Below is the unedited penultimate draft of:

This is the unedited penultimate draft of a BBS target article that has been accepted for publication (Copyright 1999: Cambridge University Press U.K./U.S. -- publication date provisional) and is currently being circulated for Open Peer Commentary. This preprint is for inspection only, to help prospective commentators decide whether or not they wish to prepare a formal commentary. Please do not prepare a commentary unless you have received the hard copy, invitation, instructions and deadline information.

For information about subscribing or purchasing offprints of the published version, with commentaries and author's response, write to: journals_subscriptions@cup.org (North America) or journals_marketing@cup.cam.ac.uk (All other countries).

Lexical Entries and Rules of Language:
A Multidisciplinary Study of German Inflection

Abstract

It is hypothesized following much work in linguistic theory that the language faculty has a modular structure and consists of two basic components, a lexicon of (structured) entries and a computational system of combinatorial operations to form larger linguistic expressions from lexical entries. This target article provides evidence for the dual nature of the language faculty from recent results in a multidisciplinary investigation of German inflection. We have examined (i) its linguistic representation focussing on noun plurals and verb inflection (participles), (ii) processes involved in the way adults produce and comprehend inflected words, (iii) brain potentials generated during the processing of inflected words and (iv) the way children acquire and use inflection. It will be shown that the evidence from all these sources converges and supports the distinction between lexical entries and combinatorial operations.

Our experimental results indicate that adults have access to two distinct processing routes, one accessing (irregularly) inflected entries from the mental lexicon and another involving morphological decomposition of (regularly) inflected words into stem+affix representations. These two processing routes correspond to the dual structure of the linguistic system. Results from event-related potentials confirm this linguistic distinction at the level of brain structures. In children's language, we found these two processes also to be clearly dissociated; regular and irregular inflection are used under different circumstances, and the constraints under which children apply them are identical to those of the adult linguistic system.

Our findings will be explained in terms of a linguistic model that maintains the distinction between the lexicon and the computational system but replaces the traditional view of the lexicon as a simple list of idiosyncrasies with the notion of internally structured lexical representations.

Keywords

grammar, psycholinguistics, neuroscience of language, child language acquisition, human language processing, development of inflection

Harald Clahsen is Professor of Linguistics and has published 5 books and over 40 research articles on language acquisition, developmental language disorders, and grammatical processing in adults. He has published 5 books and more than 40 research articles on these topics. He is the recipient of Gerhard-Hess award from the German Science foundation and an award for his 1991 book Child Language and Developmental Dysphasia.

1. Combinatorial operations and lexical entries

A core property of human language is the unlimited possibility of expression despite limited resources. Speakers and hearers have the ability to create sentences and other linguistic expressions they have never heard before, even though their vocabularies, memories, attention spans, etc. are limited. This property of human language can be explained by assuming a dual architecture of the language faculty with two separate components, a lexicon consisting of a list of lexical entries and a computational component for combining lexical entries. The lexicon is finite and specifies for each entry its category membership (`N(oun)', `V(erb)' etc.), and idiosyncratic information regarding its form and meaning. The computational component is conceived of as a finite set of rule-like operations which take lexical entries as inputs to form larger linguistic expressions, e.g. phrases and sentences. Since combinatorial operations apply recursively, they can generate an infinite number of expressions of arbitrary length. Moreover, as combinatorial operations manipulate abstract symbolic categories such as V and N rather than sounds or meanings directly, they can produce unusual sequences of words (Colourless green ideas...), novel sentences and even sentences which are structurally well-formed but completely meaningless (Mopy squitters blipped ruttily op en glurk). In this way, the distinction between lexical entries and combinatorial rules explains the `discrete infinity' of human language (Chomsky 1995). It should be pointed out that the assumption of a dual structure of the language faculty does not hinge on the adoption of a particular linguistic theory or formalism, and that, consequently, the interpretation of our empirical findings hinges only on whether the phenomena are to be accounted for by combinatorial operations (such as rules of language), or by (access to) lexical entries.

Familiar arguments in favour of combinatorial operations for language and a mental grammar come from syntactic phenomena such as strict recursion and long-distance dependencies. Consider, for example, long-distance dependencies like in wh-questions: Who did he say that John loved?. Syntactic analyses of wh-questions posit computational mechanisms, e.g. wh-movement (Chomsky 1981, 1995), to derive structural representations of such strings. In addition, a complex of constraints acting in tandem ensures that strings such as *Who did he say that John loved Mary in which an object expression follows the verb love (even though love normally takes direct objects) come out as ungrammatical. In this way, computational mechanisms such as constraints on movement correctly account for the syntactic properties of wh-questions.

Distinctions between lexical entries and combinatorial operations have also been made in psycho- and neurolinguistic research, with respect to the processing of language in real time, the acquisition of language by children and the representation of language in the brain. Linguistically, morphological roots have lexical entries, but sentences typically don't (proverbs, cliches, etc. may however be exceptions in this respect). Hence, the processing of roots has traditionally been viewed as a search process with the goal of finding an entry in the mental lexicon, whereas sentence processing involves the construction of a syntactic representation in accordance with grammatical principles (see however Boland & Cutler 1996 for some critical discussion). In research on the acquisition of language, a distinction is made between two different kinds of mechanisms: children may represent newly acquired linguistic material unanalyzed in lexical entries, or they may use combinatorial rules to construct grammatical rules (Berko 1958, Pinker 1984). Finally, the two components of the language faculty seem to elicit different brain potentials. Electrophysiological measures of neural activity have shown, for example, that semantically anomalous words elicit a characteristic brain response, the so-called N400 (Kutas & Hillyard 1980, and subsequent work), and that syntactic anomalies produce other effects (see section 4.5 below). Other brain-imaging studies suggest that different parts of the brain are active for lexical (Petersen et al. 1989, Petersen & Fiez 1993) and grammatical processing (Stromswold et al. 1996, Indefrey et al. 1996).

The view of the dual nature of the language faculty has recently come under attack from a group of researchers who seek to develop associative single-mechanism models of language and subscribe to a school of cognitive science known as connectionism (Elman et al. 1996, Rumelhart & McClelland 1986, Quartz & Sejnowski 1998). A conception of knowledge of language is advocated here that tries to make do without the machinery of internally-represented, symbol-manipulating combinatorial operations. Instead, it is argued that what looks like the application of symbolic principles or rules can better be represented in terms of associative networks operating without any directly implemented combinatorial principles. Associative networks consist of units called nodes and weighted connections between those nodes. All kinds of inputs and outputs are represented in the same way in these systems, i.e. as patterns of activation over these nodes. For some linguistic phenomena which have traditionally been viewed as following from rules of language, e.g. the English past tense (Halle & Mohanan 1985), associative single-mechanism networks have been proposed which are claimed to provide adequate accounts of inflection without invoking any kind of combinatorial operations. The most widely-known example is Rumelhart & McClelland's (1986) pattern-associator network of past tense inflection in English, which contains no inflectional rules, but which can mimic rule-like behaviour and makes errors reminiscent of those made by children; see MacWhinney & Leinbach (1991), Plunkett & Marchman (1991, 1993), Hare et al. (1995) for more recent networks of the English past tense. Some researchers have taken these simulations to argue that the knowledge of language should better be represented in terms of connectionist networks with associatively linked units rather than in terms of grammars (Bates & Elman 1993, Seidenberg 1997, Churchland 1995, Elman et al. 1996). It is claimed that if a reasonable account of linguistic phenomena can be given by a single type of mental mechanism, then - other things being equal - there would be no need for a dual architecture or for separate cognitive systems such as the `lexicon' and the `computational system'.

Apparently, however, things are not (yet) equal. Connectionist contributions to syntax have been limited (see Elman et al. 1996), and associative models of language seem to be incapable of properly implementing syntactic phenomena (Fodor et al. 1974, Fodor & Pylyshin 1988, Marcus 1998). To take a famous example, Elman's (1993) network of subject-verb agreement in English was set up to do word prediction, takes in sequences of words from a training set and predicts possible continuations. But when Marcus (1998) tested this model on words that were outside the training set, he found that it could not generalize in the way that humans do. Thus, it remains to be seen whether associative models will ever be able to successfully handle complex syntactic phenomena such as strict recursion and long-distance dependencies. Connectionist networks of the English past tense have also been shown to have severe deficiencies (Pinker & Prince 1988, Marcus et al. 1995, Marcus 1998), and it has been argued that they do not make a case for abandoning the idea of a dual structure of the language faculty including a mental grammar; see section 2 for some discussion.

The research strategy I will adopt here is to produce empirical evidence that might bear on the controversy between dual and single-mechanism models of language. My colleagues and I have studied (German) inflection in intensive detail from the perspective of different disciplines investigating its linguistic structure and historical development, how it evolves in child language acquisition, how it is produced and comprehended in real time, how it is processed in the brain, and finally, how it is affected by language disorders. We believe that such a multidisciplinary approach contributes to a better understanding of the mechanisms involved. The present target article shows that the results from these different investigations converge and that they provide new evidence for the dual nature of the language faculty.

With respect to language, most of the empirical evidence in the connectionist-symbolist debate comes from the study of the English past tense, including studies on child language acquisition (e.g. Marcus et al. 1992, Bybee & Slobin 1982), adult language processing (e.g. Stanners et al. 1979, Stemberger & MacWhinney 1986), brain-imaging studies and event-related potentials (Jaeger et al. 1996, Ullman et al. 1997, Münte et al. 1998a), connectionist simulations (e.g. Elman et al. 1996) and language disorders (e.g. Marslen-Wilson & Tyler 1997, Gopnik 1994, Clahsen & Almazán 1998). With respect to aphasia, for example, several researchers have pointed out contrasts between regular and irregular inflection in agrammatic production (Kean 1977, Lapointe 1983, Grodzinsky 1990). Perhaps the strongest evidence to date for a regular-irregular dissociation in the English past-tense system comes from a cross-modal priming study with English-speaking aphasics (Marslen-Wilson & Tyler 1997). Marslen-Wilson & Tyler found that one subgroup of subjects exhibited (partial) priming effects for irregulars, but had lost (full) priming of regular verbs; another patient showed exactly the opposite pattern. Thus, there seems to exist a double dissociation between regular and irregular past tense forms in these patients, which indicates that the processes underlying regular and irregular inflection can be selectively impaired and therefore belong to different cognitive systems. We found similar dissociations for developmental disorders in children (Clahsen & Almazán 1998). We have investigated 4 English-speaking children with William's syndrome (chronological age 11;2-15;5, mental age 5;4-7;6) and compared them to two control groups matched for mental age, normal subjects and specifically language impaired (SLI) subjects. Our findings indicate no impairments of the WS subjects on complex syntactic tasks, such as the interpretation of reversible passives and of sentences with reflexive anaphors, and on regular past tense formation; on these three phenomena, the WS subjects achieved the same level of performance as unimpaired controls, with a mean correctness score of 91%. On irregular verbs, however, the WS subjects performed much more poorly, with only 42% correct compared to 78% correct for the mental age normal controls. In SLI subjects, the results go the other way, with regular past tense forms producing higher error scores than irregular ones (see also Ullman & Gopnik 1994). It remains to be seen whether such double dissociations can be found for other inflectional phenomena and for languages other than English; see Penke (1998) for findings on selective impairments in the inflectional systems of German-speaking aphasics.

Several attempts have been made to simulate the properties of the English past tense in connectionist single-mechanism models. A detailed discussion of these models can be found in Pinker & Prince (1991), Marcus et al. (1995), Marcus (1995, 1998). The following presents a brief review. Rumelhart and McClelland (1986) conducted a connectionist simulation of the acquisition of the English past tense system in which a pattern associator that directly takes a simplified quasi-phonological representation (= 'Wickelfeatures') of the stem as input and computes a corresponding phonological representation of the past tense form as output. Correlations among pairs of features in the stem and the past are stored, and based on the strength of the association between the stem features and various output features in the training set, the model generalizes to new verbs. The challenge of this model is that it claims to represent the English past tense system including its acquisition in a unitary associative architecture without any recourse to morphological rules of inflection. However, Pinker and Prince (1988) and Marcus et al. (1992) pointed out several deficiencies of Rumelhart and McClelland's model. For example, the pattern associator can learn arbitrary input/output mappings, even those that are linguistically impossible ones, e.g. string reversals. Marcus et al. (1992) also disconfirmed Rumelhart and McClelland's claims about vocabulary development and the assumed corelation with overregularization errors.

More recently, several new connectionist models have been proposed, each addressing one of the criticisms pointed out by Pinker & Prince and by Marcus et al. MacWhinney & Leinbach (1991), for example, built a network that can distinguish past tense forms of homophones (ring - wring), something the Rumelhart & McClelland pattern associator could not do; Daugherty & Hare's (1993) model correctly produces -ed forms of denominal verbs, even when they rhyme with existing irregulars; Hare et al.'s (1995) simulation can generate regular past-tense forms of low-frequency verbs; the model proposed by Plunkett & Marchman (1993) mimics the U-shape curve of the development of -ed in child language acquisition. However, all these models have been claimed to create new problems, as pointed out by Marcus and collaborators (Marcus et al. 1995, Marcus 1995, 1998). For example, MacWhinney & Leinbach (1991) only examined a small number of verb pairs that by historical accident are homophonous and have different past tense forms in English, and they directly built features into their network for just these pairs (e.g. to lie). While this correctly reproduced the facts for these verbs, it does not account for the fact that derived verbs, even if they are homophonous to existing irregulars, are always regularly inflected (Napoleon rang the bell, and his soldiers ringed the city); Marcus et al. (1995: 211f.) for further discussion of MacWhinney & Leinbach. Daugherty & Hare (1993) dealt with the inflection of denominals by adding extra input nodes which encode how a denominal verb is related to its head noun. This produced the correct output for these verbs, but Daugherty & Hare's network had to be particularly trained on verb forms such as ringed (the city), spitted (the chicken), etc., an unrealistic requirement given that speakers of English do not seem to depend on having heard such forms in the input (Kim et al. 1994, Marcus et al. 1995: 212). Hare et al. (1995) addressed the problem of generalizing a low-frequency default by representing the default affix as a discrete atomic label for a single output unit. But -ed is language-particular, and therefore hard-wiring of -ed does not seem to be a realistic option. Moreover, as pointed out in Marcus (1998), semi-regular verbs such as to sleep and suppletive verbs such as to go pose problems for Hare et al.'s network, because the network's architecture systematically prevents massive stem changes (like went) and blends (like slept). In contrast to claims made by Plunkett & Marchman (1993), Marcus (1995:278) argued that their model's performance differs from actual children's development in important ways. For example, the Plunkett & Marchman model produced a U-shaped learning curve only after abrupt changes in the training regime, while in children's development there are no radical changes in the input. Moreover, irregularization errors (flow --> flew) are extremely rare in children (Xu & Pinker 1995), but the Plunkett & Marchman model produced them more often than -ed overregularizations.

As is clear from this brief review, many of the criticisms of the original Rumelhart & McClelland model have been addressed in more recent connectionist networks of the English past tense. The proposed additions, however, seem to have led to new problems. The fundamental deficiency of these models is that given enough training they can master the data of the training set, but in contrast to humans, their ability to generalize to novel items is limited and depends on particular statistical patterns in the input; see Marcus (1998) for further discussion.

Summarizing, we saw that despite all the evidence that has been accumulated on the English past tense, the representation and analysis of this system is still highly controversial and it is hard to see how the disagreements could be resolved. Perhaps the English past tense is not the most appropriate inflectional system for examining the distinction between memory- and rule-based representations for language. English, compared to other languages, is inflectionally poor. It has only one productive past tense suffix, the regular -ed. Also, regularity is confounded with both the presence of an overt affix and with type frequency: that is, regular verbs in English are much more (type)-frequent (= 95%) than irregular ones (= 5%), see Marcus et al. (1995), and regular past tense forms contain a segmentable affix, whereas irregular forms typically do not have affixes [1]. These two confounding features leave room for several alternative interpretations of the same set of facts. Potential differences, for example, between forms such as walk-ed and came could be effects of frequency differences and/or effects of the presence or absence of an overt affix, rather than the result of different underlying representations.

In our research we have examined two subsystems of German inflection, past participle inflection and noun plurals, from a multidisciplinary perspective. Like English past tense forms, German participles and noun plurals are easy to study, but more importantly, they do not confront us with the confounding variables (in terms of word structure and frequency distribution) that make results on the English past tense so hard to interpret. We have investigated both verb and noun inflection to allow for generalizations across individual inflectional systems. The following sections will present results from these studies.

The basic idea of a dual-mechanism approach to inflection is the postulation of two qualitatively different clusters of inflectional phenomena, lexically-based inflection versus inflection based on combinatorial rules, or - to use Pinker & Prince's (1991) terms - the distinction between irregular and regular (default) inflection. The latter is meant to capture the true productive aspects of inflectional morphology: regular (default) inflection can be easily decomposed into stem+affix and involves affixation processes which may operate on the outputs of other morphological operations (derivation, compounding); regular inflection readily extends to novel items. Lexically-based inflection, on the other hand, epitomizes the idiosyncratic aspects of inflectional morphology as well as sublevel regularities of inflectional patterns with varying degrees of systematicity such as the family resemblance patterns among irregular past tense forms of English (sing - sang, ring - rang, etc.); these irregular forms cannot be necessarily predicted from their corresponding base forms, and they are only tentatively extended to new forms.

We will adopt an approach to inflection in which the distinction between lexically-based and rule-based inflection is made explicit and constitutes a design feature of the model, Minimalist Morphology (Wunderlich 1996, Wunderlich & Fabri 1995). Minimalist Morphology distinguishes between regular inflection and lexically restricted inflection, and posits two qualitatively distinct linguistic mechanisms for them: structured lexical entries and affixation [2]. Lexically restricted (irregular) inflection is not rule-based, but rather encoded in the lexical items themselves. Affixation on the other hand, is a combinatorial process which concatenates an affix with a lexical entry. Affixation is only constrained by a small set of general (probably universal) constraints, for example, the constraint that more specific lexical entries take precedence over less specific ones in affixation [3].

With respect to the English past tense system, for example, Wunderlich & Fabri (1995) claim that regularly inflected forms such as walked are derived by affixation (Add -ed), while irregular past tense forms are represented as subnodes of lexical entries; consider, for illustration, the entries for verbs such as drive and drink:

Each node in a structured lexical entry represents a pair (<phonological string, morphological feature value>), and each subnode inherits all information of its mother, except for the features it replaces or adds; for example, the subnode [..æ..]_+pret inherits the onset dr-, the coda -nk and the categorial feature [+V] from the higher node. Subregularities among irregular past tense forms are captured through lexical templates in which stem segments are associated with segments from subnodes; for example, the [..æ..]_+pret subnode does not only occur for drink but for several other verbs with the segments -ing- and -ri- in the base entry (ring, sing, spring, etc.) which therefore constitute a lexical template.

German participle formation involves two endings [4], -n appears on all participle forms of so-called strong (= irregular) verbs, and -t appears on participle forms of all other verbs (see Appendix for details). The participle -t suffix also applies to words for which lexical entries are not readily available, such as nonsense words (faben   'gefabt'), low-frequency words (löten 'to solder'   'gelötet'), onomatopoeia (brummen 'to buzz'   gebrummt), and verbs derived from adjectives or nouns (das saubere Haus 'the clean house'; wir säubern das Haus 'we clean the house'; wir haben das Haus gesäubert `we cleaned the house'). In this regard, the participle -t behaves like the English past tense suffix -ed. By contrast, the participle -n does not apply under such circumstances. The participle -n co-occurs with phonologically unpredictable stem changes, whereas -t suffixation on weak (= regular) verbs does not involve any stem allomorphy (see Appendix for examples). Hence, -n participle forms behave like irregular past tense forms in English. To capture these differences, Wunderlich and Fabri (1995) posit two distinct mechanisms for German participle formation, a -t affixation rule and lexical entries for irregular verbs; see Appendix for further details.

German plurals are formed using five different endings (-n, -s , -er, -e and -0) along with possible vowel changes (see Appendix for examples). None of the five endings is statistically predominant, and the use of these endings with specific nouns is not readily captured by standard inflectional rules. Despite its overall irregularity the German plural system has been shown to provide a default process which applies when irregular forms are not accessible. Marcus et al. (1995) demonstrate that according to their linguistic criteria, -s plurals fall into the default cluster, even though they are extremely rare in the German language compared to other plural forms. The -s plural applies when the phonological environment does not permit any other plural allomorph. It occurs on masculine, feminine, and neuter nouns, on words that exhibit the canonical stress pattern and on those that don't, on monosyllables and polysyllables and on both vowel-final and consonant-final stems. The -s plural also generalizes to rootless and headless nouns, e.g. to nominalized conjunctions such as die Wenns und Abers `the Ifs and Buts', to eponyms and product names (Fausts, Golfs, etc.) and to nominalized VPs (die Rührmichnichtans `the Touch-me-nots'). This contrasts with other (irregular) plurals which are restricted to particular morphophonological conditions. Given the Minimalist Morphology framework, we distinguish between two kinds of morphological processes involved in German noun plural formation, an -s affixation rule and lexical entries for irregular plurals; see Appendix for linguistic details.

To sum up so far: We began by characterizing the dual nature of the language faculty according to which knowledge of language involves two distinct cognitive mechanisms, a symbol-manipulating system of combinatorial operations and lexical entries. We have chosen to examine the phenomenon of grammatical inflection from this perspective focussing on two systems of German inflection, noun plurals and participle formation. A linguistic analysis of these inflectional systems was presented which made use of two kinds of linguistic representations, affixation and structured lexical entries. Affixation covers the regular aspects of inflection and belongs to the system of combinatorial operations. The irregular aspects of inflection are encoded in lexical entries. To capture subregularities of inflectional patterns, lexical entries were claimed to have internally structured representations (`default inheritance trees') and to be linked to related entries through lexical templates.

The view of the dual nature of the language faculty has further implications beyond the level of linguistic analysis. If the idea of two distinct cognitive systems for language is real and fundamental, rather than just a convenient linguistic description or an epiphenomenon of frequency and similarity clusters in verbal memory, we would expect that rule-based inflection should dissociate from lexically-based inflection from different points of view, not just in analyses of inflectional systems in natural language grammars. We would expect to find corresponding dissociations, for example, in human language processing, with respect to neural structures in the brain, in the way children acquire these two types of phenomena over time, etc. We have used several experimental methods and different groups of subjects to investigate German inflection with respect to these domains, and as will be shown in subsequent sections, the evidence from these sources converges on the predicted dissociations, thus supporting the dual nature of the language faculty.

In the previous section, we assumed that the language faculty has a dual architecture comprising combinatorial principles and a structured lexicon. This view raises two closely related psycho-/neuro-linguistic questions: (i) How are these two mechanisms used in speech production and comprehension? and (ii) How are they represented in the brain? We will address these questions separately, taking the processing of German noun plurals and participles as the focus of our empirical investigation. Before turning to the empirical evidence, however, let us briefly consider how the mental grammar might be connected to processes involved in the production and comprehension of language.

The strongest and (to me) most interesting view concerning the grammar-processing relation is the correspondence hypothesis (originally proposed by Miller & Chomsky 1963) according to which the mental grammar is directly used in language processing. This means that grammatical rules and principles are mentally represented and that in recognition and production the language processor constructs such representations using the normal structures and operations of the grammar. The appeal of the correspondence hypothesis is that it provides a parsimonious and straightforward account of how grammatical knowledge and processing are related: the parser is said to make basically the same distinctions as the grammar (Jackendoff 1997, Phillips 1996).

With respect to morphology, psycholinguists have examined the role of morphological structure in the processing of morphologically complex words and the question of whether there is any correspondence between the linguist's decomposition of a morphologically complex word and the way it is segmented by the speaker-hearer during online production and comprehension. Current morphological processing models, however, provide conflicting answers to these questions. Some researchers (e.g. Butterworth 1983, Manelis & Tharp 1977) claim that the morphological structure of words plays no role in the way they are produced or perceived and that morphologically complex words are fully listed in memory. Recent connectionist models (Rumelhart & McClelland 1986, MacWhinney & Leinbach 1991, Plunkett & Marchman 1993, among others) are similar in spirit and can be viewed as implementations of full-listing models. All inflected words, such as irregular and regular past-tense forms of English, are said to be stored in terms of distributed representations in associative memory [5]. In contrast to this, the full parsing model of Taft and his collaborators claims that only stems have entries in the mental lexicon and that morphological variants need to be decomposed in processing before their stems can be accessed (see e.g. Taft 1979). This model assumes global affix-stripping mechanisms for processing purposes which do not necessarily correspond to the morphological structure of a complex word. Several other models of morphological processing have incorporated whole-word based representations with morphological decomposition in so-called dual-route models (Laudanna & Burani 1985, 1995, Frauenfelder & Schreuder 1992, Schreuder & Baayen 1995). In these models, the language processor is said to make use of both full-form representations of morphologically complex words and morphological decomposition. Word frequency and phonological transparency seem to be crucial factors in determining which of the routes is more efficient: highly frequent phonologically non-transparent words are more likely to have full-form representations than low-frequency transparent words. However, the extent to which these two processing routes depend upon morphological structure is viewed as controversial (see e.g. Baayen et al. 1997a).

According to the correspondence assumption, structural properties of inflected and (derived) words should converge with their processing properties, a sensible prediction if our ultimate goal is an integrated theory of brain and mental functions underlying language. Dual mechanism morphology distinguishes between regular and irregular morphological processes, the former involving affixation and the latter (structured) lexical entries. Hence, given the correspondence assumption, dual mechanism morphology leads us to expect a parallel distinction in morphological processing: inflected words that have stem+affix representations should be computed via their constituent morphemes, whereas inflected words that are represented in terms of (structured) lexical entries should exhibit associative memory effects in processing experiments. This means that for inflected words which are formed through affixation we would expect to find evidence for morphological decomposition in psycholinguistic experiments. By contrast, irregularly inflected words should produce effects which are characteristic of lexical entries in morphological processing experiments. It is an extremely strong (probably too strong) prediction that in any language one should find that linguistic and psycholinguistic properties of inflectional phenomena always converge into either of these two clusters. However, as Pinker & Prince (1991:16) stress `even partial confirmations across languages would offer new insight into the role of predictability, productivity, and statistical patterns in influencing grammar and linguistic performance'. We will examine these predictions here taking German noun plurals and participle formation as our test cases.

One crucial property of inflectional processes is that they can be easily applied to novel or unusual words under appropriate circumstances (see Bybee & Moder 1983 and Prasada & Pinker (1993) for English). The dual-mechanism model predicts that the generalization properties of regular and irregular inflection are different, however, depending on the linguistic mechanisms involved. Consider participle and noun plural formation in German. Irregular participle and noun plural forms are represented in structured lexical entries (such as those in (ii) and (iv) in the Appendix). Hence, structured lexical entries should only generalize by analogy, i.e. to novel words that are similar to existing ones. Regular -t participle and -s plural inflection, on the other hand, are based on affixation (see (i) and (iii) in Appendix) which may apply to a given syntactic category (`V' or `N') irrespective of any kind of similarity to lexical entries. Hence, affixation-based generalization should apply elsewhere, i.e. even under circumstances in which similarity-driven analogies fail, for example to unusual sounding novel words.

The generalization properties of participle and plural formation processes have been investigated in different kinds of experiments involving inflected forms of novel and uncommon words; the results support the distinction between similarity-based and affixation-based generalization.

Participle formation of nonsense words has been investigated in an elicited production experiment (Clahsen 1997). We constructed novel strong and weak verbs and presented subjects with both past tense and infinitive forms of these novel verbs. Note that past tense formation in German provides an unambiguous cue for determining verb-class membership: weak verbs form their past tenses with -te and without vowel changes, whereas strong verbs all have vowel changes and never carry the -te suffix in the past tense (see (b) - (d) in Appendix). Subjects were first given each nonsense verb in the infinitive and 1st past tense form ('step 1'). They were then asked to use the past tense form from step 1 to fill a blank in another sentence ('step 2'). Finally, they had to fill in a second blank, this time by providing the participle form of the nonsense verb ('step 3'); see the following illustration.

Step 1: presentation of nonsense verb in infinitive and 1st past tense
'Eines Tages kam mein Freund Peter zu mir und fragte mich, ob ich seinen Zatt teiden kann. Es war kein Problem für mich, und ich tied seinen Zatt.'
(One day, my friend Peter came to me and asked me whether I can teiden (=infinitive) his Zatt . That was no problem for me, and I tied (= 1st past tense) his Zatt. (Note that Zatt is a nonword as well.))

Step 2: replication of past tense form
'Es war nicht das erste Mal, daß ich einen Zatt __'
(It was not the first time that I _____ a Zatt.)

Step 3: production of participle form
'Peter sagte: Danke, daß du meinen Zatt ___ hast.'
(Peter said: 'Thank you that you have ____ my Zatt.)

The experiment produced three main results; see Fig.1. First, even in cases in which subjects correctly reproduced the strong past tense form in step 2, -n participle formation is used less often than expected: only 31% (196/637) of the items that subjects reproduced in the strong past tense form in step 2 had the expected -n ending on the participle. Second, in contrast to -n, the -t suffix appeared in nearly all of the expected cases (97% of weak verbs, 665/685 cases), and more importantly, it was heavily extended to strong verbs: 69% (441/637) of the nonsense verbs that were correctly reproduced with strong past tense forms were suffixed with -t in the participle. Third, as shown in the third column of Fig.1, 91% (178/196) of the -n participles were produced on verbs that rhyme with existing strong verbs, e.g. to teiden - tied - getieden on analogy with scheiden - schied - geschieden `to separate'.

Figure 1: Generating participles of nonsense verbs. Suffixes used on different kinds of nonsense verbs. The first column shows that the -t suffix was used on 97% of the weak verbs. The second column shows that the -n suffix was used in 31% of the strong verbs. The third column shows that 91% of the -n participle forms were verbs that rhymed with existing strong verbs.

These results show that the generalization properties of -t and -n participle formation are quite different: -t generalizes widely to all kinds of nonsense verbs while extensions of -n participle formation are narrowly restricted to novel verbs which are similar to existing ones. These differences correspond to the linguistic distinction between affixation and structured lexical entries: -t participles are affixation-based, only constrained by general principles and hence may apply to any kind of verb; -n participles are subnodes of lexical entries which can only be accessed by analogy, and hence the similarity-based extensions to novel verbs found in this experiment.

Similar results have been obtained in a paper-and-pencil judgment experiment on plural formation of nonsense nouns (Marcus et al. 1995). The experimental items were 12 monosyllabic novel nouns that rhymed with existing German nouns that take irregular plural forms, e.g. Pund on analogy with Hund - Hunde `dog - dogs', and 12 non-rhymes. Each item was first presented in a context sentence in its singular form, followed by 8 test sentences each containing all possible plural forms. Subjects were asked to rate each sentence on a 5-point scale for acceptability.

We found that irregular plural forms were judged as significantly better for nouns that rhymed with existing (irregular) nouns than for non-rhymes (p < .001), whereas -s plural forms were judged as significantly worse in the Rhyme condition than in the Non-Rhyme condition (p < .01). Thus extensions of irregular plural formation are sensitive to similarity, whereas -s plurals are applied elsewhere, even to nouns that are dissimilar to any existing German word.

We argued above that -t participle formation involves affixation, whereas the -n ending is part of lexical entries. The experiment to be reported here investigates effects of violations of affixation processes (Clahsen et al. 1997). According to our linguistic analysis, participles of weak (regular) verbs are formed by (-t) affixation. We would therefore expect that incorrectly inflected participles of regular verbs produce affixation-violation effects; for -n participle formation, however, there should be no such effects, since -n participle forms are not affixation-based.

To investigate affixation-violation effects, we adopted the sentence-matching technique from Chambers & Forster (1975) and Freedman & Forster (1985). In this task, subjects are presented with two stimuli on a screen and must decide as quickly and accurately as possible whether these two stimuli are the same or different. It has been shown that the reaction times (RTs) in this task are sensitive to grammatical wellformedness: wellformed words and sentences have shorter RTs than illformed control items. Chambers & Forster (1975), for example, found that subjects responded more quickly to word pairs such as HOUSE/HOUSE than to non-words such as HSEUO/HSEUO, even though the non-word stimuli were of the same length as the word pairs. Similar effects were found for various kinds of morphosyntactic violations, e.g. subject-verb agreement errors, illegal subjects, ungrammatical word-order patterns, etc. (Freedman & Forster 1985, Forster & Stevenson 1987, Forster 1987, Clahsen et al. 1995). In our experiment, we used the matching technique to determine whether regular and irregular inflection behave differently with respect to violations of affixation.

To prepare the subjects for our matching experiment, they were first given a booklet and asked to learn a set of 20 nonsense verbs in the infinitive and corresponding past tense forms, 10 with weak past tense forms, e.g. praupen - praupte, and 10 with strong ones, e.g. flauden - flied. After the subjects had learned the nonsense words at home, successful learning was tested using a cloze test [6]. Only those subjects who correctly reproduced the past tense forms of all nonsense verbs performed the matching task in which they were confronted with -t and -n participle forms of the nonsense verbs they had learnt before, e.g. gepraupt - gepraupen, geflauden - geflaudet. There were four types of test items as shown in Fig.2: nonsense words learnt as weak or strong verbs, and participles presented with -t or -n in the matching task.

Figure 2: Sentence-matching times for novel participles. Nonsense items that were learnt as weak verbs produced shorter mean response times when they were presented with the correct -t suffix compared to being presented with the (incorrect) -(e)n suffix. Nonsense items that were learnt as strong verbs did not produce shorter response times when they were presented with the correct -(e)n suffix.

We found that weak novel verbs incorrectly presented with -n produced longer RTs than the grammatical control condition, i.e. weak verbs correctly suffixed with -t (1,643mec. vs. 1,483msec., p < .001). For strong novel participle forms, however, there was no such effect; rather, -t suffixation, which is ungrammatical for existing strong verbs, produced slightly faster RTs than -n participle formation (1,513msec. vs. 1,552msec., p = .43). These results can be explained in terms of a dual-mechanism analysis of German participle formation in which -t (but not -n) participles are affixation-based. Given the -t affixation process, an -n participle form of a (novel) weak verb involves a violation of an affixation process and hence the observed ungrammaticality effect in the matching task. By contrast, a -t participle form of a novel strong verb does not violate any affixation process or constraint, because -t affixation may apply to any verb and because according to the linguistic analysis proposed here, there is no corresponding -n affixation process. Consequently, in the matching task -t participles of novel strong verbs do not produce longer RTs than -n participles of such verbs.

Words derived from other categories as well as words directly borrowed from other languages are unusual in that they do not have canonical lexical entries. Consider, for example, verbs derived from nouns. A denominal verb such as to spit (the chicken) in the sense of `to put the chicken on a spit' has a complex internal representation; it is based on the lexical entry of a noun, (the) spit, and is headed by a derivational affix which determines the category of the whole word (Wunderlich 1986, Olsen 1990). Hence, in terms of their linguistic structure denominal verbs do not have lexical entries (as verbs), but are rather computed by category-changing affixation [7]. When such derived words are inflected for the past tense, access to lexical entries of verbs is blocked, even though they may sound similar to existing verbs, and the regular default affix is used (Paul spitted the chicken). The ungrammaticality of irregular past tense forms in such circumstances follows from the fact that the lexical entry for the irregular verb form spat is specified for particular syntactic categories (= `V' in the case of verbs) which derived words such as `spit (the chicken)' cannot access given their morphological structure, i.e [<sub>V</sub>[<sub>N</sub> spit] -Ø]. The same applies to plurals of nouns derived from proper names and from borrowings, as for example in expressions such as the Helmuts (= Helmut Kohl and Helmut Schmidt) like cappuccinos. Linguistically, the items Helmut and cappuccino are stretches of sounds which do not have lexical entries. Rather, they bear a non-specific label `X' which needs to be converted into a proper syntactic category, yielding structures such as [[Helmut]_X-Ø]_N and [[cappuccino]_X-Ø]_N which are similar to those of denominal verbs.

In two acceptability rating experiments, one on participles and one on noun plurals, we studied how speakers of German inflect such non-canonical words (Marcus et al. 1995). In the participle study, subjects were presented with novel denominal verbs each of which appeared in two test sentences, one with a -t and one with an -n participle form. Subjects were asked to judge each test sentence on a 7-point scale for acceptability. To control for similarity-based generalizations, we made all the denominal verbs used in the experiment homophonous to existing strong verbs. Hence, if these novel denominals were inflected by analogy to existing lexical entries rather than by affixation, -n participle forms should have higher ratings than -t participles.

What we found, however, was that subjects judged -t participles of novel denominal verbs as significantly better than -n participles (mean ratings: 3.3 vs. 2.1, p < .001), even though the items were homophonous to existing strong verbs. Similar results were achieved on plural formation of derived nouns (Marcus et al. 1995). Recall from above that similarity-based generalizations were found for plurals of nonsense nouns: items that rhyme with existing nouns taking irregular plurals may be irregularly inflected by analogy. However, when the same items were presented to the subjects as proper names or borrowings, the preference for irregular plural formation disappears: -s plurals of borrowings are judged as significantly better than -s plurals of the same items used as simple nonsense nouns (p < .005), and -s plurals of proper names were judged as significantly better than irregular plurals, even for items that rhyme with existing irregulars (p < .001). So, the novel item pund, for example, which is analogous to Hund - Hunde `dog - dogs' has a clearly preferred plural form (= pund-s) when used as a proper name.

These results correspond to the linguistic distinction between affixation and structured lexical entries. Irregular participle and irregular plural forms are based on lexical entries, and these entries are specified for particular syntactic categories which derived words cannot access given their morphological structure. Affixation, however, can be applied to any element of a given category, and hence the clear preference for -t participles of denominal verbs and -s plurals of derived nouns.

Several researchers have used lexical decision experiments (LDE) to test inflected words for memory effects. LDEs are word/nonword discrimination tasks with RT as the dependent variable. Lexical decision times on non-inflected simplex words have been consistently shown to be affected by word frequency: subjects take less time to decide that high-frequency items are existing words than they do for low-frequency items (see Balota 1994 for review). This is conceived of as a memory effect: as memory traces get stronger with additional exposures, high-frequency entries can be more readily accessed than low-frequency ones. LDEs on inflected words have produced conflicting results. On the one hand, a number of studies have found an effect of root frequency (the sum of all forms with a given root) for words equated in word-form frequencies; Taft (1979) showed in an LDE that it takes less time to recognize a word like sized which has a high root frequency and a low word-form frequency than it takes to recognize a word like raked which has both a low root and a low word-form frequency. On the other hand, some studies have found word-form frequency effects in LDEs, even for what looks like regularly inflected words (Baayen et al. 1997a, 1997b), suggesting that regularly inflected forms are (sometimes) stored in memory. It is, however, not entirely clear whether the inflectional phenomena tested in these studies (Dutch and Italian plural formation) involve regular default processes; see Clahsen et al. (1997) for some discussion.

Given the dual-mechanism analysis of inflection, word-form frequency effects should be more likely to occur for irregularly inflected forms, than for regulars. This is because irregular forms are based on lexical entries, whereas affixation-based forms are typically computed on-line. I will report results from visual LDEs on German participle and noun plural formation in which this prediction was tested (Clahsen et al. 1997).

In the participle LDE, we compared 20 -n participles of different subclasses of strong verbs with 20 -t participles of weak verbs. The -t and -n participles were divided into two subgroups each according to the (word-form) frequencies of the participle forms in the CELEX database (Baayen et al. 1993), a high-frequency group (with a mean participle frequency of app. 60/Million) and a low-frequency group (with a mean participle frequency of 13/Mio.). The items were also matched for stem frequency, and this was held constant across the high- and low-frequency conditions, in order to isolate word-form frequency as the decisive factor (see Clahsen et al. 1997, Fig.3).

We found word-form frequency effects for participles of different subclasses of strong verbs, but not for -t participles of weak verbs. Fig.3 presents the mean lexical decision times for -n participles of verbs of the strong-3 subclass (see Appendix (d)) in comparison to -t participles of weak verbs:

Figure 3: Lexical decision times for participles. High-frequency irregular (-n) participles produced shorter mean reaction times than low-frequency -n participles. For regular (-t) participle forms there is no such frequency effect.

Fig.3 shows that lexical decision times for high frequency -n participles are significantly shorter than those for low frequency -n participles (p < .001). Participles of weak verbs, by contrast, do not produce a frequency effect in the lexical decision task.

In the plural study, we compared lexical decision times of -er plurals with those of -s plurals. Similarly to -s plurals, -er plurals are relatively infrequent in German (see Appendix Tab.4), but the linguistic structures involved are claimed to differ: -s plurals are based on affixation, and -er plurals have lexical entries. Thus, we would expect that lexical decision times for -er plurals (but not for -s plurals) are affected by the frequency of the plural form. Subjects were tested in a visual LDE on 20 -er and 20 -s plurals each divided into two subgroups according to their word-form frequencies in the CELEX database.

We found a strong frequency effect for -er plurals: lexical decision times for high frequency plural forms are 84 msec. shorter than those of low frequency -er plurals, (mean RTs: 671 vs. 587msec., p < .001). Regular -s plurals, however, produce similar lexical decision times irrespective of whether they are low or high in frequency, mean RTs 650 vs. 654msec., p = .62). These results show that -er, but not -s plural forms, are affected by their word-form frequencies.

The results of the two LDEs can be explained in terms of the different linguistic representations involved: -er plurals and -n participles have lexical entries, and hence high-frequency forms can be accessed more quickly than low-frequency ones. For -t participles and -s plurals, however, there are no full-form representations, as these are derived from affixation. Stems and affixes may have separate entries in the mental lexicon, and the speed with which they can be accessed might be affected by frequency. Affixation-based words as a whole, however, are not directly mapped onto corresponding entries, and hence the lack of a frequency effect for -s plural and -t participle forms in the two LDEs.

In priming experiments, two stimuli are presented to subjects, and the researcher manipulates the relation between them. Most research has been done on semantic and phonological priming to investigate the relationships among the meanings and the sound patterns of words in the mental lexicon (see Balota 1994, Lively et al. 1994 for literature reviews).

Morphological priming effects were found in several studies investigating the English past tense. In their seminal study, Stanners et al. (1979) found that when subjects were presented with a sequence of two identical words for lexical decision, e.g. walk followed by walk, the response times to the second occurrence of walk were usually faster than to its first occurrence. The repetition of a word is taken to facilitate access to its lexical entry. Interestingly, Stanners et al. found the same facilitation for earlier presentation of a regularly inflected past tense form (walked --> walk), i.e. this condition produced the same amount of priming as prior presentation of the stem itself.

All subsequent experiments have confirmed Stanners et al.'s finding of full priming for regularly inflected words (Kempley & Morton 1982, Napps 1989, Fowler et al. 1985, Marslen-Wilson et al. 1993). However, with respect to irregulars, the results are much less clear. Where Stanners et al. had seen reduced facilitation for irregulars, Kempley & Morton (1982) found no priming at all, and Fowler et al. (1985) as well as Forster et al. (1987) found full priming. Marslen-Wilson et al. (1993) investigated two subclasses of irregular past tense forms, (i) verbs such as burnt-burn and felt-feel with vowel changes and -t as the final consonant, and (ii) verbs such as sang - sing, gave - give with vowel changes only. They compared these irregular types with regular past tense forms in a cross-modal priming task. Only the regular past tense forms produced full priming. The past tense forms of semi-regular verbs (burn -burnt, feel - felt) yielded no priming, whereas those of pure vowel-change verbs such as give-gave actually led to an interference effect, with response times being significantly slower than following unrelated primes. The interference effect might be taken to reflect the presence of two lexical representations for give and gave, for example, which inhibit one another.

Taken together, the finding that regular past tense forms consistently produced full priming in all studies is compatible with the view that regular past tense forms are morphologically decomposed. Thus, the segmentation of walked into a stem (walk) plus affix (-ed) leads to activation of a lexical entry for the stem (walk) which serves as a prime for the target stem (walk) in the lexical decision task; hence the full priming for regulars. There is, however, an alternative interpretation as to why regular past tense forms in English produce full priming. Rueckl et al. (1997) argued that regular past tense forms are orthographically and phonologically more similar to their base forms than are irregular past tense forms; see for example, walked --> walk versus taught --> teach, and it might be these different form properties that account for full priming of regular past tense forms. This interpretation would make full priming effects for -ed forms compatible with single-mechanism associative models of the mental lexicon. Thus, it is not entirely clear how full priming for regular past tense forms in English should be interpreted. In addition to that, there remains the question of why irregulars have yielded such inconsistent results across studies.

Again, the confounding variables of the English past tense system make it hard to interpret experimental findings. As shown below, German inflection allows us to construct priming experiments without such confounding variables. In two experiments, we investigated German participle and noun plural inflections for morphological priming effects (Sonnenstuhl et al. 1998) by adopting the cross-modal immediate repetition priming technique from Marslen-Wilson et al. (1993, 1994, 1997). In this task, subjects hear a spoken prime (such as Ponys `ponies' or gekauft `bought') immediately followed by a visually presented form of the noun (e.g. Pony) or verb (kaufe `buy-1^stsg.') to which subjects have to make a lexical decision. Because the task is cross-modal, any priming effects from these experiments are attributable to the lexical representations themselves, rather than to effects of modality-specific access procedures.

Full and partial priming effects were determined by comparing RTs in three conditions; see Tab.1 and Tab.2 for example stimulus sets:

I. Identical primes and targets

II. Morphologically related primes and targets

III. Different primes and targets

In the control condition (III), primes and targets are neither semantically nor phonologically related, whereas in condition (I) they are identical. These two conditions provide the baseline lexical decision times (with no priming for III and full priming for I) to which the RTs of the experimental condition (II) can be compared. According to the dual-mechanism approach, we would only expect full priming, i.e. (I=II) < III, for -t participles and -s plurals, but not for -n participles and irregular plurals. The reason is that -t participles and -s plurals are decomposed into stem+affix, which both have a lexical representation of their own; hence the repetition of the same stem in conditions I and II should facilitate accessing the corresponding lexical entries. Irregular plural forms and -n participles, however, are represented on subnodes of (structured) lexical entries. This means that irregular plural and participle forms activate their corresponding base entries only indirectly, via subtrees (see (i), (ii) in section 3 and (ii), (iv) in Appendix); they should therefore produce less facilitation than identical primes (I < II < III). In other words, we predict full priming for -t participles and -s plurals and partial priming for the irregulars.

We compared priming effects of participle forms of regular verbs with those of irregular participles of the `strong-3' class (see (d) in Appendix). Note that all verbs that belong to this class, e.g. laufen - lief - gelaufen `run- ran -run', have stem changes in the past tense form, but not in the participle. This is unlike in English past tense forms where stem changes are a confounding variable for potential priming differences between regular and irregular inflection. This problem does not occur in the German study. We constructed 21 sets of triplets for each participle type (-t vs. -n), where the 1^st sg. present tense forms were always the target, presented visually for lexical decision, at the offset of one of three kinds of auditory primes, `Identical' (I), `Morphologically Related' (II) or `Different' (III). The control items for condition (III) were matched to the primes I and II for frequency and syllable length (see Tab.1 for an example stimulus set), and weak and strong-3 verbs were matched with respect to their lemma frequencies.

Tab.1: Example Stimulus Set - Participles

Three experimental versions were constructed, so that each target word occurred only once for each subject. 66 subjects were tested, 22 for each version. In addition to the experimental and control items, we included a further 534 pairs of filler items, to make sure that the target was a nonword in half of the trials and that the proportion of related items is lower than 15% overall (see Sonnenstuhl et al. 1998 for details).

The results are shown in Fig.4 for two types of verbs, weak verbs that take the -t participle suffix and strong-3 verbs with the participle -n. For both types of verbs, the `Identity' condition produces shorter RTs than the two control conditions. With respect to the `Morphological Prime' condition, however, there are pronounced differences between -t and -n participles. Regular (-t) participles produce significant morphological priming (578 vs. 605 msec., p < .001), and the size of the effect (= 27 msec.) is similar to the size of the effect in the Identity condition (= 30 msec.). RTs after presentation of -t participles do not significantly differ from those of the Identity condition (575 vs. 578 msec., p = .67). Irregular participles exhibit a different pattern of results. As shown in Fig.4b, prior presentation of an -n participle form produces shorter RTs than prior presentation of an entirely different verb (582 vs. 601 msec.), but in contrast to -t participles, the effect for -n participles is non-significant (p = .17) and RTs in the Morphological Prime condition are significantly longer than in the Identity condition (582 vs. 548 msec., p < 0.0001). These results confirm the prediction that only regular -t participles produce full priming effects; irregular -n participles, on the other hand, only lead to partial priming.

Figure 4: Cross-modal lexical priming of participles. Figure 4a presents mean lexical decision times on visual targets for -t participles, identical repetitions and unrelated controls. The morphological prime (= regular -t participles) produced a full priming effect compared to unrelated controls which is similar to an identical repetition. Figure 4b presents mean lexical decision times for irregular -n participles, identical repetitions and unrelated controls. In this case, the morphological prime, i.e. the irregular -n participles, produced a partial priming effect, i.e. longer reaction times than an identical repetition.

In the second cross-modal priming experiment, we compared -s plurals with -er plurals. Even though these two plural forms have similar (i.e. low) type frequencies in German (see Appendix, Tab.5), their lexical representations are claimed to be radically different: -s plurals are affixation-based, -er plurals are part of lexical entries. Hence, -s plurals should produce full and -er plurals partial priming effects. The design was the same as in the participle experiment. 30 sets of triplets were constructed for each plural type (-s vs. -er). Three kinds of auditory primes, `Identical', `Morphologically Related' and `Different' were used; primes were matched to control items for frequency and syllable length [9]. Uninflected singular forms presented visually for lexical decision were used as targets [9]; see Tab.2 for an example stimulus set. In addition to the experimental and control items, we included a further 220 pairs of word-word filler items and another 280 pairs of word-nonword items. There were three experimental versions, 22 subjects were tested for each version. The results are shown in Fig.5a and Fig.5b.

Tab.2: Example Stimulus Set - Plurals

Figure 5: Cross-modal lexical priming of plurals. Figure 5a shows mean lexical decision times on visual targets for -s plurals, identity primes and unrelated controls. Regular -s plurals produced a full priming effect. Figure 5b shows mean lexical decision times for -er plurals, identity primes and unrelated controls. Irregular plurals produced a longer reaction times than the identity primes.

Again, as in the participle experiment, regular and irregular inflections behave differently with respect to morphological priming. For -s plurals, there is strong morphological priming compared to the control condition (565 vs. 645 msec., p < .0001), and the effect does not significantly differ from the Identity condition (565 vs. 558 msec., p = .34). For -er plurals, however, the priming effect is much weaker; RTs are shorter than in the control condition (550 vs. 568 msec., p < .01), but significantly longer than in the Identity condition (568 vs. 531 msec., p < .01). Finally, the overall lexical decision times for nouns that take -s plurals are considerably longer for each of three conditions than the corresponding RTs for nouns that take -er plurals. These differences are due to the frequency differences between these two classes of nouns (see footnote 8).

Taken together, the results of our cross-modal priming experiments are parallel for participles and noun plurals. In both experiments, regular inflection (-s plurals and -t participles) exhibt full priming effects whereas irregularly inflected word forms (-er plurals and -n participles) only produced partial priming. These findings correspond to the linguistic differences between regular and irregular inflection posited by dual-mechanism morphology: -s plurals and -t participles are decomposed into stem+affix and therefore priming towards other corresponding word-forms can be directly mediated via the stem. Irregular plurals and participles, however, are represented on subnodes of lexical entries, and from there they can less directly prime their corresponding base forms than -s plural and -t participle forms.

Recently, some researchers have begun to investigate the brain structures involved in morphological processing by using modern neuro-imaging techniques (Jaeger et al. 1996, Ullman et al. 1997). In these studies, different areas of the brain have been shown to be active for regular and irregular inflection, but even though the results show some overlap they do not yet provide us with a coherent picture of brain activations in processing inflected words. Moreover since the techniques used (functional magnetic resonance imaging, and positron emission tomography) provide high spatial resolution but relatively low temporal resolution, they do not directly tap into the rapid processes which are involved in inflecting words.

In our research on German inflection, my colleagues and I have applied the ERP method to investigate the brain structures that are involved in morphological processing and representation. In ERP-studies, the electrical activity produced by the neurons in the brain is recorded from various points on the scalp while the subject is performing some task, such as reading a sentence. The major goal of ERP studies is to isolate those electrical components that are correlated with a particular task or stimulus. In addition to this, an ERP-study provides an on-line measure of language processing in real time. Previous ERP-studies have led to the identification of components involved in semantic and syntactic processing; see Osterhout & Holcomb (1995) and Kutas & Van Petten (1994) for overviews.

One of the most promising experimental designs for ERP-research on language is the so-called violation paradigm. Kutas and Hillyard (1980) used this paradigm in their seminal study on the N400. They found that this ERP component was elicited by semantic violations, e.g. a semantically inappropriate sentence-final word. Subsequent research has shown that the violation paradigm has proven to be extremely reliable in evoking N400 responses (see Kutas & Van Petten 1994). The violation paradigm has also been applied to study ERP components in morpho-syntactic phenomena such as violations of case, number and tense in different languages, English (Kutas & Hilyard 1983, Osterhout & Holcomb 1992, among others), Dutch (Hagoort & Brown 1994), Spanish (Kutas & Kluender 1994), German (Friederici et al. 1996, Münte et al. 1998b), and Turkish (Münte et al. 1995). Two ERP effects have been observed to these kinds of violations: a positivity with a latency of about 600 ms variably called P600, and a left anterior frontal negativity called LAN. While the exact status of the two effects is still controversial, it is safe to say that morpho-syntactic violations reliably elicit similar brain responses across different languages. In our research, we have adopted the violation paradigm to examine ERP components in morphological processing and representation. From a dual-mechanism perspective, one would expect to find different ERP effects for regular and irregular inflection. Given previous ERP research, we can make the more specific prediction that violations of regular affixation should produce similar ERP effects as other morpho-syntactic violations, i.e. a P600 and/or a left anterior negativity (LAN); misapplications of irregular inflectional patterns, however, should not elicit such components. We tested these predictions in two ERP studies, one on noun plurals and one on participles.

In the study on noun plurals (Weyerts et al. 1997), ERPs were recorded as German-speaking subjects read sentences containing nouns with correct and incorrect plural forms. Four groups of stimuli were constructed (see example stimulus set in Tab.3): nouns that normally take -n as plural marker together with the correct ending as well as with the incorrect -s plural marker, the latter resulting in a 'regularization' of an actual irregular plural (*Muskels), and nouns which normally have -s plurals with the correct ending as well as with the incorrect 'irregularized' -n ending (*Karussellen).

Tab.3: Example Stimulus Set

The electrical activity was recorded from 19 standard points on the scalp while subjects were silently reading these sentences. The ERP-recordings produced 19 sets of graphs such as those shown in Fig.6 below. The figures show the electrical activity recorded during the entire recording period, starting 100 msec before stimulus presentation and ending after 800 msec. The electrical activity is measured in microvolts. The plus and minus signs beside the axes indicate the difference of the measured voltages to the baseline condition. The baseline condition is the mean electrical activity measured 100 msec. prior to the presentation of the stimulus word; `+' means that the measured voltages are positive with respect to the baseline and `-' means that they are negative with respect to the baseline. The ERP-effects to be reported were evident from several different electrodes. For ease of exposition, however, I will only present the recordings of the electrodes F7 and Cz here; the electrode F7 is placed over the left frontal part of the brain (`Broca's area') and Cz at the central site; see Weyerts et al. (1997) for the results of the full electrode set. Fig.6 presents grand average ERPs for masculine/neuter nouns (e.g. der Muskel, see Tab.3) that take -n plurals in German as opposed to nouns that take -s plurals (see `regular' in Fig.6).

Figure 6: Event-related potentials of noun plurals. Comparison of ERPs for masculine/neuter nouns that take (irregular) -n plurals and for nouns that take regular -s plurals. Left column: Correctness effect at the left frontotemporal (F7) site for the 2 plural conditions. A significant negativity for incorrect plurals is present only for masculine/neuter. Right column: Correctness effect at the central (Cz) site for the 2 plural conditions. A phasic and earlier negativity for incorrect plurals is present only for nouns taking the regular -s ending as correct plural marker. Horizontal scale: tickmarks at 100 to 800 ms, vertical line at 0 ms (stimulus-onset): -3 to +3 µV.

These graphs show that 200 - 400 msec. after the presentation of the plural forms, the waveforms for incorrect plurals started to differ from the correct ones: from that point onwards incorrect noun plurals were associated with a more negative waveform. Interestingly, however, the scalp distribution of these negativities for the two kinds of morphological violations was rather different: For -s plural regularizations (= `incorrect masc./neuter' in Fig.6), a focal left frontotemporal effect was seen with its maximum at the F7-site, whereas the differences at Cz are non-significant in this condition. In contrast to that, irregularizations (= `incorrect regular' in Fig.6) produced an N400, i.e. a central phasic negativity with its maximum at Cz; for this condition, the differences at F7 are non-significant. Thus, this experiment has produced different ERP-responses to violations of regular and irregular inflection.

The same ERP design was used for the participle study (Penke et al. 1997). Correct regular and irregular participles were compared with incorrect ones; the latter had -(e)n on verbs that actually take -t participles (*getanz-en `dance-n'), or -(e)t on verbs that require -(e)n (*gelad-et `load-ed'). The critical words were presented in three different versions to three different groups of subjects, as part of a simple sentence, in a word list, and embedded in a story; for each version separate ERPs were recorded.

The ERP responses were very consistent across the three versions of the experiment: incorrect irregular participles (*gelad-et) elicited a left frontotemporal negativity and incorrect regulars (*getanz-en) produced no differences to the correct ones. The results are illustrated in Fig.7; here we see waveforms for regular and irregular participles at the F7 site. Fig.7 shows that regularizations, i.e. irregular verbs incorrectly inflected with the -t participle ending, are associated with a left anterior temporal negativity in all three experimental versions [10].

Figure 7: Event-related potentials of participles. Comparison of regular and irregular participles according to ending (-t versus -n) for the F7 site presented in three experimental versions, as part of a simple sentence, in a word list, and embedded in a story. Only the incorrect irregulars (those with the -t ending) are associated with a left anterior temporal negativity in all three experimental versions. Horizontal scale: tickmarks at 400 and 800 ms, vertical line at 0 ms (stimulus-onset): -4 to +4 µV.

The most salient and consistent ERP result is that regularizations were associated with a negative waveform with a focal left anterior temporal distribution starting at about 200 ms, an effect that has been discovered in previous ERP-studies and that has been called LAN. LAN-effects have been reported to occur when affixation, e.g. for subject-verb agreement is incorrectly applied (Münte et al. 1993, Osterhout & Mobley 1995), when phrase-structure rules are violated (Neville et al. 1991, Friederici & Mecklinger 1996), and in case of illegal filler-gap constructions (Kutas & Kluender 1994, Neville et al. 1991, but see McKinnon & Osterhout 1996). We found the LAN for incorrect irregulars in each of the three participle experiments and even more strikingly in a different inflectional system, noun plurals. In linguistic terms, regularizations are violations of affixation, i.e. misapplications of the participle -t or the plural -s to (irregular) verbs or nouns that would normally block the affixation process, to produce illegal stem+affix combinations such as *gelad-et and *Muskel-s. A LAN was only found in such cases. Thus, the LAN found under these conditions can be interpreted as reflecting processes involved in morphological structure building [11].

Irregularizations do not involve violations of affixation, but may rather be conceived of as unexpected or anomolous words. The ERP-results support this interpretation as the central negativity we found for (plural) irregularizations does indeed resemble the so-called N400 effect which occurs (among other cases) in response to pronounceable non-words (Kutas & Hillyard 1980, Rugg 1987).

Hence, our results indicate that the brain responds differently to violations of regular and irregular inflectional processes. Regularizations elicited signals that are typical of morpho-syntactic rule violations, whereas irregularizations produced waveforms that are typical of the reaction to anomalous words. These differences correspond to the linguistic distinction between affixation-based and lexically-based inflection.

In this section, we will discuss associative single-mechanism models of German inflection. Three recent studies have proposed connectionist implementations of German plurals and participles, Nakisa et al. (1998), Westermann & Goebel (1995) and Goebel & Indefrey (1998). In addition, Bybee (1995) and Köpcke (1988, 1993) have suggested ways in which associative schemas might handle the data on German inflection.

Nakisa et al. (1998) have made an attempt to implement the German plural system in single-mechanism pattern associator networks of different kinds. They took 8,598 plural forms from the CELEX database, split them roughly in half, and used one half as a training set and the other as a testing set. For the latter, their pattern associators produced between 70% and 82% correct plurals depending on the learning algorithm and network architecture chosen. Nakisa and collaborators consider this success rate to be `remarkably high'. In addition to pure similarity-based simulations, they also carried out a simulation in which the plural -s was removed from the pattern associator and was only applied to singular items that were phonologically distant from other singular items in their sample. This was meant to be an implementation of a dual-mechanism model and the special status of -s in the German plural system. Nakisa et al. found that this simulation failed to perform any better than their single-mechanism pattern associators. They concluded that the mental process of regular inflection can be simulated in a single-mechanism associative network by exploiting the statistics of the input, contra to claims made by proponents of the dual-mechanism model.

The reasoning, however, behind their conclusion is puzzling. We are led to believe that rules of inflection can be eliminated from the mental grammar, because after extensive training a pattern associator gets a majority of the existing German plural forms right. That the model does well is not in fact surprising, given that more than 90% of the existing plural forms of German are irregular ones (see Appendix, Tab.5). One might even ask why the network produces relatively many incorrect plurals, with 20% to 30% errors certainly not a negligable quantity. The fact that the simulation in which the -s plural was hard-wired to apply to phonologically distant nouns did not outperform the pure pattern associators does not come as a surprise, since it is based on an assumption which is decriptively inadequate for German, namely that all existing nouns that take -s plurals are phonologically distant from all other nouns. Moreover, for this simulation Nakisa et al. removed all nouns that take -s plurals from the training set, thereby creating unrealistic training materials which consisted of irregular plurals only. The crucial question then is whether any of the single-mechanism models they proposed can simulate the differences between regular and irregular plurals. Consider, for example, the generalization properties of -s plurals reported in section 4.2. We found that the plural -s applies under default circumstances, i.e. when access to lexical memory is ruled out, i.e. in cases, such as nonsense words that are dissimilar to existing words, derived forms, and words that do not have canonical lexical entries. However, in contrast to what German speakers do, Nakisa et al.'s pattern associators would extend -n and, to a lesser extent, -e plurals under such circumstances, simply because these happen to be the most frequent plural forms in the training set (R. Nakisa, personal communication). Moreover, their simulations do not explain the experimental findings reported in previous sections, e.g. the fact that irregular plurals, but not -s plurals, produced a word-form frequency effect in lexical decision. In sum, Nakisa et al.'s simulations do not present us with a descriptively adequate single-mechanism model of the linguistic and psycholinguistic properties of German noun plurals.

More promising seem to be implementations of German inflection in modular connectionist networks as suggested by Goebel and collaborators. Westermann & Goebel (1995) present a network for participle formation which consists of two separate components, a combinatorial system capable of representing sequences of feature vectors and an associative phonological lexicon. They stress that the architecture of their network is in accordance with the dual-mechanism model of inflection. The network was trained on a corpus of 538 verb tokens using a backpropagation algorithm, and it is shown that the network's generalizations of -n and -t participle formation are similar to those we found in experiments with German speakers: extensions of -n participle formation are similarity-based, whereas -t participle formation is generalized under no-similarity conditions, i.e. to items which do not resemble any existing verbs.

Goebel & Indefrey (1998) made use of the same kind of network architecture to model the German plural system. Their training corpus consisted of 6,364 noun types taken from the vocabulary of 5 to 6 year old children. After having achieved good performance for the items of the training set, the generalization properties of the network were tested on the set of experimental items used in our plural judgement study (Marcus et al. 1995). Goebel & Indefrey's model was partially successful, particularly with respect to irregular plural forms. The model extended -e plurals to novel items with masculine gender and generalized -n plurals to novel feminines, both reflecting strong tendencies among the irregular plural forms of German. The generalization properties of -s plurals, however, were not properly captured by their network. The model applied -s to novel nouns that end in the full vowel -O, i.e. to items that are similar to existing -s plurals. The crucial point, however, is what the model does to items that are dissimilar to existing nouns. Recall that speakers of German prefer -s pluralizations in such cases (Marcus et al. 1995). Goebel & Indefrey's connectionist network, by contrast, failed to generalize -s under no-similarity conditions. Thus, this model does not capture the generalization properties of -s plurals. Goebel & Indefrey speculate that it is the extremely low frequency of -s plurals combined with their default properties which make it impossible for the network to generalize -s in the way speakers of German do.

Schema-based approaches to inflection, such as Bybee (1991, 1995) and Köpcke (1988, 1993), share with connectionist approaches the view that `all types of morphological patterns can be acquired by the same process - the storage of items, the creation of connections among them, and the formation of patterns that range over sets of connections' (Bybee 1991:87). It is also argued that potential differences between morphological schemas result from their frequency distribution - a big class is more productive and forms a stronger schema than a small class. Bybee (1995) analyzed German participle formation from this perspective. She argued that -t suffixation applies to the largest number of verb stems and that when subjects are required to inflect nonsense words, they prefer to follow the majority pattern yielding overapplications of the participle -t. For German noun plurals, Köpcke (1988, 1993) presented an analysis in terms of associative schemas. The schemas are envisioned as capturing the similarities among the existing nouns that take a plural suffix; nothing has a special status as the default. To support this account, Köpcke presented results from an elicited-production task in which German adults had to pluralize novel noun stems which varied by gender (masculine, feminine, and neuter) and syllabic structure (suffixed, schwa-final, vowel-final, pseudosuffixed, or monosyllabic). He found that each type of word has a preferred suffix. For instance, feminine schwa-final nouns nearly always took -(e)n, and monosyllabic nouns tended to take -e if neuter or masculine, -(e)n if feminine. This is presented as evidence for the view that the German plural system can best be handled in terms of associative schemas.

Schema-based models of German inflection are confronted with the same problem as connectionist pattern associators in that they do not account for the linguistic and experimental differences between regular and irregular inflection. In his experiment on noun plural formation, for example, Köpcke did not test whether there is a default pluralization process in German. He presented each noun in isolation which made it possible for his subjects to form an analogy to an existing noun and its plural form. The plural judgement task reported in section 4.2 provides a way of ruling out such unwanted associations. The results show that in cases in which memory access is made impossible, speakers of German rely on the plural -s, indicating that (in contrast to what Köpcke and Bybee argued) the plural -s is a regular default in German.

Moreover, the default nature of regular affixes does not necessarily follow from their frequency distribution. This is obvious for the plural -s which applies to only 7% of nouns in German (see Tab.5 in the Appendix), whereas, for example in English the plural -s is applied to more than 99% of all nouns. For verbs, there is also a clear difference between English and German in terms of the vocabulary distribution of regulars and irregulars, even though the difference is less dramatic than for plurals. Three different frequency measures revealed that (in contrast to English), regular and irregular verb forms have similar frequencies (see Appendix). With larger samples of verbs, the gap narrows, but regular -t participle forms in German always show lower frequencies than corresponding -ed past tense forms in English. This leaves schema-based theories with no account for the fact that despite differences in frequency distribution, the generalization properties of the German -s plural and the -t participle are similar to the English -s plural and the past tense -ed.

The results from our lexical decision experiments reported in section 4.3 are also hard to explain in terms of schema-based models of inflection. We found that lexical decision times were shorter for high-frequency forms than for low-frequency forms, but this effect only occurred for irregular plurals and irregular participles, not for regular ones. Conversely, full priming effects were found for regular plurals and participles, but not for irregular ones. In the ERP experiments (section 4.5) and the sentence-matching experiments (section 4.2), violation effects were found -t participle and -s plural affixation, but not for irregular inflection. These experimental effects cannot be derived from associative schemas such as those proposed by Köpcke and Bybee for German inflection.

Summarizing, despite several attempts there is still no single-mechanism associative model that can handle the full set of facts of German plurals and participles. As the regular plural and participle affixes (-s and -t) do not generalize on the basis of frequency or similarity, it is hard to imagine how any conventional single-mechanism pattern associator could ever get the facts of German inflection right.

The findings reported above demonstrate clear processing differences between -t participle and -s plural formation on the one hand and -n participle and irregular noun plurals on the other hand. Tab. 4 summarizes the main results:

Tab.4: Summary of processing properties of plurals and participles

These findings receive, at most, a partial explanation from single-mechanism accounts of inflection in which no distinction is made between affixation-based and lexicon-based inflection. If all inflected words were stored in memory, as assumed in full-listing models of morphological processing, we would not expect to find processing differences between regular and irregular inflection; but such effects exist, as shown in previous sections, and they are left unexplained by these kinds of models. We also argued that associative networks do not properly capture the generalization properties of German inflection. In full-parsing models, on the other hand, all inflected words are assumed to be morphologically decomposed into stem+affix. If this is correct, we would expect to find rule-violation effects in the sentence-matching task and the ERP-studies for inflected words in general; legal stem+affix combinations, for example, should elicit shorter sentence-matching latencies than illegal ones. Our findings show that this is true for regular inflection, but not for irregularly inflected word-forms. This difference cannot be explained by the full-parsing model.

Two theoretical assumptions, I suggest, have to be made in order to explain the set of results in Tab.4. First, the processor may choose between accessing an inflected word from the mental lexicon or decomposing it into stem+affix. Second, these two processing routes correspond to the morphological structure of inflected words: those which are represented in terms of lexical entries are processed by accessing full-form representations, whereas affixation-based inflected words are processed by decomposing them into stem+affix combinations. The empirical results show that the structural properties and the processing properties of German participles and noun plurals converge into either the regular or the irregular cluster, thus providing support for the dual-mechanism model.

The fact that German (and English) exhibit linguistic and experimental differences between regular and irregular inflection raises the question of how general this distinction holds across languages. In two recent studies, we have investigated Italian past-tense and participle inflections with respect to this question (Say 1998, Gross et al. 1998; see also Orsolini and Marslen-Wilson 1997). In contrast to Germanic languages, Italian verbs fall into three basic morphological classes, called conjugations, and it is not clear how a dual-mechanism model might account for a three-way inflectional system. To examine the generalization properties of the three conjugation classes in Italian, Say (1998) performed an elicited production experiment, adopting the design of the elicitation task on German participles reported in section 4.2. Say found that 1st conjugation stem forms generalize freely to all types of novel verbs, regardless of phonological content, whereas 3rd and 2nd conjugation stems generalize only to those novel verbs that were similar to existing verbs of these types. Moreover, a frequency effect was found for 3rd and 2nd conjugation verbs, but not for 1st conjugation verbs: the former generalize more easily to novel verbs that were similar to existing high-frequency verbs than to those based on low-frequency existing verbs.

Thus, only the 1st conjugation stem formation process showed characteristics of a default; the 2nd and 3rd conjugation stem forms, on the other hand, showed non-default behaviour, namely frequency and similarity effects in their generalization properties. In addition to stem formation processes, we examined the role of affixation processes in Italian verb forms using the ERP violation paradigm (Gross et al. 1998); the design was adopted from our studies on German inflection (see section 4.5). We found that incorrect irregulars resulting from overapplications of regular inflection, e.g. *prendato `*tooked' instead of the correct preso `taken', elicited a widespread negativity. By contrast, incorrect theme vowels, e.g. *parlito instead of parlato `spoken' and *dormato instead of dormito `slept', produced no such effect. This difference replicates previous ERP results on German participle and plural inflection.

Further experimentation is required before any strong cross-linguistic claims on the processing of inflection can be made. This caveat notwithstanding, we think that the results mentioned above do indeed suggest that across different types of languages the mind/brain honours the regular/irregular distinction posited by dual-mechanism models of inflection, through e.g. different generalization properties and different brain potentials for the two morphological clusters.

The idea that the language faculty of an adult speaker consists of two qualitatively different components, a system of combinatorial operations and a lexicon, raises the question of how this dual structure emerges in child language development. One approach to address this question is in the spirit of Piagetian constructivism; children may construct mental representations for language on the basis of their linguistic experience and general-purpose learning mechanisms (Bates et al. 1988, Karmiloff-Smith 1992, Elman et al. 1996). However, even though much research has been devoted to this approach, it is still a puzzle how general mechanisms plus input should yield successful learning of the essentials of the adult grammar such as its dual architecture, its combinatorial mapping between semantics and linguistic forms, abstract morpho-syntactic representations, its symbolic categories and features (e.g. `V(erb)', `N(oun)), etc. One potential solution to this puzzle might be to try and eliminate such apparently unlearnable notions from representations of adult grammars and to implement an adult's knowledge of language in connectionist networks which do not contain grammatical categories and operations of the familiar kind (Elman et al. 1996, Rumelhart & McClelland 1986). The reasoning would make perfect sense: if there were no such things as a dual architecture, a combinatorial rules, `Ns' and `Vs' etc. in the adult grammar, then we would not have to worry about how children learn them. But Pinker & Prince (1988) and Marcus (1998) have shown that these kinds of connectionist networks have severe deficiencies, and argued that eliminative connectionism must be rejected; as Marcus (1998) puts it, `connectionism cannot save constructivism'. Moreover, as pointed out in the previous section, research on language and language processing in adults has produced substantial evidence in favour of the dual-mechanism view and related notions that cannot be handled by single-mechanism networks (see section 4.6). These findings would be left unexplained if grammatical categories and operations were eliminated from mental representations. This leads us back to the original question: how does the adult linguistic system emerge?

An alternative way of addressing this question is in terms of the `continuity hypothesis' which claims that the structure of the language faculty does not change over time but that development results from other factors (Pinker 1984, Hyams 1986, Weissenborn 1992, among others). The specific idea my colleagues and I have pursued in our language acquisition studies is that grammatical development may result from increases in the child's lexicon, i.e., from the set of lexical and morphological items which the child has acquired (Clahsen 1990, 1992, Clahsen et al. 1993, 1994, 1996a). Hence, even though the child's language faculty is said to be qualitatively identical to that of adults, the lexicon takes time to develop. Inflectional affixes, for example, are often unstressed, monosyllabic, not uttered in isolation, and perceptually nonsalient. For these reasons, the child may find it difficult to pick them up from the input, and consequently those elements of the child's language faculty which operate on inflectional affixes, e.g. principles of affixation, cannot become effective, though they are (latently) available throughout. In this way, the grammar of the particular language the child is acquiring develops gradually, through the interaction of available abstract knowledge, e.g. about combinatorial principles, and the child's learning of the lexicon. Evidence for continuity and lexical learning has mainly come from studies on the acquisition of syntax. Here, we will examine the development of inflection from this perspective.

The most detailed results on children's inflectional systems comes from studies of the English past tense (Marcus et al. 1992 and references cited there). This research has produced three major findings. First, only one type of inflectional error is produced: overapplications of the regular past tense affix -ed to irregular stems (*go-ed, etc.); overapplications of irregular patterns (*brang, *wope, *talken) are much less frequent and systematic (Xu & Pinker 1995). Second, overregularizations are rare at all ages (= typically less than 10%) and their occurrence is preceded by a stage at which all overtly marked past tense forms are correct. Third, overregularizations are sensitive to frequency and similarity: children make these errors more often for low-frequency irregular verbs, and they make fewer overregularization errors for irregular verbs that fall into families with more numerous and higher-frequency members (see also Bybee & Slobin 1982). Marcus et al. (1992) take these findings as evidence for the view that children (like adults) possess two distinct mechanisms for inflection, a symbol-manipulating rule system for regulars and lexical entries for irregulars. Only when the child fails to get access to a lexical entry for an irregular form is the regular rule applied, resulting in occasional overregularization errors. However, due to the confounding factors in the English past tense system, some researchers have argued for an alternative explanation, namely that children overregularize -ed not because -ed is rule-based, but simply because children have heard it used with so many different English verbs (Plunkett & Marchman 1991, 1993 among others). Associative models of language may indeed appear to look quite successful in this case and have been argued to accurately model English-speaking children's overregularizations of -ed [12]. But this success depends on an artifact of the English past tense system, namely the extremely high frequency of -ed forms. The two inflectional systems we have studied, German participle and noun plural formation, exhibit different vocabulary statistics (see Appendix) to the English past tense; results from these studies should help to assess the role of input frequencies for children's overregularizations.

We investigated the development of participle formation in longitudinal data from 9 children between the ages of 1;4 and 3;9 and two elicited production experiments covering the age period of 3;6 to 8;10 (Clahsen & Rothweiler 1993, Weyerts & Clahsen 1994, Weyerts 1997). We found that children across all age groups overapply the participle -t to strong verbs, but that they rarely overapply the participle -n to mixed or weak verbs. In the longitudinal data, there were 116 incorrect endings in children's participles, 93% of these were -t errors. Thus, children typically produce errors such as *gekommt 'come, participle' instead of gekommen, but not *geschneien 'snow, participle' instead of geschneit 'snowed'. Overregularizations of the participle -t are rare in the longitudinal data (mean rate = 10%), and there is an early stage prior to the occurrence of -t errors at which all participle forms used by the child are correct. Consider, for illustration, Fig.8 which presents a longitudinal graph for one child, Simone. On the x-axis, the period of observation, ranging from 1;10 to 3;9, is divided into 14 individual corpora which each covered approximately one month; the y-axis shows the rates of -t overregularizations for each age point. Fig.8 shows that the period from age 2;1 onwards (= Corpus 3) with a relatively steady -t overregularization rate of 5% to 10% is preceded by a period, Corpora 1 and 2, with no overregularizations.

Figure 8: Overregularizations of the -t participle suffix. Longitudinal data covering the age period from 1;10 to 3;9 divided into 14 corpora. Overregularization rates for -t were calculated against the total number of correctly inflected irregular participles. From 2;1 onwards there is a relatively constant overregularization rate of 5% to 10% compared to 90% to 95% correctly inflected participles of irregular verbs. This is preceded by a stage without any -t overregularizations.

The preference for -t overregularizations was also confirmed experimentally. In an elicited production experiment with 70 children (age range: 3;6 to 8;10), 152 incorrect endings on participles of existing verbs were produced, 90% of them -t errors. Among the -t overregularizations, we found a clear frequency effect: strong verbs that have low token frequencies elicit significantly more -t overregularization errors in participle formation than strong verbs with high token frequencies. This frequency effect indicates that participles of strong verbs form lexical entries, and that the participle -t applies in cases of unsuccessful lexical access. In another experiment, we elicited participles of nonsense verbs. Of the 454 forms with one of the two participle endings produced by the children, 93% had -t. Thus, children apply -t participle formation under default circumstances to verbs which do not have lexical entries.

We have examined various aspects of plural formation in German-speaking children (Clahsen et al. 1992, 1996b, Bartke et al. 1996, Bartke 1998). Here we focus on the question of whether children generalize inflectional processes in similar ways as adults. Recall from section 4.2 that for adults, -s plural affixation is applied under default circumstances and generalizes widely, e.g. to unfamiliar or unusual sounding words and to proper names. Extensions of irregular plural forms, however, are restricted to items which are similar to existing ones. To test the generalization properties of plural formation in children, we administered an elicited production experiment with 66 children (age: 3;1 to 8;11) and an acceptability judgment experiment with 37 children (age: 3;6 to 6;6), in which the circumstances under which adults generalize regular and irregular plural formation processes were controlled.

In the first experiment (Clahsen et al. 1996b), 16 objects were used denoting low-frequency nouns in German such as Feder `feather', Tuch ` towel', Fassung `socket', etc.; the plural allomorphs -er, -en, -s and -0 are required for adults in 4 cases each and the plural -e in 2 cases by these nouns. The children were asked to name each object, and were then prompted to produce plural forms by being presented with four instances of each object. We found an overall overregularization rate of 18.5% (n= 141) which is considerably higher than the rates reported for spontaneous speech and reflects the fact that some of the experimental items were unfamiliar to the children. Within the overregularizations the different affixes were used with different frequencies: 58.5% (n = 76) of the overregularizations were with -s plural; 26.2.% (n = 34) with -e, 14.6% (n = 19) with -(e)n, and only 0.8 % (n = 1) with -er. Thus, children (like to adults) are more likely to inflect unfamilar words with the -s affix than with any other plural form.

In a second experiment (Bartke et al. 1996), we examined children's acceptability judgments on plural forms of nonsense words. To control for similarity-based generalizations, two kinds of nonsense words were constructed: `Rhymes', items such as pund, kach etc. that rhyme with existing German nouns which take irregular plurals (Hund - Hunde `dog - dogs', Dach - Dächer `roof - roofs'); `Non-Rhymes', items such as pnähf, fneik, etc. which do not rhyme with existing German nouns. By using different context sentences, these items were introduced to the children in two conditions, as canonical root nouns and as proper names. Note that in linguistic terms, proper names are non-canonical words which do not have lexical entries and should therefore undergo -s affixation when pluralized. Recall from section 4.2 that adults do indeed prefer -s plural forms of proper names irrespective of the word's sound properties. In the child experiment, each item was embedded in a story and illustrated by picture cards, followed by the presentation of the two plural forms, one with -s and one with -(e)n, the latter of which is the most frequent plural form in German. Children were encouraged to provide a third, alternative plural if they found neither -s nor -(e)n acceptable. Fig.9 presents the results.

Figure 9: Children's ratings of plural forms. Preferences for -s plurals versus -(e)n plurals on rhyme and non-rhyme roots and names. For non-rhyme roots (Figure 9a), children preferred -s pluralizations to -(e)n plural forms, whereas there was no clear preference for rhyme roots. For names (Figure 9b), on the other hand, children preferred -s pluralizations in both the rhyme and the non-rhyme condition.

Fig.9a shows that children's ratings for -s plurals were much better in the Non-Rhyme condition than in the Rhyme condition, and Fig.9b shows that -s plurals of names were overall judged as significantly better than -n plurals for both rhymes and non-rhymes. Thus children (like adults) prefer -s plurals for unusual-sounding words (non-rhymes) and for non-canonical words (proper names).

In sum, our results on participle and noun plural formation indicate that regular and irregular inflectional processes are dissociated in child language. The -t participle and the -s plural affixes are supplied under default circumstances. They are sometimes overextended to words which are irregular in the adult language resulting in overregularization errors, and they occur when similarity-driven analogies fail, e.g. in the case of unfamiliar, unusual-sounding or otherwise non-canonical words. Overextensions of irregular patterns are more restricted; -n participle errors are extremely rare in German children, and irregular plural forms seem to be overapplied only by analogy. The generalization properties of regular and irregular inflection in children are identical to those found for adults (see section 4.2) indicating that the dual structure of the language faculty is already in place early on. What seems to change over time is only the child's vocabulary, i.e. the inventory of lexical entries. We have isolated an early developmental stage at which children do not produce any overapplication errors. At this stage, the inflectional affixes have not yet been isolated, and all inflected word forms (including those which are regular in the adult language) have lexical entries for children; the affixation component cannot yet operate at this stage and has to await the child's identification of suitable affixes. Once separate lexical entries for the -t participle and the -s plural suffixes have been created, the affixation component can become effective, and children apply affixation under the same circumstances as adults. It follows that the child does not have to learn inflectional rules; instead, the child creates lexical entries for affixes in much the same way in which s/he learns new words and represents them in the form of new entries; subsequently, combinatorial mechanisms such as affixation can operate. Thus, under this view the driving force in children's grammatical development is the child's gradual acquisition of new lexical and morphological items; the architecture, however, including its dual structure seems to be intrinsic to children's as well as adults' language faculty.

To further investigate whether the architecture of the child's language faculty is identical to that of the adult or whether it undergoes substantial developmental changes we have studied relationships between two different word-formation processes, compounding and inflection, in children and adults (Clahsen et al. 1992, 1996b).

In English, irregular plural nouns such as teeth, mice, etc. can appear as initial parts or non-heads in compounds, while plural nouns formed by adding the suffix -s are not used in this position, (e.g, teeth marks versus *nails marks; mice eater versus *rats eater). The same is true for German inflection. The sole plural forms that occur as non-head elements inside compounds are the irregulars, and never the plural -s. Hence the distribution of plurals-inside-compounds coincides with the distinction made in Minimalist Morphology between structured lexical entries and affixation. Lexical compounding is a process that concatenates lexical entries. It follows that irregular plurals (because they have lexical entries) can be fed into the compounding process, whereas affixation-based forms such as -s plurals (which do not have lexical entries) cannot be included in the compounding process. There are several other linguistic analyses to capture these facts (Kiparsky 1982, Di Sciullio & Williams 1987, Borer 1988, Wiese 1996). The idea that is common to these different analyses is that feeding relationships between plural inflection and compounding are determined by a grammatical ordering constraint. It is hard to see how children could learn this constraint directly from input data. From the perspective of `continuity' and `lexical learning', one would rather expect design features of the language faculty not to change over time. What changes over time is the identification and categorization of lexical entries from the input. Thus, children may differ from adults with respect to the lexical representations of plural forms, but whatever the child's default plural form is, as for example measured through children's overregularization errors, the ordering constraint should not be violated. What we predict, then, is a correlation between overregularization and plurals-inside-compounds: plural forms that are used in overregularizations should be omitted from the non-head elements of compounds. This would indicate that even though the children's plural forms might be incorrect in terms of the adult grammar, the ordering constraint on plurals-inside-compounds would be operative in the child's linguistic system.

Gordon (1985) has confirmed this prediction for child English using an elicited-production task which prompted children to first produce a plural form of a given item X (e.g. mouse --> mice) and subsequently a synthetic compound of the X-eater type (e.g. mouse-eater or mice-eater). Gordon found that in about 98% of the novel compounds produced by the children the plural -s is not used on the non-head element; this even holds for the youngest children, i.e. the 3-year olds. Overregularizations like *mouses sometimes occurred as a simple noun but never within a compound such as *mouses-eater. Furthermore, it was found that as soon as the children have acquired irregular plurals they sometimes use them inside compounds (as in mice-eater). Thus children behaved essentially like adults: they were willing to say mice-eater and unwilling to say *rats-eater. Gordon interprets these results as support for the early availability of the morphological ordering constraint in children. There is, however, an alternative explanation of Gordon's findings which is independently motivated by developmental principles of word formation suggested by Clark (1993). She found that 3;0 to 6;11 children tend to strip off all previous verb and noun affixes before they form a novel compound. Thus when asked to form a compound out of a phrase such as `a boy who bounces balls', children strip off the 3<sup>rd</sup> sg. -s as well as the plural -s to form the compound ball-bouncer. Clark (1993) argued that this reflects a general developmental strategy which she called the `simplicity of form' principle: young children use simple building blocks to form new words, i.e., they tend to use unadorned elements with no affixes in forming new words. The simplicity principle predicts that children should produce compounds constructed with bare nouns or bare noun and verb bases such as N+N compounds (rat-man), V+N compounds (twist-knob), and so on. Gordon's findings on compounding in English-speaking children might also be explained in this way. According to the simplicity principle, children strip off the regular plural -s inside compounds, but maintain the irregular plural form, simply because in irregulars there is no separate affix to be stripped off; compare, for example, rat-s vs. mice. In this way, Gordon's results could be accounted for without any reliance on abstract morphological principles such as the ordering constraint mentioned above.

This issue is hard to resolve by only looking at English, because of a confound between regular inflection and suffixation in English: regular plurals are marked with a suffix, while irregular plurals are not. We therefore need to examine a language in which both regular and irregular plurals have