity' of the experiment may cut it off from the rich and complex experience of real language learning. Evidence from MAL learning cannot be assumed necessarily to show the progress of natural language acquisition. Results from MAL research in general or from a single MAL experiment using a particular methodology cannot be generalised to the learning of natural languages in natural situations unless they are confirmed by evidence from natural language learning, in so far as this is possible.
For methodological reasons it was necessary to adapt the precise form of the extrapolations to be tested. First of all in order to give the learners plausible input, they could not be presented only with Verb Phrases; rather they needed to be given sentences which also had a Subject Noun Phrase. Secondly the possible learners for the experiment were speakers of English and therefore needed to be tested on non-English languages. As English has Subject Verb Object order, the two main types of MAL had to be Verb Subject Object and Subject Object Verb; it should be noted however that VSO is in some ways closer to the English SVO if the relative positions of Verb and Object are in question, ignoring the Subject. Furthermore, while VSO is found in only 9% of the world's languages, SOV is is found in 45% (Tomlin, 1984); the choice between the two languages may be biassed in some way. Thirdly, while it would have been attractive to test the learning of the Genitive, this seemed to require morphological or syntactic complexity that would involve a longer and more difficult task for the learners.
The basic design was to construct two sets of related MALs. A sample of possible sentences in each MAL was given to the Ss; a further sample of the same MAL was used for testing acquisition of the MAL; and a sample incorporating a phrase type not previously encountered tested their extrapolation.
The MALs in Set 1 tested extrapolation from a sample that showed Object position in the VP to Adjective position in the NP or to Noun Phrase position in the PP (questions i and ii). It had two extremes of SOV (MALs A and C) and VSO (MALs B and D), both of which were tested with extrapolation to the NP order, Adj N versus N Adj (MALs A and B), and the PP order, Prep N versus N Post (MALs C and D). The language that the learners were given and the tests of learning and extrapolation were 'pure' in that they involved no other aspect of syntax than the order of words in the phrase types; no other clues were available to the learners, such case-marking, etc. One justification for this is the Independence Principle, which claims that a single parameter may be validly studied in isolation from the others (Wexler & Manzini, 1987).
Set 2 tested the extrapolation from a sample that had two types of phrase, that is to say, Object position in the VP plus either Adjective position in the NP or NP position in the PP to the remaining phrase (questions iii and iv). Again there were four forms: SOV and N Post (MALs E and G), and VSO and Prep N (MALs F and H), which were extrapolated to the NP order (E and F), and to the PP order (MALs G and H).
The vocabulary for the experiments was drawn from the same pool of pronounceable monosyllables used by Morgan and Newport (1981). Arbitrary meanings and syntactic categories were assigned as follows:
Nouns: BIF (tiger), HES (lion), MIK (cow), VOT (dog), RUD (cat)
Verbs: NEB (sees), SOG (hears), KAG (follows)
Adjectives: JAX (big), TIZ (old), FAC (clever), CAV (strong)
Prepositions: DUP (near), PEL (to), LUM (on), KOR (with)
The MAL experimental paradigm that was used relied upon sentences having meaningful reference, as the bulk of the research suggests is necessary (Meier & Bower, 1986; Nagata, 1981); referential meaning was supplied through English translation equivalents rather than through the geometric shapes used in Meier and Bower (1986), Morgan and Newport (1981), and most other studies.
The presentation phase consisted of the class teacher reading aloud 36 sample sentences of the language with literal word order translations and grammatical English translations: e.g. from MAL A:
BIF HES SOG (tiger lion hears) the tiger hears the lion
The purpose of reading the sample aloud was to maintain the attention of the Ss, who had the full text in front of them. For Set 1 MALs the Ss were presented with 30 sentences of the appropriate language in this way, for Set 2 MALs with 26 sentences, the difference being due to the need to balance equal occurrences of both phrase types. The odd and even pairs in Set 1 (i.e. A/C and B/D) had the same original samples. The vocabulary was introduced in a sequence so that each new item was heard several times in the context of old items before another item was introduced. All vocabulary items had equal frequency; the sentences were otherwise randomised in order. The Ss were given two tests:
i) Test 1: whether they had learnt the word order of the language. Ss had to choose from 15 multiple-choice answers on their answer sheets. In the case of Set 1 MALs, each question was a three-way choice between the three possibilities for word order, VSO, SVO and SOV; i.e.
a) BIF MIK SOG (tiger cow hears)
The tiger hears the cow
b) BIF SOG MIK (tiger hears cow)
c) SOG BIF MIK (hears tiger cow)
The three possible answers were rotated so that each order of presentation occurred equally. Similarly Set 2 additionally tested prepositions through a two-way multiple choice between both possible positions, i.e. Prep N and N Post, or N Adj and Adj N, with both orders given equally. All Set 1 MALs had the same answer sheets; different sentences would be 'right' for each of them. Set 2 MALs necessarily had different answer sheets for this test.
ii) Test 2: whether they could extrapolate to a phrase type they had not encountered before.
For this test the students had to read the English sentences silently and decide on the best translation. Again 14 multiple choice questions asked them to decide between the two alternatives for the novel construction, either Adjective or Preposition position. As each pair of groups (A/B, C/D, E/F, G/H) had the same answer sheet with different answers being "right", it was possible for a class of Ss to be split into two groups without them being aware of it at the time, one half having the extrapolation to Adjectives, the other the extrapolation to Prepositions, while both heard the same reading by the teacher and the same test 1. The same two tests were used throughout, namely Adjective position in the NP (groups A, B, G, H), and NP position in the PP (groups C, D, E, F).
________________________________________________________________________
MAL Original sample Test 1 Test 2
A. SOV SOV/VSO/SVO Adj N vs N Adj
B. VSO " " " " "
C. SOV " " " Prep N vs N Post
D. VSO " " " " "
E. SOV + Adj N SOV/VSO/SVO + Adj N/N Adj " "
F. VSO + N Adj SOV/VSO/SVO + Adj N/N Adj " "
G. SOV + N Post SOV/VSO/SVO + Prep N/N Post Adj N vs N Adj
H. VSO + Prep N SOV/VSO/SVO + Prep N/N Post " "
Design of the 8 conditions
________________________________________________________________________
Subjects
The experiments were carried out with pupils studying foreign languages in two comprehensive schools in Colchester in Essex: in this Local Education Authority a comprehensive school is a 'creamed' school from which about 5% have been selected for grammar schools1. The tests were given to classes ranging from the first to fourth years, i.e. from an average age of 11 to average age of 14; the first years have not been reported on here. The children were all learning French; some were also learning German; all were native speakers of English. The Ss reported on here numbered 409 and were aged between 12 and 14. The experiment was conducted by their usual class teacher. The Ss were told that they were involved in some research into how people learn foreign languages and would be learning a made-up language called Xtopal; they would not have to remember all the words since they would always have a translation to look at.
The detailed results are given in tables 3 and 4, from which the other tables are drawn.
a) consistency of response
_____________________________________________________________________
TEST 1 TEST 2
MAL N Cons Non-cons p. N Cons Non-cons p.
A 37 36 1 a 30 30 0 a
B 55 54 1 a 52 52 0 a
C 37 37 0 a 29 28 1 a
D 52 51 1 a 48 47 1 a
E 54 47 7 a 47 39 8 a
F 79 53 26 b 53 46 7 a
G 39 19 20 nonsig 19 18 1 a
H 56 43 13 a 43 39 4 a
tot: 409 340 69 321 299 22
Table 1. Consistency of response
(Ss consistent at 12/15, Test 1; 12/14 Test 2)
a=sig at p.<0.001, single row chi squared test); b=sig at p.<0.01)
_____________________________________________________________________
The first result concerned, not the direction or 'correctness' of response, but the consistency. The criterion for consistency for Test 1 was to take 12 responses or better out of 15 with the same position as consistent and to count blank or double responses as inconsistent. Results are displayed in table 1. In Set 1 (MALs A-D) 178 out of 181 Ss were consistent, 98.3%; in Set 2 (MALs E-H) 162 out of 228 were consistent, 71%. All of the results for each MAL were significant (chi squared single row test, df.1) at p.<0.001 except for MAL F which was significant at p.<.0.01 and MAL G which was non-significant. It should be pointed out that non-significant still meant that 19 out of 39 Ss achieved at least 12/15 consistency.
Since the point of the experiment was to test extrapolation, those who failed to learn the original MAL were not counted for Test 2, the Ns for which are therefore the successful Test 1 Ss. Test 2 used a 12 out of 14 criterion of consistency; 157 children out of 159 were consistent, 98.7%, in Set 1, and 142 out of 162, 88%, in Set 2. The results for all the MALs were significant at p.<0.001. The first point seems quite clear: the majority of Ss extrapolated consistently from the VP order and from VP order plus either NP order or PP order regardless of whether they did it 'right'.
b) learnability of the MALs
_____________________________________________________________________
MAL N Correct Incorrect p.
A 37 30 7 a
B 55 52 3 a
C 37 29 8 a
D 52 48 4 a
E 54 47 7 a
F 79 53 26 b
G 39 19 20 nonsig
H 56 43 13 a
tot: 409 321 88
Table 2. Correctness of response (Ss correct at 12/15 Test 1)
a=p.<0.001 (chi squared single row test, df 1); b=p.<0.01
__________________________________________________________________
The next question is whether the original MALs presented to the Ss were in fact learnable. Table 2 presents the results for success, using the same 12/15 criterion for Test 1 but only counting those that were 'correct'; the total of successful Ss for the single phrase MALs in Set 1 was 159 out of 181 (88%); for the two phrase MALs in Set 2 the successful Ss numbered 162 out of 228 (71%). All of these results were significant at p.<0.001 (chi squared single row test, df 1) except for F (p.<0.01) and G (nonsignificant); again the last result still means that 19 out of 39 Ss were successful with at least 12 out of 15 sentences. All but one of these MALs was learnable to criterion by the majority of the children; and even the exception, MAL G, was nevertheless learnt to criterion by almost half the children.
Set 1 MALs (A-D) were easier to learn than Set 2 (E-H) (chi squared test, df 3, p. <0.001); one explanation for the overall difference was that the Ss in Set 1 had 30 examples of one phrase type, those in Set 2 had 13 examples of each of two phrase types; hence the Set 2 Ss had only half the number of instances for each phrase type in Set 2. Within Set 1 the differences between languages on overall success was non-significant (df 3, chi squared test), as was the difference between the two pairs AC and BD (chi squared test); BD and AC are in fact the same sample language. Within Set 2 the order of difficulty was EHFG, that is to say SOV + Adj N > VSO + Prep N > VSO + N Adj > SOV + N Post: these differences were statistically significant (p.<0.001, chi squared test). While there is little difference between the learnability of set 1 MALs, the languages of Set 2 vary more in difficulty.
c) extrapolation to new phrases
Tables 3 and 4 present the scores for each test in detail.
______________________________________________________________________ MAL Test 1 Test 2
N SOV VSO SVO Non-cons N AdjN NAdj PrepN NPost Non-cons
A SOV 37 30 0 6 1 30 28 2 - - 0
B VSO 55 0 52 2 1 52 17 35 - - 0
C SOV* 37 29 0 8 0 29 - - 2 26 1
D VSO 52 0 48 3 1 48 - - 17 30 1
Total: 181 159
Table 3. Results for Set 1 (MALs A-D)
Ss achieving 12/15 (Test 1) and 12/14 (Test 2)
*This version appeared to have one item misread by the teacher since item 4 was left blank by 19 Ss, interpreted wrongly by 4 Ss and correctly by 13; this did not however materially affect the number of consistent Ss.
________________________________________________________________________
MAL Test 1 Test 2
N SOV VSO Non-cons N AdjN NAdj PrepN NPost Non-cons
(AN) (NA)
(NP) (PN)
E SOV+Adj N 54 47 0 7 47 - - 34 5 8
F VSO+N Adj 79 0 53 26 53 - - 3 43 7
G SOV+N Post 39 19 0 20 19 0 18 - - 1
H VSO+Prep N 56 0 43 13 43 27 12 - - 4
Total: 228 162
Table 4 Results for Set 2 MALs (E-F)
Ss achieving 12/15 (Test 1) and 12/14 (Test 2) ________________________________________________________________________
Now to the four initial questions on extrapolation. Question (i) asked whether VP order was extrapolated to PP order; did the position of the Object NP influence the preference for Preposition or Postposition in Set 1 languages? Results for Test 2 show Verb Object was extrapolated to N Postposition by 63% (MAL D); Object Verb was extrapolated to N Postposition by 90% (MAL C). Given an unknown language, the Ss therefore extrapolated to a Postposition rather than a Preposition; the differences between the two MALs were nevertheless significant (chi squared test, df 2, p.<0.025).
Question (ii) asked whether VP order extrapolated to NP order; did the position of the Object NP influence the location of the Adjective in Set 1 languages? Results from test 2 show that extrapolation from Object Verb order to Adj N was 93% (MAL A); from Verb Object order to N Adj was 67% (MAL B). The order in the VP therefore had a clear differential impact on the extrapolation in the two conditions (chi squared test, df 2, p.<0.001).
Question (iii) asked whether order in the VP and order in the PP extrapolated to order in the NP: did Object position in combination with the choice of Preposition versus Postposition affect Adjective position in Set 2 languages? Test 2 results showed Object Verb order in the VP and N Post order in the PP extrapolated to N Adj in 95% (MAL G), Verb Object and Prep N orders to Adj N in 64% (MAL H). The combinations of order in two phrase types had an evident impact on the third (chi squared test, df 2, p.<0.001), though the number of successful Ss was in itself lower than for Set 1 and differed more between languages.
Question (iv) asked whether order in the VP combined with Adjective position extrapolated to Preposition or Postposition; did Object position in the VP and Adjective position in the NP affect the PP? Object Verb and Adj N orders extrapolated to Prep N in 71% of Ss (MAL E), Verb Object and N Adj orders to N Post 90% (MAL F). Again very definite preferences were felt, which differ strikingly from one MAL to another (chi squared test, df. 2, p.<0.001).
The Ss behaved in a consistent fashion. The MALs were successfully learnt by the majority of the Ss; the word order of Verb and Object, Adj and N, and Prep/Post and N were learnable from a small sample of sentences; the only exception was the comparative difficulty of MAL G (Subject Object Verb plus Noun Postposition) which had far fewer consistent Ss; it should be noted however that the consistent Ss for G were all in fact correctly rather than incorrectly consistent. The same extrapolations were made by most Ss to each MAL; a given order for one phrase type had repercussions on other phrase types. Learners do indeed extrapolate word order across phrase types. The hypothesis made by Zobl (1986) that L1 order will be a major influence when the L2 is the same type of language is not supported by any advantage for VSO MALs B and D, which have Verb and Object in the same relative position as English, over SOV MALs A and B, which do not; nor does the preference for SOV over VSO in the world's languages (Tomlin, 1984) appear to have any effect in Set 1 languages, though SOV + Adj N (MAL E) was the most learnable of the Set 2 languages.
The more detailed results on extrapolation are intriguing. The results for extrapolation to the Preposition Phrase are summed up in Table 5.
Prep. Post.
SOV (C) 7% 90%
SOV + Adj N (E) 71% 15%
VSO (D) 35% 63%
VSO + N Adj (F) 6% 90%
Table 5 Extrapolation to Preposition position
Both VSO and SOV led to Postposition (C and D). According to the CCH, however, learners should have extrapolated to Postposition from SOV, which they did (C), and to Preposition from VSO, which they did not (D). According to the head parameter, the extrapolations from SOV and VSO should have differentially led to Postposition and Preposition respectively, which they did not. While there is some difference between the results for the two MALs in that VSO has far less Postpositions, another factor seems at work. Combining the VP order with Adjective position resulted in SOV with Adj N leading to Preposition (E), and VSO with N Adj to Postposition (F); there is a dramatic switch in extrapolation for SOV and an increased amount of extrapolation to Postposition for VSO. These results go against the CCH, which makes the opposite predictions that SOV + Adj N should extrapolate to N Post and VSO + N Adj should extrapolate to Prep N. One interpretation is that Ss expect the preposition to be on the same side as the Adjective.
The information for adjectives is given in table 6.
Adj N N Adj
SOV (A) 93% 7%
SOV + N Post (G) 95% 0%
VSO (B) 33% 67%
VSO + Prep N (H) 64% 28%
Table 6 Extrapolation to Adjective position
SOV led to Adj N, VSO to N Adj, though less strongly. The presence of N Post did not change the tendency to have Adj N for SOV (G); the presence of Prep N however switched the extrapolation from slightly N Adj to slightly Adj N (H), favouring the Preposition on the same side as the Adjective. The CCH prediction would have SOV and SOV + N Post extrapolating to Adj N, which is fulfilled, and VSO and VSO plus Prep N extrapolating to N Adj, true for the former but not the latter. The head parameter as such makes no predictions for Adjective position in the NP.
We can now propose a set of extrapolation strategies that the Ss used. These strategies could be phrased in many ways both negatively and positively; the clearest form of statement in the terminology of the earlier discussion is that learners of such MALs assume;
1) if only the order of elements in the VP is known, the language has Postpositions (MALs C and D), i.e. both Object Verb and Verb Object extrapolate to N Post:
Any VP order --> Postposition
2) if only the order in the VP is known, Adjectives occur on the same side of the Noun in the Noun Phrase as Objects do in the Verb Phrase (MALs A and B), i.e. Object Verb extrapolates to Adj N, Verb Object to N Adj:
Any VP order --> same Adjective position as Object
3) if the orders in both the VP and the NP are known and are consistent (SOV + Adj N or VSO + N Adj), the Preposition occurs on the same side of the PP as the Adjective does in the NP (MALs E and F), i.e. Object Verb plus Adj N extrapolates to Prep N, and Verb Object plus N Adj to N Post:
Any VP order with the same NP order --> Preposition in the PP on the same side as the Adjective and the Object
4) if the orders in both the VP and the PP are known and are consistent (SOV + N Post and VSO + Prep N), the NP order is Adjective N (MALs G and H):
Any VP order with the same PP order --> Adj N
What explanations can be found for these strategies? One possibility, since all the Ss were native speakers of English, might be the construction of 'anti-English', differing from English as much as possible; this is only supported by the preference for Postpositions found in strategy 1. Another possibility is 'pro-English'; all languages should be the same as English; this is however only supported by strategy 4. Or since the Ss were all learning French, they might be looking for 'pro-French', perhaps accounting for their preference for Adjective N: having learnt two languages that permit Adjective Noun order predisposes the learner to think all languages have Adjective Noun order.
The Universal Grammar explanation for consistency of complement position in the phrase is directly supported only by the Prepositional Phrase (strategy 3), and contradicted by the preference for Postpositions (strategy 1), except that the Verb Object order has some influence on the strength of the preference. The Hawkins implicational universals were not themselves tested since no impossible stages were involved. Cross Category Harmony however had varying influence on the strategies, supported by (2) and (3) but not by (1) and (4); one would have expected such a strong constraint to have been more influential, as Lujan et al (1984) found, rather than for the learners to have acquired MALs with mixed settings, such as their preferences for N Post with VSO and N Adj (F) and for Prep N with SOV and Adj N (E). None of these explanations seem able to cover more than a proportion of the results, which are mostly clear and solid-seeming. Perhaps the nature of the task affected the Ss; they gave few signs of confusion, however, but appeared to attack it confidently and consistently. Perhaps they were treating it as a particular kind of problem-solving; if so, their apparent high degree of unanimity suggests they all treated it as the same problem, but one which is not transparent to the linguist or the language teacher and will need further research to establish. Or the very 'purity' of isolating order learning from vocabulary acquisition and of limiting the syntactic point so precisely may have created an artificial task. Yet the task is not distinctively different from those in other MAL research; if valid, such objections would operate against all such work. Perhaps the Ss treat the Preposition as a complement rather than a head. Or indeed perhaps the GB analysis used here is inadequate since it ignores both the possibly crucial role of the Subject of the sentence and the possible derivation of the main sentence orders from other orders via V-movement. At the moment the research appears to have demonstrated the existence of a set of extrapolation strategies that are powerful for the learner and that are not clear consequences of the two main linguistic approaches to word order. We need to see whether these strategies are in fact found in the learning of natural languages, in other words whether the experiments have uncovered a different type of learning strategy or the results are artefacts of the experimental design. Work in progress concentrates on the head parameter and expands the current approach by looking at first language acquisition data drawn from the Bristol work (Wells, 1985), at timed grammaticality judgements by L2 learners from different linguistic backgrounds, and at computer simulation of a parser that sets the value for the head parameter from data (Cook, in progress). To validate the proposed extrapolation strategies, they will need to be refined and expanded.
REFERENCES
Borer, H., & Wexler, K. (1987), 'The maturation of syntax', in T. Roeper & E. Williams (eds.) Parameter Setting, Dordrecht, Reidel
Brown, R. (1973), A First Language: The Early Stages, London, Allen & Unwin
Chomsky, N. (1986a), Knowledge of Language: Its Nature, Origin and Use, New York, Praeger (KOL)
Chomsky, N. (1986b), Barriers, M.I.T. Press
Chomsky, N. (1987), Kyoto Lectures, MS
Chomsky, N. (1988), Language and Problems of Knowledge: The Managua Lectures, MIT Press
Clahsen, H., & Muysken, P. (1986), 'The availability of universal grammar to adult and child learners - a study of the acquisition of German word order', Second Language Research, 2, 2, 93-119
Cook, V.J. (1988), Chomsky's Universal Grammar: An Introduction, Oxford, Blackwells
Cook, V.J. (in progress), The Head Parameter: A Test Case For Universal Grammar
Crystal, D., Fletcher, P., & Garman, M. (1976), The Grammatical Analysis of Language Disability, Edward Arnold
Eckman, F.R., Bell, L., & Nelson, D. (1988), 'On the generalisation of relative clause instruction in the acquisition of English as a Second Language', Applied Linguistics, 9, 1, 1-20
Gass, S. (1979), 'Language transfer and universal grammatical relations', Language Learning, 29/2, 327-344
Gass, S. & Ard, J. (1980), 'L2 data: their relevance for language universals', TESOL Quarterly, XIV/443-452
Goodluck, H. (1986), 'Language acquisition and linguistic theory', in P. Fletcher & M. Garman (eds.), Language Acquisition, second edition, Cambridge University Press
Hawkins, J.A. (1982), 'Cross-category harmony, X-bar, and the prediction of markedness', J. Linguistics, 18, 1-35
Hawkins, J.A. (1983), Word Order Universals, New York, Academic Press
Hawkins, J.A. (1987), 'Implicational universals as predictors of language acquisition', Linguistics, 25, 453-473
Hawkins, R. (1987), 'The notion of "typological markedness" as a predictor of order of difficulty in the L2 acquisition of relative clauses', paper presented to the BAAL seminar on the Place of Linguistics in Applied Linguistics, Essex
Henry, A. (1986), 'Linguistic theory and second language teaching', paper presented to CILT workshop on Acquiring Language and Learning Languages, Windsor
Keenan, E. & Comrie, B. (1977), 'Noun phrase accessibility and universal grammar', Linguistic Inquiry, 8, 63-100
Lujan, M., Minaya, L., & Sankoff, D. (1984) 'The universal consistency hypothesis and the prediction of word order acquisition stages in the speech of bilingual children,' Language, 60, 2, 343-371
Lust, B. (1983), 'On the notion "Principal Branching Direction": a Parameter in Universal Grammar', in Otsu, Y., van Riemsdijk, H., Inoue, K., Kasimo, A., & Kawasaki, N. (eds.), Studies in Generative Grammar and Language Acquisition, International Christian University, Tokyo
MacWhinney, B. (1983), 'Miniature linguistics systems as tests of the use of universal operating principles in second-language-learning by children and adults', J. Psycholinguistic Research, 12, 5, 467-477
McLaughlin, B. (1980), 'On the use of Miniature Artificial Languages in Second-Language Research,' Language Learning, 30, 1
Meier, R.P., & Bower, G.H. (1986), 'Semantic reference and phrasal grouping in the acquisition of a miniature phrase structure language', J. Mem. Lang, 25, 492-505
Meisel, J.M. (1986), 'Word order and case marking in early child language. Evidence from simultaneous acquisition of two first languages', Linguistics, 24, 123-183
Morgan, J.L. (1986), From Simple Input to Complex Grammar, MIT Press
Morgan, J.L., & Newport, E.L. (1981) 'The role of constituent structure in the induction of an artificial language', JVLVB, 20, 67-85
Nagata, H. (1981), 'Effectiveness of word order and grammatical markers as syntactic indications of semantic relations', J. Psycholing. Res., 10,5
Pinker, S. (1984), Language Learnability and Language Learning, Harvard University Press
Smith, K.H. & Braine, M.D.S., 'Miniature languages and the problem of language acquisition', mimeo, no date
Sinclair, H. (1969), Acquisition du Langage et Development de la pensee, Dunod, Paris
Tomlin, R.S. (1984), 'The frequency of basic constituent orders', Papers in Linguistics, 17, 163-196
Valian, V., & Coulson, S. (1988), 'Anchor points in language learning: the role of marker frequency,' J. Mem. Lang., 27, 71-86
Wells, C.G. (1985), Language Development in the Preschool Years, Cambridge University Press
Wexler, K., & Manzini, M.R. (1987), 'Parameters and learnability', in T. Roeper & E. Williams (eds.), Parameters and Linguistic Theory, Dordrecht, Reidel
Zobl, H. (1986), 'Word order typology, lexical government, and the prediction of multiple, graded effects in L2 word order', Language Learning, 36, 2, 159-183
Footnote
1. I am endebted to the staff and pupils of Stanway and Philip Morant Schools, Colchester, for their help with this experiment.