Language Learning: Cues or Rules?

Child language researchers have often taken gender and case paradigms to be interesting test cases for theories of language learning. In this paper we develop a computational model of the acquisition of the gender, number, and case paradigm for the German definite article. The computational formalism used is a connectionist algorithm developed by Rumelhart, Hinton, and Williams (1986). Three models are developed. In the first two, various cues to gender studied by Köpcke and Zubin (1983, 1984) are entered by hand. In the third, the simulation is given only the raw phonological features of the stem. Despite the elimination of the hand-crafting of the units, the third model outperformed the first two in both training and generalization. All three models showed a good match to the developmental data of Mills (1986) and MacWhinney (1978). Advantages of a connectionist approach over older theories are discussed. When generative grammarians write a grammar for a language, they do it by formulating a set of rules together with a set of exceptions to these rules. In their classic work, “The Sound Pattern of English” (SPE), Chomsky and Halle (1968) provided a fairly complete generative grammar for English derivational and inflectional morphology. One of the goals of that work was to express the grammar in a maximally compact form, thereby capturing the maximum number of “linguistically significant generalizations.” As a result, the rules of SPE were highly symbolic, having numerous subconditions, alternative environments, and variables. Moreover, the forms upon which those rules operated were abstract, often with no direct relation to any actually occurring phonological form. Although modern generative phonology no longer subscribes to the formulations of SPE, it still accepts the view of grammar as a set of rules based on often highly abstract symbols. The status of rules in descriptions of children’s language has often been called into question. For example, both psycholinguists such as Ervin-Tripp (1966) and Slobin (1971) and linguists such as Hockett (1968) and Givón (1984) have pointed out that there is no direct evidence that language users actually manipulate rules and rule symbols in their heads in the same way that rules are processed in a linguist’s grammar. Nor is there direct evidence that children pick up rules during actual conversational interactions. How then is it that children come to behave as if they knew the rules, even without learning rules as discrete entities? Does the mechanism underlying this learning resemble the rule itself or does it look like something very different? Our goal here is the articulation of an approach to morphological processing and structure that does not depend on rules. This approach views morphological form as arising from a competition between a large set of phonological, syntactic, and semantic cues. The competition is expressed computationally in the form of a connectionist network (Rumelhart and McClelland, 1987), rather than through the formalisms of generative grammar. A connectionist architecture is selected as a particular computational formalization of a general model of language processing and acquisition we have called the “Competition Model” (MacWhinney and Bates, in press). As a testing ground for the claims of the Competition Model, we have chosen to model the learning of the German declensional paradigm. Because of the complexities of gender, number, and case assignment in German, child language researchers have often (Braine, 1987, MacWhinney, 1978; Maratsos and Chalkley, 1980; Maratsos, 1982 and Pinker, 1984) viewed the acquisition of German declensions as a challenge to language learning theory. Rumelhart and McClelland (1987) used neural networks to model the ways in which English-speaking children learn to form the past tense of the verb. The Rumelhart and McClelland verb-learning model was able to produce overextensions such as “drawed,” as well as correct formations such as “fell” and “ran.” However, the target structures in English are fairly simple. There are only a few dozen irregular verbs and the paradigm only involves three forms the present, the past, and the perfect. In German, there is no regular or default pattern and the paradigm has 16 cells for the article alone and separate paradigms for the adjective and the noun. Moreover, whereas English appears to have some simple phonological cues that predict irregular past tense forms (Bybee and Slobin, 1982), it is generally believed that the gender of German nouns is entirely arbitrary (Maratsos and Chalkley, 1980) and cannot be predicted by any set of cues or combination of cues. Both generativists and connectionists agree that grammar must be learnable. In the generative framework, one first formulates a descriptively adequate grammar such as that of SPE. Next, one attempts to formulate a learning model that guarantees that the target structures are learnable. In constructing a learning account, the generativist is often forced to attribute to the learner a fair amount of abstract innate knowledge (Wexler and Culicover, 1980; Pinker, 1984). The postulation of innate structures is often taken to be a goal in its own right and is even used as proof of the correctness of the structural analysis. The connectionist framework we will adopt takes a different approach to learning. In that framework, learning and processing are treated directly in the same computational architecture. As the network learns, its processing abilities develop. There is no separation between learning, structure, and processing. The emphasis is upon maximizing the contribution of the learning algorithm and minimizing the recourse to innate abilities. Our goal in formulating a process model for the acquisition of German declension is not simply to raise questions regarding the value of the generative approach. We are interested in constructing a concrete cue-based alternative to that approach. Of course, we want an approach that will learn the target structures. The real test of this alternative will be its ability to simulate German children’s language learning (MacWhinney, 1978; Mills, 1986). We want to see if the model learns as the child learns. Does the model make the same errors the child makes and is the order of correct mastery of forms in the simulations like the one we find in the child? Declension in German The declensional paradigm in German is configured around three morphosyntactic dimensions. 1. Number. As in English, nouns and pronouns in German can vary in number, since they can be either singular or plural. For example, the word for “student” is Student and the plural form is Studenten. Changes in number are also marked on the article or other modifier, so that the singular form “the student” der Student becomes plural die Studenten . 2. Case. The second dimension along which nominals may vary is case. Both nouns and pronouns can be in the nominative, the accusative, the genitive, or the dative case. For example, the nominative singular form of “student” is der Student and its accusative form is den Student. Typically, subjects are in the nominative, direct objects are in the accusative, and indirect objects are in the dative. The genitive is used primarily to mark possession. Prepositions can take either the accusative or the dative and sometimes the genitive. Typically, the dative is used when the verb is is static and the accusative is used when the verb expresses motion. 3. Gender. The third dimension is gender. Nouns can be masculine, feminine, or neuter. A male student is der Student and a female student is die Studentin. The choice of der, die, or das in the nominative reflects choice of one of the three genders for the noun. To some observers (Maratsos and Chalkley, 1980) the assignment of nouns to genders has seemed entirely arbitrary. To others, like Mugdan (1977), the assignment has seemed rule-governed, but exceedingly complex. The noun itself is primarily marked only for the dimension of number. Gender is not marked on the noun and case is only marked on the noun for the genitive singular and the dative plural of certain nouns. The article does the main work of marking gender, number, and case. Simplifying the situation quite a bit, let us focus on the way in which nominal marking is achieved by the selection of the correct form of the definite article. Theoretically, a complete cross of the categories of gender, number, and case yields 24 possible cells for the full German declensional paradigm. Fortunately for the German child, gender distinctions for the definite article disappear in the plural, reducing the paradigm to 16 distinct cells. The complete paradigm for the German definite article is shown in Table 1. Although there are 16 cells in the paradigm, there are only six different forms of the definite article (der , den , dem , des , die , das ). Each form of the article occurs in at least two different cells of the paradigm, so that no form defines a unique combination of gender, number, and case. For example, the article der can mark the masculine nominative singular, the feminine genitive singular, the feminine dative singular, or the genitive plural. Masc. Fem. Neut. Plural N o m . der die das die Gen. des der des der Dat. dem der dem der Acc. den die das die Acquisition of this system is not a trivial task. Overall, this type of learning can be viewed as a three-dimensional word-class formation problem (Levy, Schlesinger, & Braine, 1987). The three dimensions to be controlled are gender, number, and case. In production, control of this system involves correct selection of article and noun markings. On the first dimension, nouns must be placed into one of three gender classes. On the second dimension of number, the child must decide on semantic grounds whether a noun should be singular or plural. If it is to be plural, the child must choose from one of eight pluralization types. On the third dimension of case, the various cues and configurations in the sentence must be grouped together so that they correctly select the case of the noun. In comprehension, the child’s task is to use the various forms of the definite article and the markings on the nouns as cues to the correct assignment of the noun to a particular gender, a particular number, and a particular case. Let us look in more detail at each of these three dimensions of this word-class formation problem. Cues to gender assignment The simplest way to solve the word-class formation problem for gender is for the learner to find a set of reliable cues that tells him when to assign a noun to a certain class. Maratsos and Chalkley (1980) have argued that German gender is so arbitrary that no set of cues would allow a child to assign a noun to its gender class. Why, for example, is “fork” feminine (die Gabel ), “knife” neuter (das Messer ) and “spoon” masculine (der Löffel )? In fact, recent work has shown that, while the German gender system is complex, it is not as arbitrary as it appears on first analysis. In a series of research reports, Klaus-Michael Köpcke and David Zubin (Zubin and Köpcke, 1981, 1986; Köpcke and Zubin, 1983, 1984; Köpcke, 1982) have conducted a broad survey of various types of German nouns and found that there is a large and powerful set of cues to German gender. Using these cues, Köpcke (1982) was able to correctly assign gender to 90% of the 1466 monosyllabic words listed in the first volume of the Duden (Grebe, 1973). The work of Köpcke and Zubin for German is parallel in many ways with that of Tucker, Lambert, and Rigault (1977) on the prediction of gender in French. Both research groups have found that there are indeed a large number of morphological and phonological cues predicting gender. The most important cues discovered by Köpcke and Zubin are given in Table 2 along with a few additional cues taken from a German grammar (Lederer, Schulz and Griesbach, 1969). There are 38 cues in all. Of these, 15 are phonological, 18 are morphological, and 5 are semantic. Some of these cues to gender are absolute. For example, if a word has a diminutive ending (i.e. -lein or -chen ), the noun is guaranteed to be of neuter gender. Other cues are more probabilistic in nature. For example, although nouns that start with an “sh” followed by a consonant tend to be masculine, there are words (e.g. die Stadt, das Spiel ) that violate this mapping. The use of this kind of cue will not guarantee a correct gender classification, but it will improve the chances of a correct classification. Type Gender Example English Phonological umlauting Masculine Der Arger Anger iror drMasculine Der Trieb Force CVMasculine Der Tabak Tobacco CCVMasculine Der Klub Club CCCVMasculine Der Strich Stroke -VC Masculine Der Beamier Official -VCC Masculine Der Hahn Rooster -VCCC Masculine Der Mark! Market monosyllabic Masculine Der Akt Nude shCMasculine Der Schrank Closet -el Masculine Der Schliissel Key -n Masculine Der Zahn Tooth -fricalive + 1 Feminine Die Nacht Night -e Feminine Die Sonne Sun -<e)s Non-feminine Das Glas Glass Morphological -ling Masculine Der Feigling Coward -ent Masculine Der Patient Patient -er Masculine Der Reiter Rider -eur Masculine Der Redakteur Editor -ei Feminine Die Malerie Painting -ie Feminine Die Phantasie Fantasy -ik Feminine Die Polemik Polemic -in Feminine Die Studentin Co-ed -ion Feminine Die Portion Portion -ilat Feminine Die Realitat Reality -sis Feminine Die Basis Basis -ung Feminine Die Zeitung Newspaper -lein Neuter Das Fraulein Young woman -ment Neuter Das Instrument Instrument -ell Neuter Das Tablet! Tablet -chen Neuter Das Madchen Girl -en Neuler Das Sehen Seeing -um Neuter Das Gymnasium High School Semantic Natural male Masculine Der Sohn Son Natural female Feminine • Die Tochter Daughter Young being Neuter Das Kind Child Superordinate Neuter Das Tier Animal Cues to selection of a plural marker Unlike the gender dimension, which has no single real-world correlate, the dimension of number maps directly onto salient features of the external world. The decision to treat a noun as singular or plural involves none of the complexities of the decision to treat a noun as either masculine, feminine, or neuter. Whereas gender is marked only on the modifiers and never on the noun, number is marked most clearly on the noun itself. But this marking is not simple (Köpcke, 1988), since there are eight different ways to mark the plural. The actual choice of one of these eight forms is governed by a set of cues that are almost as complex as those governing gender assignment. For example, the plural of die Flut “flood” is die Fluten , the plural of das Gut “estate” is die Güter while the plural of der Hut “hat” is die Hüte . There are some regularities in the assignment of these plural morphemes to a word based on the suffixes and prefixes on the stem, the mutability of the stem vowel, and the gender and animacy of the noun (Köpcke, 1988). Table 3 illustrates the eight possible ways in which nouns may be pluralized. Change Singular Plural Translation