Timed Comprehension of Binding in Advanced L2 Learners of English

Vivian Cook 
SLA Home

Draft of Language Learning, 40, 4, 557-599, 1990 [Figs have become corrupt]


Abstract

This paper is concerned with how advanced L2 learners of English interpret reflexive anaphors such as "himself" and pronominals such as "him" in sentences such as "John said Peter helps himself" and "John said Peter helps him".  Parameterised Binding Theory claims that the settings for the governing category parameter dictate whether particular anaphors or pronominals are bound to other noun phrases; the five possible settings are related in opposite hierarchies of inclusiveness for anaphors and pronominals according to the Subset Principle.  An experiment is described that tested the interpretation of "himself" and "him" across five sentence types by by 14 native speakers and 47 advanced L2 learners of English from 3 different language backgrounds - Japanese, Romance, and Norwegian.  A computer controlled comprehension task gave the subjects 40 sentences, for each of which they had to decide whether "him" or "himself" referred to "John" or "Peter" by pressing the appropriate key.  Results were: (i) anaphors were slightly more difficult than pronominals overall, (ii) pronominals were not treated as anaphors, (iii) a consistent order of difficulty was found for the five sentence types, with certain exceptions, (iv) common orders of difficulty and of response time occurred in all groups regardless of first language, again with exceptions, (v) on one view the Subset Principle was positively related to difficulty for anaphors, negatively for pronominals.  The discussion concerns the validity of comprehension tests as evidence for the Universal Grammar model.  It is suggested that the results show that parameter settings are related to performance as well as to grammatical competence

This paper is concerned with how advanced L2 learners of English interpret reflexive anaphors such as "himself" and pronominals such as "him"; in other words with whether they know that in the sentence:
   "John said Peter helps himself."
"himself" refers to the same person as "Peter", but does not refer to the same person as "John", while in the sentence:
   "John said Peter helps him."
they know that "him" may refer to the same person as "John", or to some unidentified person, but may not refer to the same person as "Peter".  In conventional grammatical terms, binding is concerned with the possible antecedents for anaphors and pronominals.  Though partly a matter of syntax that has to be accommodated within any model of L2 acquisition and use, such phenomena play a particular role in the Chomskyan Universal Grammar (UG) approach to L2 learning.  The principles and parameters model of UG theory establishes principles that are true of all languages and parameters whose settings vary from one language to another (Chomsky, 1988; Cook, 1988a).  Children learn which setting is appropriate for the language they are acquiring from the positive evidence of the sentences they hear rather than the direct negative evidence of parents' corrections or the indirect negative evidence of sentences that don't occur.

Most parameters that have been studied in L2 acquisition are like switches with two positions - on or off.  For example the pro-drop parameter describes whether a language has or doesn't have null-subject sentences (declarative sentences without overt subjects); various work has explored which of these two settings occurs first in L1 learning (Hyams, 1986) and how the L1 setting affects the L2 (White, 1986).  In Government/Binding (GB) Theory (Chomsky, 1981; 1988), which underpins the UG theory, the relationship of 'binding' between noun phrases and anaphors or pronominals is handled within a distinct module of Binding Theory; recently this theory has been seen as involving a parameter called the governing category parameter, which has five settings rather than two (Wexler and  Manzini, 1987).  The change from two-valued to multi-valued parameters raises important questions for language acquisition about the relationship between the settings: the claim is that the child prefers a setting that allows the smallest set of possible sentences consonant with the data rather than a setting that goes beyond the data presented (Wexler and Manzini, 1987); there is a hierarchy of settings for this parameter.  Consequently a general learning principle, the Subset Principle, is postulated that requires the child to act in this conservative manner.

The L2 learning concerns that will be addressed in this paper are:
- the factual basis that can be provided for L2 learners' interpretation of binding in English
- whether the L1 settings for the parameter are relevant to L2 comprehension of binding in English.  In other words, is Universal Grammar directly accessible to L2 learners or mediated indirectly via the first language?
- how the Subset Principle relates to L2 learning and processing, that is to say are the mistakes of L2 learners related to the hierarchy of parameter-settings laid down by the Subset Principle?
The evidence to be considered is how advanced adult L2 learners of English from different backgrounds and English native speakers interpret varying sentences with anaphors and pronominals in an experimental setting.

Binding Theory

Since the claims for acquisition in a UG framework are always tied in to the details of GB analysis, we first need to sketch the syntactic framework for binding in some detail.  The convention will be adopted that NPs that are bound have the same referential 'index', shown by 'i", while those coreferring with different entities have different indices, not usually shown unless specifically required.  Thus the binding relationships in the above sentences can be shown respectively as:
   "John said Peteri helps himselfi."
indicating that "Peter" is bound to the reflexive aanaphor "himself" and refers to the same person and:
   "Johni said Peter helps himi."
showing that in this case "John" is bound to the pronominal "him", rather than to "Peter".  This convention is necessary because, unlike other mistakes such as word order, mistakes in binding are in a sense invisible.  Suppose an English speaker says:
    "John said Peter helps himself."
incorrectly intending to mean that Peter helps John, i.e.:
    *"Johni said Peter helps himselfi."
A listener who used the correct form of binding would mistakenly assume that the speaker meant that Peter helped Peter rather than Peter helped John; nothing in the sentence itself shows the speaker's error.  A sentence with the wrong binding is always interpretable by a listener with the correct binding, whatever the speaker intends.  As Berwick and Weinberg (1984, p.170) put it, 'If an antecedent can be found the sentence will be grammatical, otherwise not; in both cases the sentence will be parsable.'  Binding is thus involved with the interpretation of the sentence, with deciding which is the correct antecedent, an important point when considering research methodology for investigating its acquisition.

The principles of Binding Theory in the 'classical' GB model (Chomsky, 1981), outlined for example in Cook (1988a), apply to all languages by specifying the structural boundaries inside the sentence within which two elements can be bound, known as the 'local domain'; the elements affected by binding are the different types of Noun Phrase - referring expressions such as "John", pronominals such as "him", and anaphors such as "himself".  The typical form of the Binding Principles is taken from Chomsky (1986), with minor adaptations:
    A. An anaphor is bound in a local domain
    B. A pronominal is free in a local domain
    C. A referring expression is free
How can a local domain be defined?  In the classical GB model, the local domain for any NP is 'the minimal governing category'.  The part of the sentence within which the NP occurs has to meet several requirements to constitute a minimal governing category:
i) it must be a maximal projection, i.e. a Verb Phrase (V''), a Noun Phrase (N''), a Prepositional Phrase (P''), an Adjective Phrase (A''), or an Inflection Phrase (I'').
ii) it must contain a subject.  This effectively limits it to the Inflection Phrase and the Noun Phrase, which in GB is considered to have a subject in phrases such as "Fred's love of the bottle".
iii) it must contain a lexical category that governs the NP.  The lexical categories are Noun (N), Verb (V), Adjective (A), and Preposition (P).  Government is a complex notion within GB theory defined in terms of the structure of the sentence; it is explained in detail in Cook (1988a).  In brief if two elements have the same maximal projection above them in the structure of the sentence and no other intervening maximal projections in between, one may govern the other.  In the following tree, X'' stands for any maximal phrase (V'' etc), Y'' for any other maximal phrase acting as complement within the phrase; the dotted lines show the 'ceiling' and 'floor' of maximal projections within which one element may govern another.  So X governs everything between the X'' and the Y'' but nothing below the Y'' or above the X''.    

         ...             
          |              
      ----X''----        
          |              
          X'             
      ____|______        
      |         |        
      X     ----Y''----  
                |        
                Y'       
                |        
               ...       

The simplified tree below represents the actual sentence "Margaret drinks gin" and shows the maximal projections that act as barriers to government.  In the V'' "drinks gin" the V "drink" governs anything within the ceiling of the V'' and the floor of the N'', but nothing above the ceiling such as the I', or below the floor such as the N'.  In the I'' the N "Margaret" governs everything between the I'' and the V'' but nothing below the V''.

             ---I''---
        ________|________
        |               |
        N''             I'
        |          _____|______
        N'         |          |
        |          I       ---V''---
        N                     |           
        |                     V'          
     Margaret             ____|____       
                          |       |       
                          V    ---N''---  
                          |       |
                        drink     N'
                                  |
                                  N
                                  |
                                 gin

The full Chomskyan definition of the local domain for binding is then 'the minimal governing category of alpha, where a governing category is a maximal projection containing both a subject and a lexical category governing alpha ... ' (Chomsky, 1986). 

To see how the idea of governing category is used in binding let us see whether the anaphor "himself" may or may not be bound to "Peter" in:
   "Peter says John blamed himself."
The question is whether the phrase "John blamed himself" is a local domain for "himself".  First, according to (i), it must be within a maximal projection; one maximal projection 'above' "himself" is the sentence (Inflection Phrase or I'') "John blamed himself".  Secondly, according to (ii), it must have a subject; the NP "John" is an appropriate subject.  So "John blamed himself" meets the first two requirements for an appropriate governing category for "himself" by having a maximal projection and a subject.  Part (iii) of the definition of governing category requires the maximal projection to contain a lexical category governing the N''.  In "John blamed himself" "blame" is a lexical category (V) within the same maximal phrase (the V''); there is no other maximal phrase intervening between it and the N'' "himself": so "blame" governs "himself".  "John blamed himself" has therefore met all three parts of the definition of governing category for "himself". 

We can now refer back to the Binding Principles.  Since "himself" is an anaphor, Principle A applies (An anaphor is bound in a local domain) and, since "John blamed himself" is an acceptable local domain, "himself" and "John" must be bound, i.e. "John" is the antecedent for "himself":
   "Peter says (Johni blamed himselfi)."
(From now on brackets will be used in this way to indicate the local domain for binding when necessary).

Next let us take the sentence:
    "Peter says John blamed him."
Is "him" bound to "Peter" or to "John"?  The answer hinges on whether "John blamed him" is a governing category for "him".  One of the maximal projections above "him" is the I'' "John blamed him"; this contains a subject "John"; "blame" is a lexical category within the same maximal phrase as "John" and it is not separated from it by any other maximal phrase: so "blame" governs "him".  The I'' "John blamed him" consequently meets the three conditions necessary to be the governing category for "him".  But this time Principle B applies, since "him" is a pronominal (A pronominal is free in a local domain), and ensures that "him" must be free, i.e. not be bound in this local domain.  So "him" must corefer out of this domain, say with "Peter":
    "Peteri says (John blamed himi)."
While Principle B dictates that a pronominal must not be bound within a local domain, it is neutral about what it is bound to outside this domain.  A pronominal is always free to be bound to something outside the sentence altogether; the antecedent of "him" may be someone else called, say, Christopher who is not mentioned in the sentence.  The contrast between anaphors like "himself" and pronominals like "him" is that "himself" is always bound in the minimal local domain, often a clause, and "him" is never bound in the minimal domain.

Finally the interpretation of referring expressions such as "John" and "Peter" in:
     "Peter says John blamed him."
is not tied in to anything else; they must not be bound to anything in the sentence, according to principle C (A referring expression is free).

Parameterised Binding Theory

The previous section has described the 'classical' analysis of binding.  Difficulties with this start to emerge if one looks at a wider range of languages; the technical sources drawn upon here are chiefly Wexler and Manzini (1987) and Saleemi (1988).  One problem is Icelandic sentences with the reflexive "sig" such as:
      "Jane segir a& maria elski sig."
      [Jane says that Maria loves herself]
This may mean:
      [Janei says that Maria loves herselfi]
although the English equivalent:
      *"Janei says that Mary loves herselfi."
could not have this interpretation because "herself" would be bound to "Mary" within the governing category "Mary loves herself" by Principle A.

Hence there is a difference in binding between English and Icelandic.  Wherever there is a systematic difference between languages, GB postulates a parameter to capture it.  The solution suggested by Wexler and Manzini (1987) is that the criteria used above for deciding a governing category are not universal but vary from one language to another; they postulate that the definition of governing category varies between languages.  This parameter has five settings, outlined by Wexler and Manzini (1987) as follows:
a) English permits:
   "Johni criticised himselfi."
and:
   "Johni heard the report about himselfi."
but not:
   *"Johni heard Peter's criticisms of himselfi"
because "John" is here outside the governing category for the anaphor "himself", namely "Peter's criticisms of himself", and so in breach of Principle A.  The difference affects requirement (ii); the need for a subject within the governing category (the genitival "Peter's" being treated as the subject of the N'' "Peter's criticisms of himself") is not absolute but depends on the language.  The definition of governing category for "himself" crucially depends upon the presence of a subject, whether within a sentence or within an N'': for a group of lexical items like "himself" in English a governing category must contain a subject.  In other languages this requirement may not be necessary.
b) Italian permits a range of sentences that are ruled out in English.  Taking the reflexive anaphor "se" ("himself/herself"):
   "Alicei vide Mario guardare sei nello specchio."
   [Alicei saw Mario look at selfi in the mirror]
The binding of "Alice" and "se" is permitted by Principle A in Italian though impossible in English.  Similarly the Italian sentence:
   "Alicei guardo i retratti di sei di Mario."
   [Alicei looked at Mario's portraits of selfi]
is unexceptionable while the English version is impossible.  A maximal projection does not qualify "i retratti di se de Mario" to be a governing category for the Italian reflexive "se", even though it contains a subject "Mario".  What is lacking is a finite verb form - in short INFL, the abstract Inflection constituent that gathers together such features of the sentence as agreement and tense.  For a second group of lexical items, of which Italian "se" is a member, a governing category must contain an INFL

We should now introduce a particular technical use of the word "language"; in learnability theory a language can be defined as the set of sentences generated by a grammar rather than as, say, the English language.  If all the sentences possible according to the (a) setting are compared with all those possible according to the (b) setting, more sentences are possible according to (b); the 'language' defined by (b) is 'larger' than that defined by (a) in that (b) includes not only categories that have subjects but also those that have INFL.
c) The pronominal "hann" in Icelandic can be seen in the sentence:
   "Jon skipa&i mer a& raka hann."
   [Jon ordered me to shave him]
Here "hann" comes within an embedded sentence that has an INFL; "mer a& raka hann" (me to shave him) should therefore qualify as a governing category for "hann", and, according to Principle B, "hann" must bound outside it, say with the N'' "Jon": but in fact it can not.  That is to say the following interpretation is ungrammatical. 
   "Joni skipa&i mer a& raka hanni."
The governing category for "hann" must therefore be the main sentence rather than the embedded sentence; the defining property of the governing category is not only the presence of INFL but also the presence of Tense: for pronominals like Icelandic "hann" a governing category must contain Tense
d) Let us now look at the Icelandic reflexive anaphor "sig":
   "Joni segir a& maria elski sigi."
   [John says that Maria loves (subjunctive) self]
The anaphor "sig" corefers with "Jon", although the phrase "a& maria elski sig" would prevent it by containing a subject that would qualify it as a governing category for English "himself" (a), an INFL that would qualify it for Italian "se" (b), and it has Tense to qualify for (c).  Hence, if this phrase forms a governing category for "sig" in Icelandic, there must be a further condition that Tense must be indicative rather than subjunctive.  For items such as Icelandic "hann" a governing category must contain an indicative tense.  Again (d) generates a larger 'language' than the preceding settings.
e) Finally let us take a Japanese example containing the reflexive anaphor "zibun" to which Principle A should apply:
   "Johni-wa Bill-ga zibun-oi nukunde irul to omotte iru."
   [John Bill SELF hates that thinks]
   [John thinks that Bill hates self]
Here "zibun" may be bound to "John" (or "Bill" in fact) even though it is within a phrase "Bill self hates" that meets all of the requirements (a)-(d) for its governing category by having a subject, INFL, Tense, and an indicative Tense; "zibun" is bound only within the whole sentence, i.e. wherever the main or "root" tense can be found.  For Japanese "zibun" and similar items a governing category must have "root" tense.  (e) as usual defines a larger 'language' than (d).

The full definition of the parameter from Wexler & Manzini (1987, p.53) can now be displayed:

Governing category parameter
gamma is a governing category for alpha iff
   gamma is the minimal category which contains alpha and
     a. can have a subject, or
     b. has an INFL, or
     c. has a Tense, or
     d. has an indicative Tense, or
     e. has a root Tense

A further point about this parameter is seen in the last example:
   "Johni-wa Bill-ga zibun-oi nukunde irul to omotte iru."
   [John Bill SELF hates that thinks]
   [Johni thinks that Bill hates selfi]
"Zibun" may be bound either to "John-wa" or to "Bill-ga".  In other words anaphors with the (e) setting may be ambiguous if there are several Noun phrases within the sentence; the whole sentence is the governing category and both "John-wa" and "Bill-ga" are within it and so available for binding.  In English "himself" with an (a) setting is necessarily bound to "Bill"; languages that have, say, the (e) setting do not exclude the English interpretation but are not restricted to it; they have a wider range of possible interpretations.

The different parameter settings for governing category do not strictly speaking distinguish whole languages such as English or Japanese from one another but apply to individual anaphors or pronominals: English "himself" has value (a), Icelandic "hann" has value (c), Japanese "zibun" has value (e), and so on.  The (a) setting applies to the English anaphor "himself" not to English as whole; the (d) setting applies to the Japanese anaphor "zibun", not to the Japanese language, and so on.  This view of parameters sees them as linked to lexical items rather than to grammatical principles.  It proposes that everything that is acquired by the child is in the lexicon; principles are common to all languages, parameter settings are learnt along with the lexical items of a particular language; 'there is only one human language, apart from the lexicon, and language acquisition is in essence a matter of determining lexical idiosyncrasies.' (Chomsky, 1989, p.44).  This reflects the developing view within UG theory that parameters are properties of lexical items rather than of syntax.

Finally there is variation between languages not only in the range over which items may be bound in the sentence but also in what they may be bound to.  In languages such as Japanese an anaphor such as "zibun" must be bound to a subject; in other languages such as English it may be bound to any noun phrase.  This necessitates a further parameter called the proper antecedent parameter, given in Wexler and Manzini(1987, p.64) as:

      Proper antecedent parameter
        A proper antecedent for alpha is
          a. a subject beta; or
          b. an element beta whatsoever

Thus in both English and Japanese it is possible to have:
    "Johni told Bill about himselfi"
where "himself" is bound to a subject, but it is only in English that it is possible to have:
    "John told Billi about himselfi"
where "himself" is bound to an object; the equivalent Japanese sentence with "zibun" would be incorrect.

The Subset Principle

Originally applied to parameter-setting by Berwick (1985), the Subset Principle has become chiefly known through the work of Wexler and Manzini (1987).  The problem is: given that children learn only from positive evidence of sentences that actually occur, how do they know which of the five settings for the governing category parameter to choose for each anaphor and pronominal?  Since the children's misconception would not be visible, it would remain uncorrected and they could remain with the wrong setting for ever more.  In other words children could easily adopt a more general setting than the one that actually applies to the item.  To ensure that they do not end up in a blind alley they proceed cautiously, making the least assumptions from the data that are possible, that is to say always assuming the minimum 'language' (a subset of the whole language that would be possible) rather than the maximum 'language' (a superset including all the possibilities).  According to Wexler and Manzini, such a limitation does not come from within linguistic theory or Universal Grammar; it is instead part of a separate module, a learning component.  It can be stated as the Subset Principle (Wexler and Manzini, 1987, p.61):

        the Subset Principle
  The learning function maps the input data to that value of a
  parameter which generates a language:
        a) compatible with the input data: and
        b) smallest among the languages compatible with the input data

The learner always starts with the setting for the parameter that yields the smallest 'language'.  The relationship of each setting to the next can now be put in terms of the Subset Principle.  Binding Principle A specifies the local domain within which binding can occur for anaphors; each of the settings for the governing category parameter generates a larger 'language' for anaphors, i.e. the sentences described by (a) are a subset of those described by (b), the sentences described by (b) a subset of (c), and so on.  For anaphors the Subset Principle ensures that the learner chooses only a minimal local domain; it may be that the learner guesses too small, in which case positive evidence will put him or her on the right track; but the learner cannot guess too big without violating the principle. 

For pronominals such as "him" the Subset Principle makes the opposite prediction.  Principle B specifies the freedom of pronominals, that is to say the fact that they can not be bound within the governing category.  Hence the smallest 'language' is one in which only sentences with free pronominals are permitted, i.e. (e); the next smallest 'language' is one in which all sentences with free pronominals and all those with indicative tense are permitted i.e. (d); and so on.  The Subset Principle applies as much to pronominals as to anaphors but makes the opposite prediction because of the way in which the 'languages' generated by the settings differ in the subset relationship for pronominals and anaphors: for pronominals (e) is a subset of (d), (d) a subset of (c), and so on, while for anaphors (a) is a subset of (b), (b) a subset of (c), and so on; 'the subset hierarchy for pronominals ... is a mirror image of the subset hierarchy for pronominals' (Saleemi, 1988, p.113).

The Subset Principle can also be applied to the proper antecedent parameter introduced above.  This has two values according to whether an antecedent (a) has to be a subject or (b) may be anything whatsoever.  Taking anaphors as an example, the (a) setting generates a 'language' which permits only subject antecedents and is therefore included in the larger (b) 'language' that permits any possible antecedent.  The English item "himself" has the minimal (a) setting on the governing category parameter, the maximal (b) setting on the proper antecedent parameter; the Japanese "zibun" has the maximal (e) setting on the governing category parameter, the minimal (a) setting on the proper antecedent parameter.  Parameters are assumed to be set in isolation from each other, according to the Independence Principle (Wexler and Manzini, 1987, p.46); the Subset Principle therefore applies separately to the governing category and proper antecedent parameters, yielding two different sets of inclusion relationships.  The Subset Principle is a general property of human language learning which can be applied to the setting of any parameter.  It should be pointed out however that it is a highly technical notion.  It is not just claiming that simpler things are learnt before more complicated things but that learners progress through languages that are subsets of larger languages rather than from supersets to subsets, with 'language' being defined as a set of sentences produced by a grammar.  While it may be intuitively appealing to take the Subset Principle as a general concept about language acquisition and to apply it to phenomena in, say, vocabulary or semantics, this is a metaphorical use unconnected with its proper formulation.

Background in L1 and L2 research

How do Binding Theory and the Subset Principle relate to L1 acquisition?  In terms of sheer occurrence Wells (1985, p.168) found that reflexive pronouns were the last type of headword to appear in the noun phrase in English children, way after other pronouns.  In a UG model that uses classical non-parameterised Binding Theory the child is taken to know the Binding Principles; what the child acquires amounts to the knowledge of which words are anaphors, which pronominals, and which referring expressions.  Comparatively few mistakes might be expected to occur and no particular difference in acquisition between pronominals and anaphors would be anticipated.  Read and Hare (1979) found English children already interpreted the binding of "himself" within the clause by the age of six to some extent and this increased gradually up to twelve.  Some empirical research has claimed that anaphors are easier for children than pronominals.  Deutsch, Koster, and Koster (1986) found that Dutch children aged 6 to 10 were consistently better at reflexive anaphors than at pronominals.  Solan (1987) tested English children aged 4 and 7 with a similar result.  He explained the comparative difficulty of pronominals by suggesting that children initially have access only to Binding Principle A, which covers anaphors; Finer (1987) argues that it is not so much that Principle B is missing but that children do not know it applies to particular pronominals.  Lust at al (1989) have used evidence from children at ages from 3 to 10 in English, Dutch, Spanish, and Chinese to claim that anaphors require a smaller core domain to define than pronominals and so are easier.  On similar lines Jakubowicz (1984) argued that children have a stage at which they treat pronominals as bound within a local domain, i.e. as if they were anaphors.

A second point is the difference in difficulty between various types of sentence.  Solan (1987) used tensed clauses, infinitival clauses, and noun phrases; the greater difficulty of infinitival over tensed clauses he ascribed tentatively to problems with discovering the subject in infinitival sentences; the greater difficulty of noun phrases over infinitival sentences he attributed to the possible ambiguity of expressions such as "The horse's picture" and to the minimum distance principle of falling back on the nearest NP.

The sparse L2 research strikes a similar note.  White (to appear) presents a general account of the Subset Principle applied to L2 learning.  Dulay and Burt (1978) found a sequence of appearance of "he/self" before "him/them" before "herself/hisself/theirself" before "himself/themselves" etc.  Finer and Broselow (1986) used a comprehension task to investigate how Korean learners with an (e) setting for the reflexive "vcaki" equivalent of "himself" understood tensed and infinitival English sentences with "himself" and "him".  Though their discussion does not develop this point, their figures show an error rate for tensed clauses of 9% for anaphors and 46% for pronominals, for infinitival clauses of 42% for anaphors and 21% for pronominals; in the tensed sentences anaphors were understood more readily than pronominals, in the infinitival less readily.  The result they concentrate on is the Koreans' use of a (c) or (d) setting that was found in neither their first nor their second language; they argue on the one hand that the Koreans do not have access to the Subset Principle via their first language, as this would mean transferring the (e) setting, on the other that they are not using the Subset Principle from scratch, since that would have meant them adopting the conservative (a) setting.  Their way out of this dilemma is to see it as an interaction between the governing category parameter and the proper antecedent parameter, for which Korean has an (a) setting, and English a (b) setting; if the 'subject of the infinitive is taken as the direct object of the matrix verb', the Koreans can be seen to conform to the Subset Principle as it applies to the proper antecedent parameter.  Hirakawa (1989) found Japanese learners of English successfully set the proper antecedent parameter to its correct English value but had continuing problems with the governing category parameter; in sentences such as:
   "Tom says that Paul told Bob to introduce himself"
98% of English people tested chose "Bob" as the antecedent for "himself", compared to only 53% of Japanese learners.  Thomas (1989) found both Chinese and Spanish learners of English allowed binding out of governing categories defined by the (a) setting, showing both lack of transfer from the L1 and a breach of the Subset Principle.

The research  design and method

So far as L2 learning research is concerned, a number of unresolved or half-resolved questions remain:

i) is the relative difficulty of pronominals compared with anaphors confirmed in L2 learners?  Parameterised Binding Theory predicts variation between specific lexical items; the Subset Principle predicts possible differences for anaphors and pronominals; classical Binding Theory however predicts no difference.  The difficulty of pronominals for L2 learners could provide additional evidence for the independence of pronominals and anaphors, rather than the complementary nature assumed in classical binding.

ii) are pronominals treated as anaphors?  In terms of classical Binding Theory, the equivalence of pronominals or anaphors might be caused by the learners' imperfect knowledge of the Binding Principles themselves, as Solan (1987) argues, or by them incorrectly assigning pronominals to the anaphor class, as Jakubowicz (1984) claims; in this case we should expect to find circumstances in which anaphors had the same binding as pronominals for all sentence types.  In terms of parameterised Binding Theory, pronominals, not anaphors, would have the incorrect setting.  These claims might be distinguished by looking at variation across sentence types; if this phenomenon occurs in one sentence type but not in another, it would reflect the variation in the governing category parameter rather than imperfect knowledge of the Binding Principles.

iii) are there differences in binding across a range of sentence types?  While L1 research has already considered more than two sentence types (Deutsch et al, 1986; Solan, 1987), the L2 research has mostly looked at a single pair, for example tensed and infinitival clauses (Finer and Broselow, 1986); usually it has concerned anaphors in isolation from pronominals (Thomas, 1988; Hirakawa, 1989).  We need to see whether there is a continuum of difficulty across different sentence types reflecting the L2 variation in the governing category parameter.  For this a broader base of sentence types is needed as well as the inclusion of both pronominals and anaphors.

iv) will the learner's L1 affect the results and in what way?  Finer and Broselow (1986) used Korean learners of English with an (e) setting for anaphors, Hirakawa (1989) Japanese learners also with an (e) for "zibun"; this only represents one of the possible five settings for the governing category parameter.  Do L2 learners with different L1 settings produce the same or different results?  This time a broader base of Subjects is indicated.

v) how is the Subset Principle involved in L2 learning?  On the one hand this is a matter of whether the setting of the governing category parameter in the L1 affects the L2 as in (iv), i.e. whether L2 learners only have indirect access to the Subset Principle via the L1 rather than direct access; on the other hand this could show whether the type of governing category involved affects difficulty of understanding for different groups of learners.  A wider range of sentence types and of Subjects would allow a more general picture of the effects of the Subset Principle in L2 learning to emerge, either in terms of the transfer of setting from the L1 or in terms of the relative difficulty of certain settings.  Together these amount to a research design that compares groups of L2 learners with different L1s over a range of sentence types, extending the previous work in two different dimensions.  For good measure a native group would serve as control.

Turning to research methodology, the main problem in investigating binding is the invisibility of mistakes; while a normal grammatical mistake is visible, a wrong interpretation cannot be seen directly.  Looking at actually occurring learners' sentences is therefore ruled out, since it would never be clear which of the interpretations they had in mind.  A test of grammaticality judgements would not distinguish wrong from right interpretations, since each sentence would be grammatical with some interpretation.  The technique of elicited imitation has been employed in some first language work with binding, for example by Lust et al (1989); getting learners to repeat sentences does not however test the wrong versus right interpretations but assumes that the wrong interpretation will yield more overt mistakes, a rather indirect sign of incorrect binding.  It seems preferable to use a comprehension task that can directly show the Subjects' interpretation.  For the present experiment it was decided to use a forced choice comprehension task similar to those in Solan (1987) and Finer and Broselow (1986) but to adapt it to computer control and to add a measure of response time.  The essential technique is to present subjects with a sentence and to ask them to say as quickly as possible which out of two possible people is referred to in it; the time they take is measured.  This provides a controlled form of comprehension since the choice of A and B is always the same, the situation and the time available never vary, and the scoring is automatic.  The additional assumption is that the time involved in coming to a decision about the interpretation of the sentence may in itself reflect its difficulty to the learner, regardless of whether the answer is right or wrong.  The technique of timed comprehension yields two sets of scores: the answers the Subjects give and the response time taken.  Further discussion of these two measures is provided later.

The experiment used an Acorn BBC B microcomputer fitted with the word processing program WORDWISE and a timing program called EXMORE developed at the University of Exeter Psychology Department (Mitchell & Barchan, 1984); this accurately times presentation and response, records which key is pressed, and calculates the average timings for different conditions.  Subjects were presented with a sentence in the centre of the screen and with the request to "Press A if him/himself refers to Peter, B if it refers to John"; the letter B was reassigned to the key G to equalise the areas of the keyboard they had to cover; the keys A and G were accordingly labelled A & B and the other parts of the keyboard obscured with a cloth.  After the Subjects had answered by pressing one of the two keys, the screen went blank and the message "Press Spacebar to go on" appeared; the program went on to the next sentence if the Subject had not answered after twenty seconds.  The sentences were presented in a randomised order that was always the same.  The Subjects were told that the aim of the experiment was to compare how learners of English from different nationalities understood various sentences; the computer would score their answers and would also record how long they took, so they should be as quick as possible.  The experimenter took them through four practice sentences to check that they understood the instructions.

Materials

A set of materials was constructed consisting of ten sentence types.  The list below gives short names that will be used in the later discussion.

A. "Peter shot him."     simple   sentence
B. "Johni shot himselfi."                                "          "
C. "Peter said that John voted for him."   tensed       "
D. "Peter said that Johni voted for himselfi."           "          "
E. "John asked Peter to pay for him."                  infinitival  "
F. "Peter asked Johni to include himselfi."              "          "
G  "Peter discovered John's report on him."            noun  phrase
H. "John reported Peteri's criticisms of himselfi."      "    "
I. "Peter saw John save him."                          
J. "Peter saw Johni save himselfi."                  

Half of the sentences contain the pronominal "him", half the anaphor "himself"; three further sentences of each type were created, using similar vocabulary items with the NPs "John" and "Peter" and rotating verbs and nouns.  The complete list is given in the Appendix.

Let us start with the sentence types with "himself".  First comes the 'simple' sentence with no embeddings:
   "(Johni shot himselfi)." (B)
"John shot himself" fits all five settings of governing category; the anaphor "himself" must be bound to "John" by Principle A.  Whatever the L1 of the learner, no differences are to be expected if the L1 setting for the equivalent reflexive anaphor is transferred to English.  Next comes the 'tensed' sentence:
   "Peter said that (Johni voted for himselfi)." (D)
"John voted for himself" meets the requirements of governing category for settings (a) to (d); in (e) settings the anaphor "himself" could be bound either to "John" or "Peter" by Principle A since the whole sentence could be the governing category.  If the L1 setting for the equivalent reflexive anaphor is transferred to English, then L2 learners with an (e) setting for the equivalent anaphor are likely to have more problems with (D) sentences.

Next is the 'infinitival' type seen in:
  "Peter asked (Johni to include himselfi)." (F)
The phrase "John to include himself" includes a subject "John" and an INFL "to" and so is a governing category according to both (a) and (b); hence English "himself" is bound to "John" by Principle A, as is Italian "se" with a (b) setting in the equivalent sentence.  The sentence does not include any form of Tense; an L1 with an equivalent anaphor with settings (c), (d), or (e) will not therefore treat it as a governing category, but see the main sentence as the governing category, binding "himself" either to "John" or to "Peter" as in:
    "(Peteri asked John to include himselfi)."
There might therefore be differences between learners with settings (a) and (b) as opposed to (c), (d), and (e), inasmuch as the sentence would have two possible antecedents for "himself" for them.

Finally comes the 'noun phrase' sentence type:
   "John reported (Peteri's criticisms of himselfi)."  (H)
With setting (a) as in English the governing category will be "Peter's criticisms of himself" as it contains a subject "Peter"; hence "himself" is bound to "Peter" by Principle A.  "Peter's criticisms of himself" does not meet the criteria for the other settings since it has no INFL and no form of Tense; so "himself" in (H) could in these languages be bound either to "John", i.e.:
   "(Johni reported Peter's criticisms of himselfi)."
or to to "Peter", i.e.:
   "(John reported Peteri's criticisms of himselfi)."
This time there should be strong differences between learners with (a) settings and those with all the rest.  To sum up, English always requires the most local domain, having an (a) setting; other languages permit less and less local domains and hence have greater choice of antecedent.

Alongside these sentence types with anaphors are the equivalent with pronominals:
    "(Peter shot him)." (A)
    "Peteri said (that John voted for himi)." (C)
    "Johni asked (Peter to pay for himi)."  (E)
    "Peteri discovered (John's report on himi)." (G)
These fall within the province of Principle B in that the pronominals must be free within their governing category.  Thus they behave in the opposite fashion to their anaphoric equivalents.  A pronominal is always free to refer outside the sentence altogether, i.e. in:
    "Peter said that John voted for him."
"him" may corefer either with "Peter" or with someone not otherwise mentioned in the sentence. 

In addition we tested the sentence types:
    "Peter saw (Johni save himselfi)." (J)
and:
    "Peteri saw (John save himi)."  (I)
In English the governing category for "himself" is "John save himself" because it has a subject "John".  Koster (1984) describes the behaviour of the Dutch reflexive zich in sentences such as:
    "Billi zag  Peter  naari zich    toe   komen."
    [Bill  saw  Peter  to    himself prt.  come]
    [Bill saw Peter come to himself]
Despite the presence of a subject "Peter" in the embedded clause, and despite the absence of INFL or Tense, "zich" is bound to "Bill".
    "(Billi saw Peter come to himselfi)."
This type was added as a supplementary test to the versions of the 4 main types.  The four practice sentences at the start of each test consisted of two simple anaphor sentences and two simple pronominal sentences.

Subjects

The Subjects for the experiment were all advanced adult learners of English, who were studying in England at the time1.  14 were speakers of Romance languages (6 Spanish, 6 Italian, 1 French, 1 Portuguese), 16 were speakers of Japanese, 17 were speakers of Norwegian.  In addition 14 native speakers of English provided a control group, who were mostly students of applied linguistics.  These groups were chosen to reflect different L1 settings on the Wexler and Manzini hierarchy.  The Romance languages have values for anaphors that incorporate (b), e.g. Italian "se"; Norwegian has a (c) value for the anaphor "seg" (since according to Hellan (1988) it requires Tense but not indicative tense); Japanese "zibun" is the classic example of the anaphor that takes (e).  All the Subjects could be classed as advanced professional users of English, many of them teachers, though their precise level could not be measured by other means.  The research is not then concerned with stages of development as such but with the behaviour of advanced learners.

Results

The overall results are given in tables 1 and 2, for comprehension and response times respectively.  The results are presented in terms of the five research questions outlined earlier.

i) are pronominals more difficult than anaphors ?
Starting with error rate, pronominals indeed have a higher error rate for all groups for the simple sentence types A/B, where the error rate is low in any case, and for the I/J sentence types.  For all others however pronominals have lower error rates, for types G/H for all groups, E/F for all groups except Natives, and C/D for all groups except Norwegians.  The overall error rate for pronominals is 10.1%, for anaphors 12.6% (chi squared test on totals, df 1, p.<0.001); in general anaphors are slightly more difficult than pronominals.  Turning to response times, the results for pronominals show a slightly slower speed than for anaphors, on average 5.188 seconds for pronominals, 4.813 seconds for anaphors; this difference is not however statistically significant, nor is that between the pairs BD/AC and EG/FH (ANOVA tests).  So the anticipated difficulty for pronominals is found only with error rates for simple and tensed sentences, with certain exceptions; the other sentence types show increased error rates for anaphors, particularly infinitival and noun phrase sentence types.  The difference between anaphors and pronominals occurs on particular sentence types rather than across the board and goes mostly in the wrong direction.  One cause for the marginally longer time of 0.375 seconds for pronominals might be the research design itself; deliberately the Subjects were allowed only a two-way choice between Peter and John for "him" even though it might refer to somebody not mentioned in the sentence, as is possible for any pronominal.  Some of the native Subjects indeed pointed out after they had been tested that they felt that this had sometimes led to extra processing time.  The discrepancy between the present results and those surveyed in Lust et al (1989) is presumably linked to the task involved: elicited imitation as a task may involve a form of processing in which the smaller domain for anaphors is helpful; comprehension may involve processes in which it is not.

ii) are pronominals treated as anaphors?

The evidence for pronominals being treated as anaphors would be results for pronominals that were the mirror of those for anaphors: while the score for anaphors for a given sentence type should tend to 100% correctness, the score for pronominals should tend to 100% incorrectness.  A quick glance at the overall results in table 1 shows that such complementarity is never the case; the highest error rates for pronominals are 24% (Japanese on I), 28.7% (Japanese on G), 23.4% (Japanese on E), and 20.6% (Norwegian on I), the rest being much lower; mostly where pronominal error rates are high, so are anaphor rates, for example E/F (Japanese, 23.4% and 40% respectively).  The overall higher error rate for anaphors on sentence types E/F and G/H hardly implies that pronominals are being treated as anaphors.  Since the research design concerned only advanced learners, the possibility of such an interpretation at earlier stages of L2 development is not ruled out.  However arguments based on the misinterpretation of pronominals as anaphors are clearly otiose at this stage.

Figure 1. Error rates and timings for anaphor

Figure 2  Error rates and timings for pronominals

iii) are there differences across a range of sentence types?

The results for the anaphor sentence types B, D, F, H are given in the graph in Fig 1.  Let us take 11% as an arbitrary cut-off point for random noise, chosen as it seemed to represent the limit for non-systematic errors in natives.  The order of difficulty for errors in natives is (B/D/F)H (chi squared test on raw scores, p.<0.02); for Romance speakers and for Norwegians (B/D)FH (chi squared test, p.<0.001); for Japanese BDFH (chi squared test, p.<0.001).  For response times the order from quickest to slowest is BDFH for all groups (ANOVA natives, F=8, p.<0.01; Romance F=16, p.<0.01; Japanese F=27, p.<0.01) except the Norwegians who have BFDH (ANOVA, F=22, p.<0.01).  So, although there are differences in the overall level of performance between the groups affecting where the cutoff point comes, the difficulty order between the sentence types is the same for both comprehension and response times for all three groups, with the exception of the Norwegian group's performance on DF for response time.  The overall order for anaphors is therefore BDFH.  To translate this back into the actual sentence types, the easiest is the simple B "John shot himself"; the next the tensed D "Peter said that he voted for himself"; the next the infinitival F "Peter asked John to include himself"; the last the noun phrase H "John reported Peter's criticisms of himself". 

The graph for the pronominal "him" in the four main sentence types ACEG is given in Fig 2.  Using 11% again as the arbitrary cutoff point for random noise, the order for natives is (A/C)EG (nonsignificant, chi squared test); for Romance (A/C/E)G (chi squared test, p.<0.01); for Japanese (A/C)EG (chi squared test, df 2, p.<0.05), and for Norwegian (A/C)GE (nonsignificant); the overall order for comprehension is ACEG (chi squared test, df 3, p.<0.001).  The order in response times is ACEG (ANOVAs natives, F=11, p.<0.01; Romance F=2, nonsignificant; Japanese, F=16, p.<0.01)), except for the Norwegians who have AECG (F=6.47, p.<0.01).  Put back into sentences, the easiest is the simple A "Peter shot him", next the tensed C "Peter said that John voted for him"; next the infinitival E "John asked Peter to pay for him"; finally the noun phrase G "John discovered Peter's report on him".

The results for anaphor sentence type J can be added for purposes of completeness.  Only the Japanese error rate is higher than the cutoff point, putting it between B and D; in response times it comes between D and F.  The pronominal sentence type I had higher error rates than usual for Japanese and Norwegian, placing it last for Norwegians, and between A & C for Japanese; its response times put it last for natives and Norwegians, between E and G for Romance and between A and C for Japanese.

To sum up, there is a clear difference between the difficulty of the sentence types across groups, consistent for both errors and response times for anaphors and for pronominals.  While the cutoff point for error rates varies slightly, the response times form a smooth progression for all but the Norwegian group on E and F.

iv) does the L1 of the learner affect the results and in what way? 

In one sense the research does not permit quantitative comparisons between groups since the level of English was not necessarily equal.  But tests of significance show that the error rates of the groups are not different on A, B, C, E, G and J; they differ on D (chi squared test, p.<0.01), F (p.<0.001), and H (p.<0.05); putting aside the native group, only D and F remain significant.  In other words the groups behave the same on the pronominal sentences but have differences on two of the anaphor types, D and F.  It does not seem on the basis of the results here that large quantitative differences exist between the groups, even if this could only be firmly established by an experiment that controls for language level more precisely.  Again this does not mean that at earlier stages such differences could not be found.

It is however perfectly proper to compare order of difficulty of sentence types as opposed to absolute measures; while the error rates for sentence types varies between groups, speakers from all backgrounds show a similar order of difficulty, as we have already seen.  A common difficulty order for anaphors and pronominals on both measures for all groups with few exceptions suggests slight effect from the L1.

Having made these general points and bearing in mind the caveat about possible differences in level of English between the groups, let us look at some more detailed differences between them.  For anaphors there are significant differences in error rates between the four groups for two sentence type, D & F; with D the Japanese are much worse than the other groups, perhaps due to the (e) setting for "zibun"; for F the groups are more spread out, again with the Japanese being worst.  Though the native group is always faster than any of the non-native groups, the timings of all three non-native groups are close for each sentence type, with the exception of the Norwegians being faster on the infinitival F and slower on the noun phrase H.  For pronominals there are no significant differences in error rate between the four groups; the response times were similar for all non-native groups with Norwegians slightly faster on all but A and with the natives consistently faster than the rest.  Again a quick glance at Figures 1 and 2 overwhelmingly confirms the basic similarity between all groups for both measures and for both pronominals and anaphors.

How do these results relate to the L1 setting of the governing category parameter?  This means looking at the differences between the groups in terms of their L1 settings for the governing category parameter for the equivalent anaphors and pronominals.  Take the Japanese 43.7% error rate on NP sentences (H); this might be thought to be a clear effect of an (e) setting transfer; but looking at the other groups, the Romance with a (b) setting have 35.7%, only 8% better, and for both groups their most difficult sentence type.  (However let us not forget that all the L1s actually behave the same since none of them have the English (a) setting in this context).  The only significant difference in error rate was for the infinitival F sentence type, which should not have distinguished the groups in terms of the governing category setting.  The faster times of the Norwegians on the infinitival F is hardly explained by their (c) setting for "seg", which should have made it more difficult for them than for Romance speakers rather than the reverse.  So at this level of English effects of L1 setting are not prominent; in quantitative comparisons the overall high error rate for the Japanese might be attributed to their (e) setting for "zibun"; the slightly lower error rate for the Romance might be due to their (b) setting.  Yet the overall similarities in difficulty order for all groups show no effects of setting; perhaps results such as the Norwegians' comparatively fast speed for the infinitival anaphor F are an indirect consequence of their (c) setting.  But again the point that comes across from the results is their uniformity.

(v) how is the Subset Principle involved in L2 learning and using?

Let us now see whether the difference in results between sentence types can be related to the Subset Principle.  The main caveats to be made are first that the sentences are all English and hence have the settings for the governing category parameter for English, and second that for many of the sentences the difficulty for the Subjects is that their equivalent L1 setting yields two possible antecedents in English rather than uniquely giving the wrong answer.  The direct effects of the Subset Principle are hard to see.  Since English has the (a) setting for the anaphor "himself" any other setting represents a mistake and a breach of the Subset Principle by generating too large a 'language'.  A possible order of difficulty if L2 learners were dominated by the Subset Principle for anaphors would be BDFH, this being in fact the actual order.  But pronominals should then have the reverse order, GECA, rather than the order actually found, namely ACEG.  The changes in sentence type have similar effects for both anaphors and pronominals.  If the Subset Principle were involved in any way, it would make opposite predictions for anaphors and pronominals, which are not found. 

Is there any justification for the view that L2 learners have settings intermediate between their L1 and L2?  To my mind this would be settled by findings of much greater consistency among L2 learners, rather than error rates that never exceed 45%, and of much greater differences between learners with different L1s, rather than the overall similarities seen here.  Nevertheless it should be pointed out that Subjects here were advanced speakers tested at a single moment in time.  The true test of the Subset Principle may be whether L2 learners start with a minimal setting and gradually expand it, rather than whether they retain certain difficulties after they become fluent speakers.  A biassing factor in this research, and indeed in all the L2 research, is that it is based on L2 learners with settings for their equivalents to "himself" other than (a), the most restricted environment for anaphors, learning English, where "himself" is (a).  It would be useful to investigate the reverse phenomenon of how English L1 speakers learn languagse with other settings, say Japanese.

Conclusions

The general issue of the use of timed comprehension judgements can now be raised.  The Universal Grammar theory relies on declarative statements about language knowledge and is not concerned with procedural processes involved in language comprehension and production: it is concerned with grammatical competence rather than with performance.  Its typical form of evidence consists of the native speaker knowing that a particular sentence fits his or her knowledge of the language.  Hence UG theory has difficulty in deciding what counts as permissible evidence to show what the speaker knows when dealing with actual language development rather than with the logical argument of language acquisition.  Cook (to appear) discusses the problems for UG theory of using observational data taken from actual conversations and argues that only certain kinds of statement can be validly made from such data, essentially those supported by sentences that are representative of the learner's speech rather than isolated examples and by positive data of what is actually present in the sentence rather than of what is absent from it, such as 'missing' words or endings, since these could be due to a multitude of causes other than grammatical competence.  L2 learning research in a UG paradigm has tended to use grammaticality judgements, for example White's work with pro-drop (White, 1986).   A grammaticality judgement is no more central to UG than any other experimental technique; it is tarred with the same brush of performance (Carroll et al, 1981).  As we saw earlier, grammaticality judgements are inappropriate for binding since incorrect binding still leaves a grammatical sentence (under some other interpretation). 

So a comprehension test was employed here, as used by Solan (1987) and Finer and Broselow (1986).  This has already shifted the focus from knowledge of language to understanding of sentences; like grammaticality judgements, comprehension is an experimental technique that involves performance, and hence only indirectly linked to grammatical competence.  A second measure of response time was introduced, a standard psycholinguistic technique.  Overall there was a significant correlation between comprehension and response time (Kendall's tau, 0.51, p.<0.05); the two measures also correlated in terms of the difficulty orders for sentence types for pronominals and anaphors, as a glance at the graphs in figs 1 and 2 confirms.  Error rates in comprehension and timings do seem to measure the same performance difficulty.  The same is true of grammaticality judgements; measurements of the time taken go well with success or failure of judgement, though not necessarily meaning that correct judgements take less time (Cook, 1989).  While a purist may attack a timed comprehension task as inconsistent with a UG approach, this objection applies equally to all L1 and L2 learning research that uses anything but information about whether a speaker 'knows' that a sentence fits his or her grammar - observational data, grammaticality judgements, and all the rest.

Let us then consider the results as a performance task.  One point that immediately emerges is the 'cognitive deficit' of advanced L2 learners.  Though their level of performance in terms of errors is hardly distinguishable from native speakers, they take consistently longer to do it.  Another task can be added to the range surveyed in Cook (1988b) in which non-native performance is consistently below that native performance on reaction times, mental arithmetic, and so on.

More particularly, let us follow up the tentative suggestion that a BDFH order of difficulty for anaphors and ACEG for pronominals corresponds to the sequence of settings from (a) to (e) in the governing category parameter.  Looked at in processing terms one possible cause of this might be sentence length.  The simple A and B sentence types have less words and consist of a single sentence so there are straightforward reasons for them being easiest and quickest: the other sentences however vary from 6 to 7 words in length, and all include embedded sentences or NPs; the shorter sentences were not found to be easier than the longer ones.  So length is not in itself crucial.  Another possibility might be that the presence of certain clues aids processing:
- finding a subject verb construction helps most (A/B);
- finding a tense helps next (C/D);
- finding an INFL helps next (E/F);
- not finding any of these hinders most (G/H). 
The sentence types J/I can now be put back into the sequence: these have subject verb but no INFL or Tense on the surface; hence they should occur next most easily after simple subject-verb sentences, as they indeed do for anaphors (though not for pronominals).  The hierarchy of sentence types for binding can be arranged in order of difficulty from those with subject verb to those without INFL or Tense.  Indeed Koster (1984) has proposed a locality principle that each position from (a) to (e) adds another specification for the governing category.

Let us go further outside orthodox UG theory to consider an alternative way of looking at parameters.  As Solan (1987, p.200) puts it, 'A number of differences between various of the sentence types, while inviting explanations based on deep principles of universal grammar, appear to be better explained with reference to ease of processing and language particular phenomena.'  The sequence of performance difficulty just outlined has a suspicious resemblance to the settings for the governing category parameter expressed in a slightly different way.  One interpretation is that language processing reflects the settings for the governing category parameter which are still latent in the mind; difficulty is caused the more that the sentence could have ambiguous meanings for different settings.  In this view the setting of multi-valued parameters does not just take place once-and-for-all during L1 or L2 acquisition but the different settings are somehow still active in using the language.  Cook (to appear) argues that the alternative settings for the two-valued pro-drop parameter are latent in native English speakers in that they produce null-subject sentences in performance and understand null-subject sentences in certain registers, for example the prose of J.P. Donleavy.  Cook (1989) shows that, while advanced Japanese learners make few mistakes with grammaticality judgements about "on the table" versus "the table on", their L1 parameter setting influences how fast they do so, in this case by making them quicker to make a decision about "the table on" than non-native speakers with other L1 settings.  Cook (in progress) has been developing a computer simulation that treats Universal Grammar as a parser that contains the principles of word order but continuously sets and resets word order parameters from evidence as it processes language; in other words it treats parameters as part of the parsing process.  The present results might suggest that the alternative settings for another parameter still influence language processing; they are still latent within the processing mechanism - what chaos theory might call a 'strange attractor' (Gleick, 1987).  A bilingual is not just two monolinguals with two different settings; rather he or she possesses a single overall grammar that encompasses two or more settings for each parameter.  A performance model of parameters needs a wider discussion and justification than can be presented here; it is not dissimilar to the arguments in Hawkins (to appear) about the relevance of parsing to the grammar.

In terms of the aims set for this paper the study of binding in L2 learning has been put on a broader factual base: anaphors tend to be more difficult and to take longer to process than pronominals at the advanced level of English investigated; pronominals are not treated as anaphors; there are differences in difficulty order across a range of sentence types; the L1 of the learner has slight effect; though the design of the present research makes such links tentative, the Subset Principle can be related to some extent only to anaphors.  As always a wider range of research is needed to extend the approach from the comprehension of advanced learners to the stages during which such comprehension develops, to the other parameters involved, and to languages other than English as the L2.  One lesson to be learnt is perhaps familiar: all statements about L2 development depend on some test of performance; we have always to take into account the relevant aspects of performance by native and non-native speakers before drawing conclusions for the theory of Universal Grammar.  Another lesson is also salutary; to put the study of Universal Grammar in L2 learning on a proper foundation means tackling areas of syntax that are central to the development of Government/Binding theory; because of the massive interaction in GB between different aspects of the grammar and because of the difficult argumentation connecting acquisitional evidence and UG such research is complex and fraught with problems of interpretation. 


References

Berwick, R.C. (1985), The Acquisition of Syntactic Knowledge, MIT Press

Berwick, R.C., & Weinberg, A.S. (1984), The Grammatical Basis of
   Linguistic Performance, MIT Press

Carroll, J. M., Bever, T.G., & Pollack, C. R. (1981), 'The non-uniqueness  of linguistic intuitions', Language, 57, 2, 368-383

Chomsky, N. (1981), Lectures on Government and Binding, Foris, Dordrecht

Chomsky, N. (1986), Knowledge of Language: Its Nature, Origin and Use, New  York, Praeger

Chomsky, N. (1988), Language and Problems of Knowledge: The Managua
   Lectures, MIT Press

Chomsky, N. (1989), 'Some notes on economy of derivation and
   representation', MIT Working Papers in Linguistics, 10, 43-74

Cook, V.J. (1988a), Chomsky's Universal Grammar: an Introduction,  Blackwell, Oxford [Latest edition 2007]

Cook, V.J. (1988b), 'Cognitive processing and second language learning',  polyglot, 9, fiche 2

Cook, V.J. (1989), 'Word order and timed grammaticality judgements', paper read at the LARS Conference, Utrecht

Cook, V.J. (to appear), 'Observational data and the UG theory of language  acquisition', in I. Roca (ed.), Logical Issues in Language
Acquisition, Foris
Cook, V.J. (in progress), 'Universal Grammar considered as a parser that  acquires language'

Deutsch, W., Koster, C., & Koster, J. (1986), 'What can we learn from  children's errors in understanding anaphora?', Linguistics, 24, 203-225

Dulay, H.C., & Burt, M.K. (1978), 'From research to method in bilingual education', in J. Alatis, (ed.), Georgetown University Round Table, 551-575

Finer, D. (1987), 'Comments on Solan', in T. Roeper & E. Williams (eds.), Parameters and Linguistic Theory, Dordrecht, Reidel

Finer, D. & Broselow, E. (1986), 'Second language acquisition of reflexive binding', NELS, 16

Gleick, J. (1988), Chaos, Penguin, Harmondsworth

Jakubowicz, C. (1984), 'On markedness and binding principles', NELS, 14

Hawkins, J.A. (to appear), 'A parsing theory of word order universals', Linguistic Inquiry

Hellan, L. (1988), Anaphora in Norwegian and the Theory of Grammar, Foris,  Dordrecht

Hirakawa, M. (1989), 'A study of the L2 acquisition of English
   reflexives'.  Paper presented at the Conference on the Interaction of Linguistics, Second Language Acquisition, and Speech Pathology,
   University of Wisconsin-Milwaukeee, April

Hyams, N. (1986), Language Acquisition and the Theory of Parameters,  Dordrecht, Reidel

Koster, J., (1984), 'On Binding and Control,' Linguistic Inquiry, 15, 3,  417-459

Lust, B., Mazuka, R., Martahardjone, G., and Yoon, J.M. (1989), 'On  parameter setting in first language acquisition: the case of the
   Binding Theory' paper presented at GLOW, Utrecht

Matthei, E. (1981), 'Children's interpretation of sentences containing
   reciprocals', in Tavakolian, S.L. (ed.) (1981), Language Acquisition
   and Linguistic Theory, MIT Press

Mitchell, D.C., & Barchan, J. (1984), EXMORE: Exeter Module for On-line Reading Experiments, University of Exeter, Psychology Department

Read, C., & Hare, V. (1979), 'Children's interpretation of reflexive
   pronouns in English', in F. Eckmann & A.J. Hastings (eds.), Studies in
   First and Second Language Acquisition
, Newbury House, pp. 98-116

Saleemi, A. (1988), Learnability and Parameter Fixation, Ph.D., University   of Essex
Solan, L. (1983), Pronominal Reference, D.Reidel, Dordrecht

Solan, L. (1987), 'Parameter setting and the development of pronouns and  reflexives', in T. Roeper & E. Williams (eds.),  Parameters and Linguistic Theory, Dordrecht, Reidel, pp.

Thomas, M. (1989), 'The interpretation of English reflexive pronouns by non-native speakers', Studies in Second Language Acquisition, 11, 3

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

White, L. (1986), 'Implications of parametric variation for adult second  language acquisition: an investigation of the pro-drop parameter', in V.J. Cook (ed.), Experimental Approaches to Second Language Acquisition, Oxford, Pergamon

White, L. (to appear), 'Implications of learnability theories for second language learning and teaching', in M.A.K. Halliday, J. Gibbons, and H. Nicholas (eds.). Learning, Keeping, and Using Language, Amsterdam, John  Benjamins

Footnote

1. I am grateful to the students and staff of the Norwegian Studies Centre, University of York, the Bell's School, London, and the Summer Course for Japanese Teachers at the University of Essex for their cooperation and help with this experiment; to Dr D. Mitchell of Exeter University for permission to use the EXMORE program; and to Lydia White, Mark Newson, and anonymous reviewers for crucial comments on earlier drafts.

Appendix: Test sentences
Letters refer to sentence types explained in the text

( 1.A)   Peter shot him
( 2.A)   John hurt him
( 3.A)   Peter blamed him
( 4.A)   John painted him
( 5.B)   John shot himself
( 6.B)   Peter shot himself
( 7.B)   Peter blamed himself
( 8.B)   John painted himself
( 9.C)   Peter said that John voted for him
(10.C)   John says that Peter talked about him
(11.C)   Peter believes that John laughed at him
(12.C)   John believed that Peter helped him
(13.D)   Peter said that John voted for himself
(14.D)   John says that Peter talked about himself
(15.D)   Peter believes that John laughed at himself
(16.D)   John believed that Peter helped himself
(17.E)   John asked Peter to pay for him
(18.E)   Peter asked John to include him
(19.E)   John told Peter to support him
(20.E)   John told Peter to mention him
(21.F)   Peter asked John to include himself
(22.F)   John asked Peter to pay for himself
(23.F)   John told Peter to support himself
(24.F)   Peter told John to mention himself
(25.G)   Peter discovered John's report on him
(26.G)   John discovered Peter's betrayal of him
(27.G)   Peter reported John's comments on him
(28.G)   John reported Peter's criticisms of him
(29.H)   John reported Peter's criticisms of himself
(30.H)   Peter reported John's comments on himself
(31.H)   John discovered Peter's betrayal of himself
(32.H)   Peter discovered John's report on himself
(33.I)   Peter saw John save him
(34.I)   Peter heard John kick him
(35.I)   John saw Peter hit him
(36.I)   John heard Peter praise him
(37.J)   Peter saw John save himself
(38.J)   John saw Peter hit himself
(39.J)   Peter heard John kick himself
(40.J)   John heard Peter praise himself