A grammatical account of lexical exceptions in phonological acquisition 1

USE LISTED ERROR: A grammatical account of lexical exceptions in phonological acquisition 1 Anne-Michelle Tessier Department of Linguistics, University of Alberta Two questions about learning a phonological grammar Q1. How do children learn the specifics of their target language s phonology? In OT terms, how do they manage to learn the right amount of markedness? And how and why do they learn it gradually? Q2. How can grammatical theories understand lexical exceptions in development? What I will claim today: A1. To gradually learn all and only the right structures of a target grammar, a learner must rely heavily on stored errors A2. Those stored errors used to build the grammar may also help to provide a formal understanding of developmental lexical exceptions I call this approach Error-Selective Learning (Tessier 2006, 2007) Error-Selective Learning requires a lot of error memorization Can we find any independent support, from elsewhere in phonological development, that children are storing and using their previous errors? II. Data: Two kinds of lexical exceptions in development Fossilized forms: lexical items that seem stuck at an earlier stage of development Precocious forms. lexical items that seem to foreshadow a later stage Intuitive idea about these forms: they are somehow outside the current grammar Foreshadowing today s proposal Lexical exceptions do escape the core grammar... meaning the current ranking of Markedness and Faithfulness constraints However! their exceptionality is encoded in the grammar and it is created, and eventually resolved, via normal error-driven learning Overview of Today s Talk I. Background: gradual phonological learning and the importance of errors 1) A common trajectory of English coda acquisition a) initial state: /tost/ to (an error) b) intermediate state: /tost/ tos (an improved error) c) final state: /tost/ tost (the target form!) Two views of how to learn a phonology gradually i) accumulate a grammar - use and forget errors as you go ii) accumulate errors build and forget grammars as you go This talk assumes the latter idea: the centrality of errors learners make errors, learners are selective about which errors they memorize, and learning means reasoning from sets of accumulated errors to each new grammar 1 For discussion, help and suggestions that have accumulated over many years in various parts of this work, I thank Michael Becker, Christopher Cox, Karen Jesney, John McCarthy, Joe Pater, Yvan Rose, Shelley Velleman and Kie Zuraw, as well as three anonymous NELS reviewers. All remaining errors are sadly mine. III. The connection between stored errors and lexical exceptions A learner that stores its errors has two ways to produce a word: i) if it has made an error on the word before, it can simply reproduce that error ii) if it has no stored error for this word, it will have to use its current grammar I will propose a constraint that encodes a laziness preference for the former option: 2) USELISTEDERROR (to be defined below; cf. Zuraw 2000 s USELISTED) Today s claim: Used together, a set of stored errors, the rankable constraint USELISTEDERROR, and Error-Selective Learning can provide an account of lexical exceptions in phonological development This talk will walk through this claim, and begin to discuss its implications. 1

I. An idea about learning phonology, restricively and gradually 1.1. A view of the learning challenge 3) What s in an error? observed, target English output: toast! [tost] (says mother) current grammar s output: to! [to] (says child) (a) an error tableau: /tost/ NOCODA *COMPLEXCODA *PHARYNGEAL MAX-C tost *! *! to * (b) analyzing the error 2 winner ~ loser NOCODA *COMPLEXCODA *PHARYNGEAL MAX-C tost ~ to L L e W What should the learner learn from 3b)? Eventually, English learners learn that both simple and complex codas are tolerated by the grammar, but that e.g. pharyngeals (and many other things) are not In the meantime, though, many English learners stop off for some months at a stage where they produce only singletons despite evidence from words like toast Upshot: they must be learning restrictively, and also gradually 3 2 This comparison format is taken originally from Prince (2002), where such objects are called Elementary Ranking Condition vectors. 3 The familiar OT method of imposing restrictive learning are ranking biases: famously Markedness >> Faithfulness (e.g. Smolensky 1996; Pater 1997; Demuth 1996; Gnanadesikan 1995/2004) However, it turns out that (i) this bias is not the only one necessary (see various arguments in e.g. Smith 2000; Hayes 2004; McCarthy 1998; Tessier 2007) and (ii) these biases must enforced persistently i.e. at each step of re-ranking (Broiher 1995, Prince and Tesar 2004 Hayes 2004; cf. Ito and Mester 1999.) The class of algorithms that I use, from Prince and Tesar 2004, can incorporate as many biases as necessary, and impose them persistently throughout learning. See Prince and Tesar 2004; Hayes 2004;Tessier 2007. 1.2 Error-Selective Learning: restrictive AND gradual (Tessier 2007) The key: two kinds of stored errors temporary ones, made by the current grammar: the Cache permanent ones, that drive the move to each new grammar: the Support The Set-up Begin with a very restrictive grammar (including e.g. All Markedness >> All Faith) Day-to-day, use your current grammar and make errors (like 3b) Store those errors in the Cache... but continue using the same grammar Eventually: after you have accumulated enough evidence that some markedness constraint is being too restrictive a learning cycle occurs 4) The Error-Selective Learning Cycle a) Choose from Cache an error(s) that will minimally change your current grammar b) Add just that error to the Support, and empty the Cache c) Use a re-ranking algorithm to build a new grammar that resolves all the errors, including the new one, in the Support (here, using a version of Biased Constraint Demotion algorithms(prince and Tesar 2004; see also Hayes 2004.) Learning is all about making, storing, analyzing errors re-ranking is the easy bit II. Fossilized and Precocious Forms in developing phonologies Well-documented data is a bit thin on the ground: lots of anecdotal evidence, but little systematic study (for good reasons). My goal here: to work out a theory about two types of lexical exceptions in development. Typological study and more cross-linguistic data are necessary future projects 4 4 One empirical note about the scope of the data: Many case studies report various regressions in the time-course of development (e.g. Bleile and Tomblin, 1991; Jespersen, 1922; Macken 1980; Macken and Ferguson 1979; Menn, 1976, 1983; Stemberger et al 2001.) Many (most?) regressions seem to have reported lexical exceptions. In today s talk, these exceptions to regressions will end up being precocious forms. 2

2.1 Fossilized forms An early stage of the learner s grammar did not permit a target structure Now, the learner has acquired that structure... except in some word(s) Schematically: 5)a) Stage n: NOCODA >> MAX toast /tost/ [to] dog /dɑg/ [dɑ] bee /bi/ [bi] b) Stage n+1: MAX >> NOCODA toast /tost/ [tos] dog /dɑg/ [dɑ] fossilized form bee /bi/ [bi] Example: Trevor (Compton and Streeter 1977) s pronunciation of his name 6) Usually /tr/ clusters retained except in Trevor, where /tr/ [tʃ] truck [tɹʌk] (2;2.23) [tʃɛvə:] (2;3.04) train [tɹe:n] (2;2.23) [tʃɛvə:] (2;4.03) try [tɹai] (2;4.03) [tʃɛvə:] (2;4.13) tree [tɹi:] (2;4.13) 2.2 Precocious forms The child grammar doesn t allow a structure permitted in the target... except some word(s) DO allow it! Schematically: 7) Stage n: NOCODA >> MAX toast /tost/ [to] dog /dɑg/ [dɑg] precocious form bee /bi/ [bi] Today s example of precociousness: Velar Fronting (VF) velar segments (k, g) banned in particular contexts, particular word-initially usually fronted to coronal place (t, d) 8) Velar Fronting in E s grammar (Inkelas and Rose 2007) kiss /kɪs/ [tɪs] again /əgɪn/ [ədɪn] VF is a common process among English-learning children and Tweety Birds 9) I thought I saw a pussy cat! [aɪ tɑt aɪ tɑ ə pʊɾi tæt] (On non-fictional VF in child phonology, see Bills & Golston 2001, Chiat 1983, Brett, Chiat & Pilcher 1987, Dinnsen 2002, Ingram 1974; Inkelas and Rose 2007; Stoel- Gammon & Stemberger 1994, Stoel-Gammon 1996) Inkelas and Rose (2007) s approach, whose spirit I adopt: VF is child-specific, driven by child s articulatory pressures I will encode such pressures in an OT constraint, simply called *K... the real constraint must be positionally-sensitive 5 Precocious forms in a VF grammar (Bleile and Tomblin 1991) 10) Stage n: Word-initial, stressed velars generally fronted... but, an exceptional set were not fronted 11) Regular VF Precocious Words 6 candy /kændi/ [tændi] clown /ˈklaʊn/ [kaʊn] okay /oˈke/ [oˈke] cookie /ˈkʊki/ [ˈkʊki] Is this really grammar? Yes! (What does that mean?) 5 Inkelas and Rose show that E applied velar fronting to initial and stressed syllables; the Bleile and Tomblin data is compatible with that characterization as well. 6 The full set of precocious words reported: clown, kitty cat, cow, cookie, (ice) cream, Gumbi, but the three listed in (11) are those that surpassed the numerical criteria set prior to the study. 3

III. Deriving lexical exceptions in error-selective learning 3.1 A proposed constraint: USELISTEDERROR 12) USELISTEDERROR Assign a violation to any output form that is non-identical to the input s stored loser form These loser forms include both errors in the current Cache AND the Support USELISTED is a formal pressure encoding a kind of conservatism/laziness starts at the top of the ranking, and is demoted via errors in the Support Where we re going adding this constraint means there are now three possible stages in which a fullyunmarked grammar (M >> F) becomes fully-faithful (F >> M) this new intermediate stage can derive fossilization of a particular error 3.2 Fossilized forms: the simple effect of USELISTEDERROR 13) Initial state: USELISTEDERROR >> MARKEDNESS >> FAITH e.g. USELISTEDERROR >> NOCODA >> MAX 14) First time processing dog : no stored error, no effect of USELISTED /dag/ USELISTEDERROR NOCODA MAX no stored loser a) dag *! b) da * This error gets added to the Cache, as usual, and similar errors pile up on new words. What happens next time the learner tries again to say dog? Now the Cache does have a stored loser for this word, so: 15) Next time; because of stored (14), USELISTED is violated! /dag/ UseListedError NoCoda Max stored loser: [da] a) dag *! *! b) da * 16) A fragment of the Cache, near the end of Stage One... winner ~ loser USELISTEDERROR NOCODA MAX a) dag ~ da e L W b) dag ~ da stored loser: da L L W 3.2.1 Moving to the next stage In Error-Selective learning, once learning is triggered: The learner must pick an error from the Cache to add to the Support and learn from The Error-Selective learner chooses an error that will make minimal changes to the current grammar. Depending on what the rest of the Cache contains, either error could be chosen. 7 What if the chosen error does NOT violate USELISTEDERROR? 17)a) T he error in 16a), added to the Support winner ~ loser USELISTEDERROR NOCODA MAX dag ~ da e L W b) The new grammar that the BCD algorithm builds: USELISTEDERROR >> MAX >> NOCODA This Stage 2 grammar has fossilized the dag ~ da error When the Stage 2 grammar processes any other word: 18) MAX >> NOCODA usually preserves codas... /kaet/ USELISTEDERROR MAX NOCODA no stored loser a) kaet * b) kae *! But to reach Stage 2, the error in 16a) has been added to the Support: so there IS a stored loser for dog! and due to high-ranking USELISTEDERROR, the grammar has to use it 19) USELISTEDERROR >> MAX retains the old error /dag/ USELISTEDERROR MAX NOCODA stored loser [da] a) dag *! * b) da * 7 Curious or suspicious souls who are wondering about the formal details of how minimal change is defined, and about how learning is triggered, are welcome to ask (or seetessier 2007 ch. 3.) 4

Upshot: if the learner happens to learn to produce codas via an error like (16a), rather than (16b), they will end up with a fossilized coda-less form 20) Stage 1: All codas deleted USELISTEDERROR >> NOCODA >> MAX Stage 2: Most codas preserved, except fossilized error USELISTEDERROR >> MAX >> NOCODA 3.3 Precocious forms: USELISTEDERROR plus child-specific constraints Recall the case of precocious forms escaping velar fronting via *K 21) Regular words Precocious words candy /kændi/ [tændi] clown /ˈklaʊn/ [kaʊn] okay /oˈke/ [oˈke] Where we are going: why is this set of words is exceptional? hypothesis: connected to the fact that *K is a child-constructed constraint adding errors to the support before constructing the *K constraint can create precocious forms later on 22) Stage 1: No *K constraint constructed yet 23) The Support at this stage (USELISTEDERROR not yet relevant) winner ~ loser *DIPHTHONG *COMPLEX ONSET MAX IDENT- PLACE cow kaʊ ~ ka L e W e clown klaʊn ~ kaʊn e L W e At some point, the learner is overcome by increasing articulatory demands... and so decides to lessen the burden by adding a *K constraint to their grammar Necessary assumption: when a new markedness constraint is built, it is ranked as in the initial state thus, it appears above all Faith constraints... and below USELISTEDERROR 24) Stage 2: USELISTEDERROR >> *K >> IDENT-PLACE This grammar is now one that applies velar fronting to new words: 25) New kind of error being created by this grammar: VF /kændi/ USELISTEDERROR *K IDENT-PLACE a) kændi (*) *! b) tændi *... but not to those old errors in the Support of (23)! 26) Errors in the Support created before *K do not undergo VF! / klaʊn / USELISTEDERROR *K IDENT-PLACE stored loser: [kaʊn] a) klaʊn *! * b) kaʊn * c) taʊn *! * At Stage 2, we have a set of precocious forms with unfronted velars. In the previous section, we saw how USELISTEDERROR retains an old pronunciation But when a new constraint has been added to the grammar, USELISTEDERROR can actually allow old structure, which has only recently been deemed marked, to survive 3.4 Moving on 3.3.1 Moving beyond fossilization Recall the fossilized stage: USELISTEDERROR >> *K >> IDENT-PLACE To finish learning: an error on the fossilized word (here, dog ) needs to get into the Support 8 27) The error that gets the learner to the final stage winner ~ loser USELISTEDERROR NOCODA MAX dag ~ da stored error: da L L W 28) The final grammar, that the learning algorithm builds from 27) MAX >> USELISTEDERROR, NOCODA 8 A technical point about how errors that overcome fossilization get into the Support. Given the implementation of ESL in Tessier (2007), this will require one of the following three tweaks: getting the Violation Threshold down to one (which seems unlikely), or counting error tokens (which might work), or treating errors driven by UseListed as special (which might be best.) 5

29) The final grammar /dag/ MAX USELISTEDERROR NOCODA stored loser: [da] a) dag * * b) da *! 30) Stage 1: USELISTEDERROR >> NOCODA >> MAX Stage 2: USELISTEDERROR >> MAX >> NOCODA (fossilization) Stage 3: MAX >> USELISTEDERROR, NOCODA 9 3.3.2 Moving beyond precocious forms Later stages that include precocious forms seem to be more complicated, compared to fossilizations. As an example: for E, the velar fronting child, the next stage was one in which fronting was uniformly applied, even in the previously-precocious forms. 31) Amount of Velar Fronting among Precocious Words Sessions # VF tokens # Unfronted tokens Stage 2 1-6 3 16 Stage 3 7-22 58 3 Stage 4:... two weeks after the study, a post-test found no more VF at all. Discussion of how these later stages ought to be approached is included in an appendix follow-up questions welcome. 3.5 Section Summary the learner s biased, initial state: USELISTEDERROR >> M >> F if a new M constraint is introduced into this ranking (like *K) we get precocious forms, from errors made before *K was created USELISTEDERROR >> *K >> F if the learner learns an intermediate grammar: USELISTEDERROR >> F >> M we get fossilized forms, from errors made at the initial state 9 Note: if the learner had added the other error, 16b, to the Support at Stage 1, they would have not fossilized any errors and instead acquired all codas at the same time, moving to Stage 3. 3.6 Discussion of this approach to lexical exceptions 3.6.1 Important aspects of the learner for this approach error storage: otherwise USELISTEDERROr cannot make sense two error-storage facilities: both Cache and Support the ability to construct constraints, during the process of phonological learning 3.6.2 The Inspiration for the USELISTEDERROR constraint Zuraw (2000): USE-LISTED while this was my inspiration, it is completely different! this constraint means use the listed winner it begins at the bottom of Zuraw s grammar; rises to the top during learning prevents low-ranked stochastic generalizations from affecting novel words, so they will only influence nonce words My proposed constraint, USELISTEDERROR, does the reverse: it restricts just the known forms (known in the sense of having a stored error)...... while nonce words that have no stored entry are taken at current face value IV. Open and Unresolved Questions The main question: how can we look for empirical support for this account? With respect to precocious forms: processes that are considered child-specific, and so require extra constraints to be constructed by the learner, should show these lexical exceptions the easiest place to look: child Consonant Harmony 32) Examples of Trevor s Consonant Harmony (from Pater and Werle 2001) dog /dɑg/ [gɑg] cup /kʌp/ [kʌk] coat /kot/ [kok] cat /kæt/ [kæk] Anecdotally, it is reported that children will start out with isolated words with two places of articulation, and then later regress those forms to conform with CH: Example: Moskovitz (1980) reports earliest pronunciations of e.g. truck that did not undergo CH (at 8 months), but were regressed to [kak] with CH at a later stage. The question now: when and how often do these early words resist CH? (for a while?) 6

4. Conclusions (Some) lexical exceptions in phonological development can be modeled as part of normal grammatical learning The Proposal A learner that stores, accumulates, and selectively learns from its errors is wellsuited to this task -- if provided with a formal pressure to be lazy in recycling errors: USELISTEDERROR The Implications These analyses provide new uses of stored errors, as required in an approach like Error-Selective Learning beyond the need for restrictiveness As in many areas, Building a formal view of lexical exceptions in phonological development may provide a clearer lens through which to examine the range, scope, and implications of these frequent, but perhaps under-studied, phenomena Thank you very much. References Bernhardt, Barbara. H., and Joseph. P. Stemberger (1998) Handbook of Phonological Development: From the Perspective of constraint-based nonlinear phonology, San Diego, Academic Press. Bills, S. & C. Golston. 2001. Prosodic and Linear Licensing in English Acquisition. To appear in Proceedings of the Western Conference on Linguistics. Bleile, Ken and J. Bruce Tomblin. 1991. Regressions in the Phonological Development of Two Children. Journal of Psycholinguistic Research 20 no. 6, 483-499. Brett, L., S. Chiat & C. Pilcher. 1987. Stages and Units in Output Processing: Some Evidence from Voicing and Fronting Processes in Children. Language and Cognitive Processes 2:165-177. Broiher, Kevin. 1995. Optimality Theoretic Rankings with Tied Constraints: Slavic Relatives, Resumptive Pronouns and Learnability. Ms, MIT. ROA-46. Chiat, S. 1983. Why Mikey s Right and My Key s Wrong: The Significance of Stress and Word Boundaries in a Child s Output System. Cognition 14:275-300. Compton, A.J. and M. Streeter. 1977. Child Phonology: Data Collection and Preliminary Analyses. In Papers and Reports on Child Language Development 7. Stanford University, Stanford, California. (See also Pater 1997 on this corpus.) Demuth, Katherine, & E. Jane Fee. 1995. Minimal Words in Early Phonological Development. Ms., Brown University and Dalhousie University. Demuth, Katherine (1996). Stages in the acquisition of prosodic structure. In E. Clark (ed.), Proceedings of the 27th Child Language Research Forum, pp. 39-48. Stanford University: CSLI. Dinnsen, Daniel. 2002. On the Composition and Treatment of Children s Phonological Error Patterns. Paper delivered at the GLOW Workshop on Language Acquisition. Utrecht Institute of Linguistics OTS, Utrecht University. April. Gierut, Judith and Daniel Dinnsen. 1987. On predicting ease of phonological learning. Applied Linguistics 8: 241-263. Gnanadesikan, Amahlia. 1995/2004. Markedness and Faithfulness Constraints in Child Phonology. In Kager, Rene, Joe Pater & Wim Zonneveld,(eds.), Fixing Priorities: Constraints in Phonological Acquisition. Cambridge, U.K.: Cambridge University Press. Hayes, Bruce. 2004. Phonological Acquisition in Optimality Theory: the early stages. In Kager, Rene, Joe Pater & Wim Zonneveld,(eds.), Fixing Priorities: Constraints in Phonological Acquisition. Cambridge, U.K.: Cambridge University Press. Ingram, David. 1974. Fronting in child phonology. Journal of Child Language 1: 233-241. Ito, Junko and Armin Mester. 1999. The structure of the Phonological Lexicon. In N. Tsujimura, (ed.), The Handbook of Japanese Linguistics. Oxford: Blackwell. pp. 62-100. Jespersen, Otto. 1922. Language: Its nature, development and origin. New York: H. Holt & Co. 7

Kiparsky, Paul and Lise Menn. 1977. On the acquisition of phonology. In John Macnamara (ed.), Language Learning and Thought. New York: Academic Press, pp. 47-78. Levelt, C. C. and R. van de Vijver. 2004. Syllable types in cross-linguistic and developmental grammars. In R. Kager, W. Zonneveld, J. Pater (eds.) Fixing Priorities: Constraints in Phonological Acquisition, Cambridge, Cambridge University Press. Macken, Marlys. 1980. The child s lexical representation: The puzzle-puddle-pickle evidence. Journal of Linguistics 16, 1-17. Macken, Marlys and Charles Ferguson. 1983. Cognitive aspects of phonological development: Model, evidence and issues. In K. Nelson (ed.) Children s Language. vol 3. Hillsdale, NJ: Erlbaum. Menn, Lise. 1976. Pattern, control, and contrast in beginning speech: A case study in the development of word form and word function. Ph.D. dissertation, University of Illinois, Urbana. Menn, Lise. 1983. Development of articulatory, phonetic, and phonological categories. In B. Butterworth (ed.) Language production. New York: Academic Press. McCarthy, John J. 1998. Morpheme structure constraints and paradigm occultation, In M. Catherine Gruber, Derrick Higgins, Kenneth Olson, and Tamra Wysocki, eds., Proceedings of the Chicago Linguistic Society 5, Vol. II: The Panels, Chicago, CLS. Moskovitz, B. A. 1980. Idioms in phonology acquisition and phonological change. Journal of Phonetics 8, 69-83. Pater, Joe. 1997. Minimal Violation and Phonological Development. Language Acquisition 6(3): 201-253. Pater, Joe and Adam Werle. 2001. Typology and variation in child consonant harmony. In C. Féry, A. D. Green and R. van de Vijver, eds., Proceedings of HILP 5, University of Potsdam. Prince, Alan 2002. Entailed Ranking Arguments.. Manuscript, Rutgers University. ROA-500. Prince, Alan and Bruce Tesar. 2004. Learning Phonotactic Distributions. In R. Kager, Joe Pater & Wim Zonneveld,(eds.), Fixing Priorities: Constraints in Phonological Acquisition. Cambridge, U.K.: Cambridge University Press. Smolensky, Paul. 1996. On the comprehension/production dilemma in child language. Linguistic Inquiry 27: 720-731. Smith, Jennifer L. 2000. Positional faithfulness and learnability in Optimality Theory, In Rebecca Daly and A. Rehl, eds., Proceedings of ESCOL99, Ithaca, CLC Publications. Stoel-Gammon, Carol. 1996. On the Acquisition of Velars in English. In B. Bernhardt et al. (eds.) Proceedings of the UBC International Conference on Phonological Acquisition. Somerville: Cascadilla Press. 201-214. Stoel-Gammon, Carol. & Joseph Stemberger. 1994. Consonant Harmony and Underspecification in Child Phonology. In M. Yavas (ed.) First and Second Language Phonology. San Diego: Singular Publishing Group, Inc. 63-80. Tesar, Bruce. 2000. Using inconsistency detection to overcome structural ambiguity in language learning. Technical Report RuCCS-TR-58, Rutgers Center for Cognitive Science, Rutgers University. ROA-426. Tesar, Bruce and Paul Smolensky. 2000. Learnability in Optimality Theory. Cambridge, MA: MIT Press. Tessier, Anne-Michelle. 2006. Stages of Phonological Acquisition and Error-Selective Learning. In Donald Baumer, David Montero, and Michael Scanlon (eds.), Proceedings of WCCFL25. Somerville, MA: Cascadilla Press. Tessier, Anne-Michelle. 2007. Biases and Stages in Phonological Acquisition. Ph.D. dissertation, University of Massachusetts Amherst. ROA #883-1106. Zuraw, Kie. 2000. Patterned Exceptions in Phonology. Ph.D. dissertation, UCLA. Available at [http://www.linguistics.ucla.edu/people/zuraw/dnldpprs/diss.pdf] 8

Appendix to Section 3.3: Later stages beyond precocious forms The relevant example: for E, the velar fronting child, the next stage was one in which fronting was uniformly applied, even in the previously-precocious forms. 33) Amount of Velar Fronting among Precocious Words Sessions # VF tokens # Unfronted tokens Stage 2 1-6 3 16 Stage 3 7-22 58 3 Stage 4:... two weeks after the study, a post-test found no more VF at all. Analyzing later stages 34) Stage 3: *K >> USELISTEDERROR >> IDENT-PLACE With this ranking, the Support still has no errors that will demote *K. And in addition, the velar fronting constraint is now most important: 35) Now even old errors in the Support undergo VF / klaʊn / *K USELISTEDERROR IDENT-PLACE stored loser: [kaʊn] a) klaʊn *! * b) kaʊn *! c) taʊn * * 37)a) This grammar has no VF in old errors / klaʊn / IDENT-PLACE USELISTEDERROR *K stored loser: [kaʊn] a) k(l)aun (*) * c) taun *! b)... or on new forms either /kɑfi/ coffee IDENT-PLACE USELISTEDERROR *K a) kafi * c) tafi *! An open question: How can precocious form become targets of the child-specific process? In other words: why regress all the way in the *K case? Why would USELISTEDERROR ever get demoted below Markedness? A possible beginning of an answer: because *K gets invented for special functional reasons, its specialness carries over to the behaviour of the BCD algorithm Just as a learner can decide to invent this constraint, perhaps they can also decide that it needs to be obeyed all the time, despite the errors it causes. Moving onto the end state on the basis of this (and other errors), the learner will eventually determine that IDENT-PLACE is most important, reaching the final stage 36) Stage 4: IDENT-PLACE >> USELISTEDERROR, *K 9