Symposium
Chairs: LouAnn Gerken and Rebecca L. G—mez
Discussant: James L. Morgan
The four papers to be given in this symposium present cutting edge researchusing artificial grammar learning to explore humans' ability to findsyntactic structure in the linguistic signal. More importantly, the fourpapers articulate competing interpretations of data on infant artificialgrammar learning and what it tells us about the process of learning one'snative language.Papers by Saffran and Altmann demonstrate the potential power ofstatistical learning mechanisms. Saffran provides a convincingdemonstration that 'words' extracted statistically from speech function assuch in infants' representation of language. Infants also haveexpectations about how statistically extracted words should participate insyntactic contexts. Specifically, when newly extracted words are placed inEnglish sentence frames, infants show a different pattern of preferencebehavior than when words are placed in a nonsense syllable frame.Altmann discusses how statistical learning mechanisms can account forlearners' abilities to generalize from sentences produced in one vocabularyto sentences produced by the same grammar in new vocabulary. He proposesthat recent demonstrations by Gomez and Gerken and by Marcus can beaccomplished without reference to mental rules. Altmann also describes astudy in which statistically-based generalization from scenes to sentences(instead of from sentences to sentences) could serve as the basis forlearning the grammars of actual human languages.Papers by Marcus and by Gomez and Gerken explore more traditional, rule-andcategory-based representations of language. Marcus presents datademonstrating generalization based on discrimination of AAB and ABBpatterns. He discusses nine non-rule-based interpretations of these dataand why he believes these are incorrect. In doing so, he attempts toilluminate what is meant by a linguistic rule. Gomez and Gerken proposethat the debate about whether infants abstract algebra-like rules misses animportant point about language learning, namely, that linguisticgeneralization is rooted in categories (instead of pattern-based rules).They present results from several experiments with adults and infantsdemonstrating the conditions underlying formation of grammaticalcategories.The papers in this symposium will promote discussion about the nature ofhuman language and how it is acquired. The issue of statistically-basedversus rule-based characterizations of language will figure prominently.Another issue concerns findings of novelty versus familiarity preferencesin infant artificial grammar learning studies. Which type of preference aresearcher can expect to find is becoming a serious question in the studyof infant language research. The discussant, James Morgan, has studied bothartificial grammar learning and infant language perception. He has alsoengaged in statistical modeling of language input to children and istherefore ideally suited to highlighting key theoretical issues in the ruleversus statistics debate. In short, the symposium promises to engenderheated but illuminating discussion of critical questions in human languagelearning.
Details of individual items:
paper
To what extent do laboratory language learning tasks tap infants' abilitiesto acquire natural languages? We addressed this question by examininginfant word segmentation: the discovery of word boundaries in continuousspeech. One type of information which may be useful in word segmentation isstatistical patterns in language input, namely the probabilities with whichsyllables co-occur. Prior research suggests that 8-month-olds can detectthese statistical patterns. However, the representational structure of theoutput of this learning process is unknown. Consider the sequence 'tibudo'.Following segmentation of 'tibudo', infants might continue to treat'tibudo' as a relatively coherent sound sequence with high internalprobabilities, but with no particular status with respect to their nativelanguage. Alternatively, infants might now treat 'golabu' as a potentialword in their native language.This study was designed to determine whether statistical learning actuallygenerates novel word-like units, rather than probabilistically-relatedsound strings. Eight-month-olds heard a two minute continuous stream oftrisyllabic nonsense words with no acoustic cues to word boundaries (e.g.,'patikugolabudaropitibudogolabupatiku...'. Listening preferences for wordsand part-words (trisyllabic sequences spanning a word boundary) were thenmeasured. These targets were embedded in two different types of sentenceframes. Infants in the English condition heard words and part-words insimple English contexts (e.g., 'I like my tibudo.'). If infants aretreating the newly segmented sound sequences as novel English words, thenwe predicted that infants would show a familiarity preference for wordsover part-words when the targets were embedded in English, as words shouldbe more natural and coherent -- and thus more interesting -- thanpart-words in English sentences.To ensure that the predicted familiarity preference for the English groupwas due to the native language status of the frames, rather than an effectof embedding targets in any type of frame, a second group of infants weretested on the same targets embedded in nonsense frames (e.g., 'zy fike nytibudo.').When embedded in nonsense frames, unlike English frames, wepredicted that infants would continue to show a novelty preference forpart-word targets consistent with dishabituation, as in previous work inwhich test words and part-words were presented in isolation (Saffran,Aslin, & Newport, 1996).In accord with our predictions, listening preferences were affected by thecontext (English versus nonsense frames) in which the items from thefamiliarization phase were embedded during testing. The results suggestthat the statistical learning mechanisms used in word segmentation generateword-like representations for subsequent incorporation into the infant'snative language. Moreover, the current findings strongly suggest that thestatistical learning process does contribute to real-world languageacquisition: sound sequences defined only by their statistical propertiesas words were indeed treated as 'better English' than those which were not.In this fashion, the statistical structure of the input heard during abrief laboratory learning experience interacted with infants' currentknowledge of English. Infant learning mechanisms are apparently capable ofshaping what is initially purely statistical information into thebeginnings of lexical representations.
paper
In a recent article, my colleagues and I (Marcus, Vijayan, Bandi Rao, &Vishton, 1999) argued that seven-month-old infants had the ability to learnabstract algebraic rules. In these experiments, infants were habituated fortwo minutes on sentences like 'la ta la' and then tested on sentences like'wo fe wo' ('consistent' with the habituation sentences) and 'wo fe fe'('inconsistent' test sentences). In each of three experiments, we foundthat infants looked longer at the inconsistent test sentences; from this,we proposed that infants appeared to be learning abstract algebraic rules.Although some found this interpretation interesting, many have challengedit (Altmann & Dienes, 1999; Christiansen & Curtin, 1999a; Christiansen &Curtin, 1999b; Dominey & Franck, 1999; Eimas, 1999; Gasser & Colunga, 1999;McClelland & Plaut, 1999; Negishi, 1999; Seidenberg & Elman, 1999a;Seidenberg & Elman, 1999b; Shultz, 1999). In this talk, I aim to synthesizethese responses with three more positive commentaries (Berent, 1999;Kuehne, Gentner, & Forbus, 1999; Shastri & Chang, 1999) and several repliesby myself (1999a; 1999b; 1999c; 1999d; 1999e), as well as with relatedexperimental work by Gomez and Gerken (1999).Specifically, I will reconstruct what we meant by a rule, compare andcontrast nine competing computational models of our results (Altmann &Dienes, 1999; Christiansen & Curtin, 1999a; Dominey & Franck, 1999; Gasser& Colunga, 1999; Kuehne, Gentner, & Forbus, 1999; Negishi, 1999; Seidenberg& Elman, 1999a; Shastri & Chang, 1999; Shultz, 1999), and consider therelation between the abstract grammars used in our experiments and the'real' grammars that children learn.
paper
Language acquisition is about more than simply learning a language; it isabout learning the mapping between that language and the world which thatlanguage can describe. In an attempt to mimic some of the problems thatface the young language learner, cartoons were created in which objectsinteracted in various ways. For each cartoon, a corresponding sentence wasrecorded that encoded the objects, their roles, and the nature of theinteraction. Adult participants first watched all the cartoons, thenheard all the sentences (without any cartoons), and finally, were tested onthe equivalent of a picture-verification task. Each cartoon was pairedeither with its matching sentence, or with a mismatched sentence matched onlength, and participants had to say which pairs were correctly orincorrectly matched. Participants performed significantly above chance,despite the fact that the sentences were in a language that was entirelynovel to them (composed of nonsense syllables obeying an underlying grammarthat was 'head-final' - cf. Japanese).This finding extends prior demonstrations of 'transfer' of grammaticalknowledge across domains (e.g. Altmann et al. 1999; Gomez & Schvaneveldt,1994; Reber, 1969). Typically, in such studies, participants are presentedwith exemplars of a grammar instantiated in one vocabulary, and are thenpresented with new exemplars, instantiated in another vocabulary. Some ofthese obey the grammar, but others do not. Participants are generally ableto discriminate grammatical from ungrammatical.I shall argue that the processes that enable the mapping between domainsobserved in such studies could in principle underlie the mapping betweensentence and world that is presumed to underpin early language acquisitionin children. I shall describe the principles underlying a neural networkcapable of performing the same feat as our participants in the transferstudies (Dienes, Altmann, & Gao, 1999) and shall consider the widerimplications of these principles for theories of language acquisition.During the course of this discussion, I shall draw on recent work by Gomezand Gerken (1999), and subsequently Marcus (Marcus et al., 1999), who havedemonstrated that very young infants are able to effect transfer acrossvocabularies. Whereas Marcus argues that this ability can only beexplained by recourse to a mechanism able to induce abstract algebraic-likerules, I shall describe a simulation of the Marcus data which suggests thatsuch explicit rule formulation is not required (Altmann & Dienes, 1999).All that is required is a mechanism that is able to map statisticalregularities within one domain onto statistical regularities withinanother. To accomplish this, variation within one domain must be predictiveof variation within the other. This predictive relationship defines thecorrespondence between the domains, and is responsible for enabling themapping process. It defines also the mapping that children must induceduring language acquisition. I shall conclude that an algebraic rule-likesystem could not, in fact, account for the mapping between language andworld that must be effected in early language acquisition.
paper
Recently, Marcus et a. (1999) showed that 7-month-olds exposed to speechsamples of strings with AAB or ABB word patterns discriminated strings withthe training pattern from those with a new pattern despite a change invocabulary between training and test. Gomez and Gerken (1997, 1999) showedsimilar abstraction with older infants, using a more complex grammar.Marcus et al. argued that the only system capable of accounting for suchlearning is one involving abstraction of algebra-like rules, e.g. 'thefirst item X is the same as the second item Y.'These findings are important for demonstrating that infants can abstractbeyond specific word-strings. However, the implications for a theory oflanguage acquisition are limited. Consider three types of abilitiesavailable to the learner: stimulus-based processing, pattern-basedgeneralization and category-based abstraction.Stimulus-based processing involves noting statistical relations amongparticular elements in the physical stimulus. Saffran et al. (1996)reported that infants use the transitional probabilities among syllables insequence to determine whether particular syllables belong to the sameword-like constituent.Pattern-based generalization can be described by means of algebra-likeoperations over physical stimuli. Recognizing ba-po-ba and ko-ga-ko asinstances of the pattern ABA entails noting the identity relation betweenthe first and third syllables within the sequence. This type ofgeneralization might allow a learner to represent some aspects of language,but its role is limited because the variables involved are stimulus bound,generating very specific rules (if element 1 element 3, thengrammatical).In contrast, category-based generalization requires operations overabstract variables. Compare the pattern-based representation ABA to thecategory-based Noun Verb Noun. Although superficially similar, theseexamples differ critically. Abstracting the pattern ABA from ba-po-bainvolves noting that the first and third elements in sequence arephysically identical. Learners must ignore physical similarity, however,when identifying members of abstract categories. 'Dogs chase cats' and'John loves books' share the same category-based structure, despite thephysical dissimilarities among category members.We conducted a series of studies designed to investigate category-basedabstraction in a distributially defined artificial language. Categoriesplay an important role in language acquisition because once a novel word isidentified as an instance of a category, that word then inherits all theproperties associated with that category. (Maratsos, 1982). For instance,any English speaker hearing 'The wug hit the ball' knows that 'wug' canalso be paired with 'a' as in 'A wug hit the ball.' This is becauseEnglish speakers have identified 'the' and 'a' as members of the samecategory. Therefore, any word that can be paired with 'the' can also bepaired with the 'a.'Infants and adults were exposed to sentences defined by the grammarS-->{aXbY or bYaX}, where a, X, b, and Y were instantiated by distinctwords, e.g., a-->{ud, im} and X-->{vot, pel, jic, rud}. Learners wereexposed to a subset of legal category combinations to see whether theywould generalize to new legal combinations. We discuss the conditionspromoting such abstraction and the implications for language acquisition.