Criterial Features and the CASP Model of SLA

Criterial Features and the CASP Model of SLA John A. Hawkins Department of Linguistics, UC Davis

From 2005 -2011 I co-directed a research program on second language learning at Cambridge University (the “English Profile Programme”) with three components: (a) Cambridge Learner Corpus [CLC]: 40+ million words of written English from learners around the world; (b) Computational techniques: CLC was first searchable lexically, with 76 error codes; subsequently tagged for parts of speech and parsed using an automatic parser, RASP, Briscoe et al 2006; (c) New research methodology: designed to yield practical benefits for learning/teaching/assessment, and to make a theoretical contribution to SLA.

The key theoretical and methodological innovations include: (i) the concept of criterial features, as a means of distinguishing levels of proficiency in the learning of a second language; (ii) the development of a general learning model, CASP (“complex adaptive system principles of SLA”) informed by the criterial features of the CLC and by other empirical studies in the literature. These two innovations, especially CASP, were developed on the basis of joint work with my principal collaborator, Luna Filipović (then Cambridge, now University of East Anglia). Further assistance from members of the Cambridge team is recognized in the Acknowledgements at the end of this talk.

Criterial features are properties of learners’ English that are distinctive and characteristic of L 2 proficiency at the different levels. The CLC scripts have been graded by Cambridge examiners and assigned grades A-F at each of six levels of proficiency, following the Common European Framework of Reference (CEFR), see the Council of Europe 2001 : Levels: C 2 C 1 B 2 B 1 A 2 A 1 Mastery [CPE] Effective Operational Proficiency [CAE] Vantage [FCE] Threshold [PET] Waystage [KET] Breakthrough When searching for criterial differences between levels we focussed on the scripts of students who had achieved passing grades of A-C at each level.

The questions that motivated this research are: 1) how much of the grammar and lexicon of English do learners actually know and/or produce at each of these CEFR levels? 2) what empirical patterns and principles are there in these developing second language stages of English? and 3) what are the practical benefits, for learning, teaching and assessment of gathering this information?

Electronic corpora of learner English make it possible for us to answer these questions. The CLC is the biggest learner corpus of English or of any language. It gives us empirical evidence for developmental stages in the learning of new constructions, words and word meanings. It gives us quantitative data on learner errors in syntax, morpho-syntax and lexical choice. It was originally searchable only lexically, in conjunction with the error codes. Subsequently it has been tagged and parsed using the automatic parser, RASP.

Sample Error Codes in the CLC RN Replace noun Have a good travel (journey) RV Replace verb I existed last weekend in London (spent) MD Missing determiner I spoke to President (the) I have car (a) AGV Verb agreement error The three birds is singing (are) IV Incorrect Verb Inflection I spended last week in London (spent) FJ Wrong Adjective Form The situation got worst (worse) UQ Unnecessary Quantifier A little bit quite common (quite common) DY Derivation of Adverb It happened fastly (fast)

Briscoe’s RASP (Robust Accurate Statistical Parser) • • • identifies parts of speech (Po. S) probabilistically tagging generates a parse forest representation containing all possible subanalyses with associated probabilities weighted Grammatical Relations yielded by the nbest parses of the input.

Criterial features can be found in all areas of English: syntax, morphology, phonology, the lexicon, semantics, and discourse. They distinguish higher proficiency levels from lower levels in an efficient way. In this talk some of these features are illustrated, as is theoretical learning model derived from them, “CASP”.

An analogy: Languages change over time, and when historians of English examine Old English, Middle English and Early Modern English, they focus on important differences between these stages, not on what stayed the same. Similarly in a learning context, we are interested in changes from one level to another.

There are different types of criterial features. Here we focus on just two: Positive Criterial Features These refer to positive, i. e. correct, linguistic properties of English that have been acquired at a certain L 2 level and that generally persist at all higher levels. A property P (e. g. a new construction type) acquired at B 2 may differentiate that level and higher levels from [A 1, A 2, B 1] and will be criterial for the former. Or P may be acquired at C 2 and differentiate this level from all lower levels.

Negative Criterial Features These are incorrect properties of English, or errors, that occur at a certain level or levels and with a characteristic frequency. Both the presence versus absence of the errors, and especially their frequency (the "error bandwidth"), can be criterial for the level(s).

Examples of Positive Criterial Features (from Hawkins & Filipović 2012 and using data from the CLC) The A levels (A 1 and A 2) Simple intransitive clauses (NP-V) and the slightly more complex transitive (NP-V-NP) sentence types are present from the beginning: He went. (NP-V) A 1 He loved her. (NP-V-NP) A 1

Modal auxiliary verbs like may, might, can and must appear first at A 1 or A 2, but only in some of their senses. Can is first attested in the PERMISSION sense at A 1 and in the POSSIBILITY sense at A 2: And if you want, you can bring pencils or pens. (PERMISSION) A 1 In this magazine you can see all the new C. D. [s] and all the dates of the concerts. (POSSIBILITY) A 2

Noun Phrase sequences of Pronoun plus Infinitive are found at A 2: something to eat nothing to do A 2 as are postnominal modifiers with participial –ed: beautiful paintings [painted by famous Iranian painters] A 2

Lexical verbs appearing at the A levels are typically among the most basic and frequent verbs of English; they appear first in their most basic and frequent senses. Verbs attested at A 1 include: catch, eat, give, put, take and walk

New lexical verbs appearing at A 2 include: break, cut, hit, push, stand, and fall again typically in their most basic and literal senses. For break this includes its primary physical sense: I broke a beautiful glass. A 2 for cut it includes the following example in its primary sense: First I cut the cake with my mother. A 2

The B Levels (B 1 and B 2) The new features at B 1 involve more complex syntax, e. g. an “Object Control” structure such as: I ordered him [to gather my men to the hall] B 1 him is both the object of ordered and the logical subject of gather here. This is a criterial construction for B 1 and higher levels which distinguishes them from the A levels.

Structures like the following with finite or non-finite subordinate clauses and movement of the WH-word (how, where, etc) to the front of its clause are also first attested at B 1: I don’t know [how I could have done it] I did not know [where to look for it] B 1

And postnominal modifiers in participial –ing become productive at B 1: I received your mail [asking for the sales report] B 1

Structures with a finite subordinate clause positioned to the right of predicates like is true and seems with a subject it are also criterial for B 1 and higher levels: It’s true [that I don’t need a ring to make me remember you] B 1 i. e. so-called “Extraposition” structures

A large number of new lexical verbs appear for the first time at B 1 including: divide, fit, grab, spill, stick and tear And the meanings of the verbs that appeared first at A 1 and A 2 begin to expand from their basic senses.

break appears for the first time in the extended sense of INTERRUPT at B 1: At last I managed to break the routine of the city … B 1

Constructions that are criterial for B 2 and higher levels include “secondary predications” go and paint the houses yellow and blue B 2 with yellow and blue predicated of the direct object houses

“Extraposition” structures with a non-finite subordinate clause positioned to the right of its predicate are B 2 It would be helpful [to work in your group as well] B 2

And so-called “Pseudocleft” structures with an initial what functioning as subject of its verb: What fascinated me was [that I was able to lie on the sea surface] B 2

“Subject-to-Subject Raising” constructions appear first at B 2 with most of the higher verbs and adjectives that trigger this rule, for example prove: The car has proved [to be one of the most important inventions of our century] B 2 Similar examples are found at B 2 with other raising verbs and adjectives (The car happened to be …, The car appeared to be …, The car turned out to be …, The car is likely to be …, etc)

New lexical verbs at B 2 include acquire, capture, drag, rush, spread, swallow and new meanings and uses are attested for the verbs that appeared earlier.

For break, first attested at A 2, these include new collocations such as break a promise or break the law: For cut, also an A 2 verb, they include new meanings at B 2 such as REDUCE in cut the cost

The C Levels (C 1 and C 2) “Subject-to-Object Raising” constructions with the verb believe appear first at C 1 and are criterial for the C levels: I believe her [to be this country’s best representative] C 1

Passivized Subject-to-Object Raising constructions such as the following with assumed are also criterial for C 1: the low cost of membership and entry was assumed to be an advantage. C 1

Sequences of two prenominal –s genitives are found at C 1: in the bride’s family’s house C 1 Structurally: in [[[the bride’s] family’s] house]

New lexical verbs appearing first at C 1 include accumulate, boast, quote, reassure, shape and stain along with new meaning possibilities for the verbs already introduced. E. g. break appears first in the idiomatic sense of break the bank at C 1.

New features appearing at C 2 include less common Subject-to-Object Raising constructions with higher predicates such as presume, declare and remember: He presumed work [to be the way to live] C 2

New lexical verbs at C 2 include stagger, sway, limp, saunter, raid, squander New meanings for break at C 2 include original figurative senses such as the attested break the wall that surrounds him.

Negative Criterial Features One major distinguishing feature of the C levels can be seen in the low frequencies for “negative features” or error types such as those illustrated above. There are significant improvements in ALL of the syntactic and morpho-syntactic error types at the C levels.

By contrast, at the B levels improvements are relatively modest, and for many error types the scores actually get worse, especially at B 2, before they get better again at C 1.

The error codes involve morpho-syntactic errors of inflection, derivation and grammatical form, syntactic errors of omission, positioning and co-occurrence, and errors of appropriate lexical choice. It is clear that learners at the C levels are increasingly mastering these rules of English, whereas B-level learners are not (see Hawkins & Filipović 2012 for details).

We must now ask: WHY do we see these patterns in the data and why do we see the criterial features changing the way they do at the different levels? In particular, WHAT is it about the features of the higher proficiency levels that makes them late acquired rather than early?

It cannot simply be that learners are imitating the words and constructions they are explicitly taught in their textbooks. First, because there are many different textbooks and teaching methods around the world. But secondly because learners learn more than they are explicitly taught, from their reading materials, papers, magazines, movies, TV, conversations, etc. I. e. second language learning shares many similarities with first language learning, but not all obviously.

For example, more frequently occurring words and constructions are learned before less frequent ones, and simpler words, constructions and meanings are learned before more complex ones, in both first and second language acquisition.

E. g. learning English nouns and verbs with high frequencies of use is easier than learning those with lower frequencies, because they are encountered more frequently (greater exposure); frequent lexical items are overrepresented at first in L 2 English, moving gradually to L 1 English norms (see Hawkins & Buttery 2009, Hawkins & Filipović 2012)

The constructions of English that are learned earliest are those that occur most frequently in the input, as reflected in e. g. the British National Corpus. This could be established by comparing the CLC with the British National Corpus (BNC), see Williams (2007).

The new constructions that are criterial for A 2, B 1 and B 2, in Williams’ (2007) data, appear to be learned in direct proportion to their frequency in the input, as reflected in the BNC. The more exposure, the earlier the acquisition and the easier the learning. This is shown in Tables 1, 2 and 3. Table 1 lists the new construction types found, for example, at B 1. Tables 2 and 3 give the frequency correlations between the CLC and the BNC for the different levels.

Table 1 New B 1 Verb Co-occurrence Frames NP-V-NP-NP She asked him [his name] NP-V-VPinfin (Wh-move) He explained [how to do it] NP-V-NP-V(+ing) (Obj Control) I caught him stealing NP-V-NP-PP (P=to) (Subtype: Dative Movement) He gave [a big kiss] [to his mother] NP-V-NP-(to be)-NP (Subj to Obj Raising) I found him (to be) a good doctor NP-V-NP-Vpastpart (V=passive) (Obj Control) He wanted [the children] found NP-V-P-Ving-NP (V=+ing) (Subj Control) They failed in attempting the climb NP-V-Part-NP-PP I separated out [the three boys] [from the crowd] NP-V-NP-Part-PP I separated [the three boys] out [from the crowd] NP-V-S (Wh-move) He asked [how she did it] NP-V-PP-S They admitted [to the authorities] [that they had entered illegally] NP-V-Part She gave up NP-V-S (whether = Wh-move) He asked [whether he should come] NP-V-P-S (whether = Wh-move) He thought about [whether he wanted to go]

Table 2 Frequencies for Verb Co-occurrence Frames in English Corpora (including BNC) Average Token Frequencies in the BNC for the new Verb Co-occurrence Frames appearing at the learner levels A 2 1, 041, 634 B 1 38, 174 B 2/C 1/C 2 27, 615

Table 3 Frequency Ranking Average Frequency Ranking in the BNC for the new Verb Co-occurrence Frames appearing at the learner levels A 2 8. 2 B 1 38. 6 B 2/C 1/C 2 55. 6

These kinds of data enable us to set up the following principle of second language learning (for which there also well-attested parallels in first language learning, see e. g. Tomasello 2003, Diessel 2004, Mac. Whinney 2005):

(1) Maximize Frequently Occurring Properties (Ma. F) Properties of the L 2 are learned in proportion to their frequency of occurrence (as measured, for example, in the BNC): more frequent exposure of a property to the learner facilitates its learning and reduces learning effort. I. e. more frequent properties will result in earlier L 2 acquisition, more of the relevant properties learned, and fewer errors, in general. Infrequency makes learning more effortful, with precise predictions depending on other factors.

Highly correlated with frequency is another principle of second language learning (shared with first language learning) that involves the relative simplicity or complexity of structures and meanings. The criterial grammatical features of earlier levels are, in general, simpler than those of later levels.

In phonology also simpler consonants and consonantal distinctions are acquired earlier than more complex ones (see e. g. Eckman 1984). Simpler and more basic meanings for verbs are acquired earlier than more complex and derived extensions in meaning, figurative uses, etc.

The verb break in its basic physical sense at A 2 break in the sense of INTERRUPT (break the routine) B 1 break an agreement, promise, etc. B 2 break the bank (idiomatic) C 1 break the wall that surrounds him (original figurative) C 2

(2) Maximize Structurally and Semantically Simple Properties (Ma. S) Properties of the L 2 are learned in proportion to their structural and semantic simplicity: simplicity means there are fewer properties to be learned and less learning effort is required. I. e. simpler properties will result in earlier L 2 acquisition, more of the relevant properties learned, and fewer errors. Complexity makes learning more effortful, in general, since there are more properties to be learned, with precise predictions depending on other factors.

In second language learning we also see ‘transfer’ effects from the first language, either positive (when the transfer results in a correct L 2 property) or negative (when it results in an error). This is one thing that differentiates second from first language acquisition.

E. g. speakers of languages with definite and indefinite articles find it easier to acquire the article system of English than do speakers of languages without articles (see Hawkins & Buttery 2009, 2010)

Errors involving missing definite and indefinite articles in the L 2 English of the CLC are consistently low when the L 1 s also have articles. Recall: MD I spoke to President (the) I have car (a)

Table 4 (next slide) shows missing determiner error rates for “the” and “a” at all proficiency levels for French, German and Spanish as first languages. All three languages have an article system. (Data from Hawkins & Buttery 2009) The figures indicate the percentage of errors with respect to the total number of correct uses. For instance a percentage of 10. 0% would indicate that a determiner was omitted 1 in every 10 times that it should have appeared. We see generally low error rates for these languages, without significant deviation between levels.

Table 4 Missing Determiner Error Rates for L 1 s with Articles French German Spanish A 2 4. 76 0. 00 3. 37 Missing “the” B 1 B 2 4. 67 5. 01 2. 56 4. 11 3. 62 4. 76 C 1 3. 11 3. 22 C 2 2. 13 1. 60 2. 21 A 2 6. 60 0. 89 4. 52 Missing “a” B 1 4. 79 2. 90 4. 28 C 1 4. 76 3. 62 5. 16 C 2 3. 41 2. 02 3. 58 B 2 6. 56 3. 83 7. 91

Table 5 (next slide) shows missing determiner error rates for “the” and “a” at all levels for Turkish, Japanese, Korean, Russian and Chinese as first languages. These languages do not have an article system. There is a general linear improvement, i. e. a decline, in error rates across the levels with increasing proficiency (shown from left to right). Chinese shows an interesting inverted U-shaped progression, especially in the case of missing “a”.

Table 5 Missing Determiner Error Rates for L 1 s without Articles Turkish Japanese Korean Russian Chinese Missing “the” A 2 B 1 B 2 22. 06 20. 75 21. 32 27. 66 25. 91 18. 72 22. 58 23. 83 18. 13 14. 63 22. 73 18. 45 12. 41 9. 15 9. 62 C 1 14. 44 13. 80 17. 48 14. 62 12. 91 C 2 7. 56 9. 32 10. 38 9. 57 4. 78

Table 5 continued Turkish Japanese Korean Russian Chinese Missing “a” A 2 B 1 24. 29 27. 63 35. 09 34. 80 35. 29 42. 33 21. 71 30. 17 4. 09 9. 20 B 2 32. 48 24. 26 30. 65 26. 37 20. 69 C 1 23. 89 27. 41 32. 56 20. 82 26. 78 C 2 11. 86 15. 56 22. 23 12. 69 9. 79

One of the learning principles proposed in Hawkins & Filipović (2012) to account for these data is Maximize Positive Transfer:

(3) Maximize Positive Transfer (Ma. PT) Properties of the L 1 which are also present in the L 2 are learned more easily and with less learning effort, and are readily transferred, on account of pre-existing knowledge in L 1. Shared L 1/L 2 properties should result, in general, in earlier L 2 acquisition, in more of the relevant properties being learned, and in fewer errors, unless these shared properties involve e. g. high complexity and are impacted by other factors. Dissimilar L 1/L 2 properties will be harder to learn by virtue of the additional learning that is required, again in general.

More generally, Filipović & Hawkins (2013) provide a multi-factor model of learning, supported and informed by data in the CLC, and comprising a set of interacting principles such as those illustrated.

The model is a type of “complex adaptive system” (see Gell-Mann 1992) and is called the “CASP” model, short for “complex adaptive system principles of SLA”.

The principles interact, sometimes reinforcing each other (e. g. early acquired frequent items are also often simple), sometimes competing to produce variable outputs and alternative interlanguages. Some of the principles are more general, others more specific. Two of the more general principles are: (A) Minimize Learning Effort and (B) Minimize Processing Effort.

(A) Minimize Learning Effort (Mi. L) Learners of a second language (L 2) prefer to minimize learning effort when they learn the grammatical and lexical properties of the L 2. Learning effort is minimized when shared properties of the L 1 can be transferred directly into the L 2 (Ma. PT), when properties of the L 2 are frequently occurring in the L 2 input (Ma. F), when structural and semantic properties of the L 2 are simple rather than complex (Ma. S), and when there are fewer linguistic items to be learned in a given grammatical or lexical domain.

(B) Minimize Processing Effort (Mi. P) Learners of a second language (L 2) prefer to minimize processing effort when they use the grammatical and lexical properties of the L 2, just as native speakers do. E. g. even when more complex properties have been learned at an acquisition stage, L 2 learners will still prefer to use simpler properties, just like native speakers do.

Principles (A) and (B) are principles of least effort. If these were the only principles determining learning and production our learner corpora would reveal increasingly minimal outputs. Clearly, they do not. MLUs (i. e. mean length of utterance figures) increase at each higher proficiency level (cf. Hawkins & Filipović 2012: ch. 2. 2) as greater use is made of less frequent and more complex structures and meanings: A 2 B 1 B 2 C 1 C 2 7. 9 10. 8 14. 2 17. 3 19. 0

The reason is that learners are trying to increase their expressive power in the L 2, and to behave like native speakers, which means learning and using the mix of infrequent and frequent, and complex and simple, linguistic items, just like native speakers do.

(C) Maximize Expressive Power (Ma. E) Learners of a second language (L 2) prefer to maximize their expressive power, i. e. to formulate in the L 2 whatever thoughts they would wish to express in the L 1, and to perform the same language functions as L 1 users. Successive stages of acquisition reveal more native-like L 2 outputs with increasingly complex and less frequent structures for the expression of increasingly complex thoughts, in partial opposition to principles (A) Mi. L and (B) Mi. P.

(D) Maximize Communicative Efficiency (Ma. C) Learners of a second language (L 2) prefer to maximize their communicative efficiency in relation to the hearer and his/her mental model. Communication is efficient when the message (M) intended by the speaker (S) is calibrated to the hearer's (H) mental model in such a way as to achieve accurate comprehension of M with rapid speed. This requires sometimes more, sometimes less, processing effort, in partial opposition to principle (B) Mi. P. Principle (D) can help us explain some interesting differences between ‘positive’ and ‘negative’ transfers, i. e. transfers from the L 1 that are correct rather than error-ful.

The three more specific principles introduced earlier follow from these more general principles: (1) Maximize Frequently Occurring Properties (Ma. F) Properties of the L 2 are learned in proportion to their frequency of occurrence: more frequent exposure of a property to the learner facilitates its learning and reduces learning effort. (2) Maximize Structurally and Semantically Simple Properties (Ma. S) Properties of the L 2 are learned in proportion to their structural and semantic simplicity: simplicity means there are fewer properties to be learned and less learning effort is required. (3) Maximize Positive Transfer (Ma. PT) Properties of the L 1 which are also present in the L 2 are learned more easily and with less learning effort, and are readily transferred, on account of pre-existing knowledge in L 1.

Let’s now consider a fourth specific principle, involving negative transfers from the L 1. CASP can explain some interesting puzzles here in the SLA literature: (4) Permit Negative Transfer (PNT) Properties of the L 1 which are not present in the L 2 can be transferred, resulting in errors, as learners strive to achieve an expressive power and communicative efficiency in L 2 comparable to that in their L 1 (see principles C and D), while minimizing learning effort (principle A) and/or processing effort (principle B). I. e. when grammatical and lexical properties are shared, transfers from L 1 into L 2 result in positive or correct properties in the L 2. When properties are not shared, and the transfer still takes place, this results in negative or incorrect properties in the L 2.

One major difference between positive and negative transfers is that there are severe limitations on expressive power and on communicative efficiency that can be conveyed by linguistic properties when they are not part of the L 2 and are not used by its native speakers. When native speakers communicate with L 2 learners they tolerate and compensate for departures from the native language conventions. But when learners depart too radically from these conventions, they are not understood by native speakers. Learners accordingly acquire a sensitivity to the native speaker’s ability to compensate for these violations in conventions of grammar and use. This, we believe, plays a major role in determining whether and when negative transfer can occur and when errors will be found, see principle (5):

(5) Communicative Blocking of Negative Transfer (CBN) The transfer of negative properties from L 1 to L 2 is filtered in proportion to communicative efficiency (principle D): the more an L 1 property impedes efficient communication in L 2, the less negative transfer there is.

In phonology substitution of L 1 consonants like [t] or [s] or [f] for L 2 [θ] in English thin minimizes learning and processing effort for learners whose L 1 s do not have this consonant, and these substitutions generally succede communicatively (Lado 1957). In syntax Spanish Pro-Drop (e. g. *is a beautiful country for it is a beautiful country) is often transferred into early L 2 English to express the proposition in question and the removal of the subject it does not impede communicative success. This structure is simpler than its English counterpart with an overt subject, and transfer is not blocked, as predicted by our principle (3) Ma. S. Similarly, many article omission errors do not diminish expressive power and communicative success, and at the same time they minimize learning and processing effort through the transfer of L 1 structures, see above.

By contrast, Chinese prenominal relative clauses do not result in errors whereby the English man whom the woman loves is changed into its Chinese prenominal counterpart *the woman loves whom man This Chinese structure is complex and typologically marked cross-linguistically (Hawkins 2004, Eckman 2011), and if Chinese learners used it in L 2 English, they wouldn’t be understood. More generally, many structures and meanings will often not transfer from L 1 to an L 2 even when they are shared, and will be blocked by principles (2) Ma. F, (3) Ma. S and (5) CBN. I. e. our interacting CASP learning principles make predictions for when an L 1 feature will transfer, and when it will not.

Consider the basic word orders of English and Japanese. These languages have mirror-image patterns, head-initial versus head-final, that are both frequent and productive across languages: [went [to [the cinema]]] versus [[[the cinema] to] went], (Greenberg 1966; Dryer 1992; Hawkins 1983, 1994, 2004)

Head-final Japanese orders are not transferred into L 2 English by Japanese learners, just as head-final Chinese noun phrases are not transferred into L 2 English. Why not? Because that would result in extreme communicative inefficiency: speakers using Japanese or Chinese word orders in English L 2 would simply not be understood! By contrast, headinitial word order variants of Spanish that lack precise counterparts in English (e. g. , I read yesterday the book) are negatively transferred into L 2 English, and they do not impact efficient communication.

We predict that because Japanese is a head-final language, the contrast with the mirror-image word order patterns of English is considerable and transferring head-final patterns into a headinitial language like English, and vice versa, would significantly impair communication. This is why it is imperative for Japanese learners of English, and English learners of Japanese, to acquire correct basic word orders in their L 2 s early. But speakers of L 1 languages with flexible SVO like Spanish do not have the same incentive, because even when they transfer incorrect orders from their L 1 into a fundamentally similar headinitial English L 2, communication is not significantly impaired. The CLC error data for Japanese and Spanish learners reflect this difference perfectly: word order errors are rare or non-occurring for the former, but common for the latter.

(6) Order of Second Language Acquisition (OSLA) The order of acquisition for properties of the L 2 is in accordance with general principles (A)-(D), and with the more specific principles and patterns that are supported empirically. These principles can be incorporated within a multi-factor model of SLA, the CASP model, and used to define possible versus impossible, and likely versus unlikely, interlanguage stages proceeding from a given L 1 to a given L 2. These principles operate collectively to make constrained predictions for the acquisition of properties of L 2 English and of other languages, and for their relative sequencing. Their interaction is complex, because there are several such principles, which sometimes compete and sometimes cooperate, because they are gradient, and because they have different relative strengths.

Some Practical Applications of this Research Once criterial features of the different proficiency levels and the interacting principles of the CASP model have been defined, they can be put to use for learning, teaching and assessment purposes.

For Learning and Teaching NB! Our criterial features are taken only from candidates who scored passing grades at each level. Hence learners who are studying for the relevant level can now be told explicitly what their successful peers have mastered, and teachers can incorporate these features in their teaching. Grammatical and lexical properties of English, and teaching materials and methods built around them, can be calibrated to the criterial features of each level, making learning more efficient.

Learners striving for B 1 can be introduced to Object Control structures that are first attested at B 1 like I ordered him [to gather my men to the hall] B 1 and to subordinate clauses with WH-movement: I don’t know [how I could have done it] I did not know [where to look for it] B 1

They can be introduced to the lexical verbs that successful candidates in B 1 exams know, e. g. divide, fit, grab, spill, stick, tear and to the expanding meanings of verbs learned earlier: e. g. break appears for the first time in the extended sense of INTERRUPT at B 1: At last I managed to break the routine of the city … B 1

Similarly for the other levels and their features. More generally, grammar and vocabulary can be introduced in a sequence that reflects their frequency in the input and their inherent simplicity/complexity, as revealed both through the CLC and through native speaking corpora like the BNC.

For Assessment This research provides content that can help to validate the scores that examiners of English have provided independently. The assignment of a level and a grade to a sample of learner English currently relies on judgments that examiners make based on their experience and training. Examiners have learned to assign scores with good inter-examiner agreement, but there is still a certain amount of intuition that they bring to the task. Examiners are implicitly rather than explicitly aware of what to look for in many cases.

An individual script, let us abbreviate it as S, by a candidate taking an exam at level X can be searched for the presence versus absence of criterial features derived ultimately from all passing scripts at X, and from those at the immediately lower level X-1 or at the immediately higher level X+1.

Script S may contain several constructions and lexical items that are features of B 2 and higher levels. This establishes that S is at least B 2. The script might contain no uniquely C-level features, however. These levels are eliminated, therefore, and B 2 is supported. S may even contain a unique B 2 feature. This all supports B 2.

Criterial features can also be used in the preparation of diagnostic grammar tests that assign students to their appropriate levels of instruction based on their command of English grammar.

L 1 -specific ELT materials can be written for different groups of learners. For learners whose first languages have no definite and indefinite articles, English language materials can encourage explicit and implicit learning in this area.

The learning stages, transfer effects and error types characteristic of Spanish learners of English can be reflected in textbooks and teaching materials designed specifically for them. Similarly for Chinese learners, and Japanese learners, and Russians, etc.

Theoretical interest of this work The criterial features we are extracting from the corpus are of direct benefit for theoreticians studying language acquisition. They provide a new set of empirical patterns that can inform predictive and multi-factor theories of learning such as CASP and they enable us to test and refine principles of frequency, complexity and transfer, as illustrated in this talk. (see Hawkins & Filipović 2012 and Filipović & Hawkins 2013 for further details)

References Briscoe, E. , J. Carroll and R. Watson (2006) ‘The second release of the RASP system’. In Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions, Sydney, Australia. Council of Europe (2001) Common European Framework of Reference for Languages: Learning, teaching, assessment. CUP, Cambridge. Diessel, H. (2004) The Acquisition of Complex Sentences. CUP, Cambridge. Dryer, M. S. (1992) ‘The Greenbergian word order correlations’, Language 68: 805 -55. Eckman, F. R. (1984) ‘Universals, typologies, and interlanguage’. In: W. E. Rutherford, ed. , Language Universals and Second Language Acquisition, John Benjamins, Amsterdam, 79 -105. Eckman, F. R. (2011) ‘Linguistic typology and second language acquisition’. In: J. J. Song, ed. , The Oxford Handbook of Linguistic Typology, OUP Oxford, 618 -633. Filipović, L & J. A. Hawkins (2013) ‘Multiple factors in second language acquisition: The CASP model’, Linguistics 51(1): 145 -176. Gell-Mann, M. (1992) ‘Complexity and complex adaptive systems’. In J. A. Hawkins & M. Gell-Mann, eds. , The Evolution of Human Languages, Addison-Wesley, Redwood City, CA. Greenberg, J. H. (1963) ‘Some universals of grammar with particular reference to the order of meaningful elements’, in J. H. Greenberg (ed. ) Universals of Language, MIT Press, Cambridge, Mass. . , 73 -113.

Hawkins, J. A. (1983) Word Order Universals. Academic Press, New York. Hawkins, J. A. (1994) A Performance Theory of Order and Constituency. CUP, Cambridge. Hawkins, J. A. (2004) Efficiency and Complexity in Grammars. OUP, Oxford. Hawkins, J. A. & P. Buttery (2009) ‘Using learner language from corpora to profile levels of proficiency: Insights from the English Profile Programme’. In L. Taylor & C. J. Weir, eds. , Language Testing Matters, Proceedings of the 3 rd ALTE Conference 2008, CUP, Cambridge, 158 -175. Hawkins, J. A. & P. Buttery (2010) ‘Criterial features in learner corpora: Theory and illustrations’, English Profile Journal 1. Hawkins, J. A. & L. Filipović (2012) Criterial Features in L 2 English: Specifying the Reference Levels of the Common European Framework. CUP, Cambridge. Lado, Robert 1957. Linguistics across cultures: Applied linguistics for language teachers. Ann Arbor, MI: University of Michigan. Mac. Whinney, B. (2005) ‘A unified model of language acquisition’. In J. F. Kroll & A. M. B. de Groot, eds. , Handbook of Bilingualism: Psycholinguistic Approaches, OUP, Oxford. Tomasello, M. (2003) Constructing a Language: A Usage-based Theory of Language Acquisition, Harvard University Press, Cambridge, Mass. Williams, C. A. M. (2007) ‘A preliminary study into verbal subcategorisation frame usage in the CLC’, MS, RCEAL, University of Cambridge.

Acknowledgments The findings reported here are based on joint work with Luna Filipović (Cambridge Linguistics Department, now University of East Anglia), see Hawkins & Filipović 2012 and Filipović & Hawkins 2013. The assistance of many other researchers and collaborators was vital. Special thanks to: Ted Briscoe of the Cambridge Computer Lab and his colleagues for use of the RASP parser; Paula Buttery of DTAL, Cambridge (see Hawkins & Buttery 2009, 2010); Dora Alexopoulou of the Education First Research Unit, DTAL, for advice and input; Andrew Caines of DTAL, Cambridge, for help with numerous syntactic searches; Annette Capel of CUP for help with the wordlist searches; Kristen Kennedy of UC Davis for help with MLU calculations; Mike Milanovic & Nick Saville of Cambridge ESOL for theoretical and practical guidance and financial support (see below); Mike Mc. Carthy of CUP and Penn State U for advice and input; Roger Hawkey and Angeliki Salamoura of Cambridge ESOL for advice and English Profile Programme co-ordination; Lu Gram of the Computer Lab for help with error calculations and other searches; Caroline Williams of RCEAL, Cambridge for verb subcategorization data; and to CUP’s computational linguists who prepared "The <#S> Compleat|Complete Learner Corpus Document" 2006, from which the error codes and examples sentences are taken.

Financial Support The work reported here was made possible by generous financial support from Cambridge Assessment and from Cambridge University Press, within the context of the Cambridge English Profile Programme, and by research funds and a seed grant for international outreach from the University of California, Davis. This support is gratefully acknowledged.

Links to Learner Corpora for English and Other Languages The EF (Education First) Cambridge Open Language Database for English (30+ million words): http: //corpus. mml. cam. ac. uk/efcamdat/ Sylviane Granger’s summary of learner corpora for English and other languages worldwide: https: //www. uclouvain. be/en-cecl-lcworld. html