Abstract: Recent experimental work indicates that by the age of ten months, infants have already learned a great deal about the phonotactics (legal sounds and sound sequences) of their language. This learning occurs before infants can utter words or apprehend most phonological alternations. I will show that this early learning stage can be straightforwardly modeled with Optimality Theory. Specifically, the Markedness and Faithfulness constraints can be ranked so as to characterize the phonotactics, even when no information about morphology or phonological alternations is yet available. I will also show how later on, the information acquired in infancy can help the child in coming to grips with the alternation pattern. I also propose a procedure for undoing the learning errors that are likely to occur at the earliest stages. There are two specific formal proposals. One is a constraint ranking algorithm, based closely on Tesar and Smolensky’s Constraint Demotion, which mimics the early, “phonotactics only” form of learning seen in infants. I illustrate the algorithm’s effectiveness by having it learn the phonotactic pattern of a simplified language modeled on Korean. The other proposal is that there are three distinct default rankings for phonological constraints: low for ordinary Faithfulness (used in learning phonotactics); low for Faithfulness to adult forms (in the child’s own production system); and high for output-to-output correspondence constraints.

Abstract: We document that monetary policy inertia can help alleviate problems of indeterminacy and non-existence of stationary equilibrium observed for some commonly-studied monetary policy rules. We also find that inertia promotes learnability of equilibrium. The context is a simple, forward-looking model of the macroeconomy widely used in the rapidly expanding literature in this area. We conclude that this might be an important reason why central banks in the industrialized economies display considerable inertia when adjusting monetary policy in response to changing economic conditions.

Abstract: This article addresses the learnability of auditory icons, that is, environmental sounds that refer either directly or indirectly to meaningful events. Direct relations use the sound made by the target event whereas indirect relations substitute a surrogate for the target. Across 3 experiments, different indirect relations (ecological, in which target and surrogate coexist in the world; metaphorical, in which target and surrogate have similar appearance or function, and random) were compared with one another and with direct relations on measures including associative strength ratings, amount of exposure required for learning, and response times for recognizing icons. Findings suggest that performance is best with direct relations, worst with random relations, and that ecological and metaphorical relations involve distinct types of association but do not differ in learnability.

Abstract: Many constraint-based accounts of language-internal phonological inconsistencies appeal to the idea that the phonological lexicon is stratified into distinct subcomponents. This article adopts the perspective of the language learner in reexamining two assumptions related to this approach: 1 ) that the membership of any lexical item can be determined on the basis of distributional evidence (strong lexical stratification); and 2) that ranking inconsistencies should be modeled through multiple faithfulness constraints indexed to different strata (indexed faithfulness). It is argued that a phonological grammar with strong lexical stratification cannot be acquired given the lack of positive distributional evidence available to the learner. When the strong lexical stratification hypothesis is retracted, a computational problem emerges for the acquisition of indexed faithfulness. Taking these issues into consideration, the article explores a revised model of lexical stratification that can be learned based on alternation data and general learning mechanisms.

Abstract: The number of concepts in a model has been frequently used in the literature to measure the ease of use in creating model schemas. However, to the best of our knowledge, nobody has looked at its effect on the readability of the model schemas after they have been created. The readability of a model schema is important in situations where the schemas are created by one team of analysts and read by other analysts, system developers, or maintenance administrators. Given the recent trend of models with increasing numbers of concepts such as the unified modeling language (UML), the effect of the number of concepts (NOC) on the readability of schemas has become increasingly important. In this work, we operationalize readability along three dimensions: effectiveness, efficiency, and learnability. We draw on the Bunge Wand Weber (BWW) framework, as well as the signal detection recognition theory and the ACT theory from cognitive psychology to formulate hypotheses and conduct an experiment to study the effects of the NOC in a data model on these readability dimensions. Our work makes the following contributions: (a) it extends the operationalization of the readability construct, and (b) unlike earlier empirical work that has focused exclusively on comparing models that differ along several dimensions, this work proposes an empirical methodology that isolates the effect of a model-independent variable (the NOC) on readability. From a practical perspective, our findings have implications both for creators of new models, as well as for practitioners who use currently available models for creating schemas to communicate requirements during the entire lifecycle of a system.

Abstract: Crain, S. & Thornton, R. Investigations in Universal Grammar. Cambridge: MIT, 1998.Stephen Crain (C) & Rosalind Thornton (T) have garnered a well-deserved reputation for their unwavering commitment to language learnability as a constraint not only on theories of child language and language development but also on experimental design and the interpretation of experimental findings. In his well-known defense of children's early knowledge of syntactic constraints, Crain (1991) argued for the widely-held position that the best solution to the learnability problem is to assume that grammatical knowledge which cannot be learned on the basis of experience is specified in advance as part of the human biological endowment for language in the form of a UNIVERSALGRAMMAR (UG) (Chomsky, 1965). With respect to experimental design, C&T have strongly maintained that even young children know UG constraints but perform poorly in some experiments-due to the extralinguistic demands associated with experimental tasks, particularly those involved in presupposition accommodation and complex response planning. C&T specifically design their experiments to reduce the impact of extralinguistic demands on children's linguistic performance while at the same time providing felicitous environments for adultlike performance.

Abstract: In this paper I investigate the assignation of syllabic structure to segments in second language learners invoking principles of parsing and learnability. Drawing on the model of Phillips (1996) in parsing, and the work of Fodor (1999) and Dresher (1999) in learnability, I discuss the implications for second language learning. I also look at the question of whether new phonological structure can be triggered if that structure is lacking in the first language. Drawing on evidence from the acquisition of (1) phonological features in Japanese, Chinese and French learners of English, as well as English learners of Czech, and (2) moraic structure by an English speaker learning Japanese consonant and vowel length contrasts, I argue that second language learners can trigger new prosodic structure. The process of acquisition is a combination of acquiring new structures, and mapping the interfaces of different levels of structure via a phonological parser.

Abstract: The aim of this paper is to try to understand the process of children’s language acquisition by using the theory of inference of formal grammars. Toward this goal, we introduce an extension of Marcus External Contextual grammars which constitutes a Mildly Context-Sensitive language family, and study their learnability in the limit from positive data. Finally, we briefly indicate our future research direction.

Abstract: 1. The claim The present paper will reconsider the acquisition of free anaphors in French. Hamann et al. (1996) have pointed out that object clitics do not appear as easily in French child language as subject clitics. When the amount of object clitics is compared with the total amount of object phrases, there is delayed rise in the use of object clitics. The phenomenon is less marked for subject clitics. Further study by Jakubowicz et al. (1998) was directed at SLI children. The acquisition delay for object clitics was clearly reaffirmed. Since subject/object asymmetries have been a concern for some time in the theory of syntax, Hamann et al.’s original observations led to proposals that derive the acquisition delay from the argument licensing of clitics. French subject and object clitics do not form a homogeneous class. Kayne (1975:86f) assumes the subject pronoun cliticized to the head Vfin at at PF, whereas object clitics move from the canonical object position to a position in front of the VP. Jakubowicz et al. (1998) argue that the subject/object asymmetry in acquisition follows from that syntactic difference. The subject clitic would not only express the subject’s D-features (person/number/gender), but also the verbs Ifeature ‘finite’. This additional I-function would enhance the learnability of the subject and so explain the comparitively delayed rise of object clitics.

Abstract: In general a grammar describes a (possibly infinite) set of sentences with a finitestructural description. Computational Grammar Induction (CGI) deals with the cre-ation of computational models for identification of these infinite sets on the basis ofa finite set of examples. CGI is a field in its own right, with its own internal researchquestions, many of which have no direct impact on the study of human language.Yet it is clear that computational models created by the CGI community might beof interest to the linguistic community because human language after all appears tobe an infinite set, the description of which is learned efficiently in a relative shorttime. There are various domains in which learnability of a language might be inter-esting for linguists: e.g., first language acquisition, second language acquisition, orthe automatic extraction of grammars from corpora.In this article we will focus on first language acquisition with some suggestions forextensionstootherareas.

Abstract: Context-free grammars cannot be identified in the limit from positive examples (Gold 1967), yet natural language grammars are more powerful than context-free grammars and humans learn them with remarkable ease from positive examples (Marcus 1993). Identifiability results for formal languages ignore a potentially powerful source of information available to learners of natural languages, namely, meanings. This paper explores the learnability of syntax (i.e. context-free grammars) given positive examples and knowledge of lexical semantics, and the learnability of lexical semantics given knowledge of syntax. The long-term goal is to develop an approach to learning both syntax and semantics that bootstraps itself, using limited knowledge about syntax to infer additional knowledge about semantics, and limited knowledge about semantics to infer additional knowledge about syntax.

Abstract: The child's linguistic input often includes evidence for incorrect grammatical analyses. For instance, the child acquiring English will hear sentences such as "Here comes the train," but English is not in general a V2 language. How does the child know which evidence to trust? Could one V2 sentence shift a major parameter? Or perhaps block a child’s progress, because the input contains an unresolvable contradiction? Although classic learnability theory assumes that a single example can suffice to change a parameter, we argue that there has to be a method to filter out certain sentences. Otherwise, the child will be paralyzed by contradictory (e.g. V2 and non-V2) input. In this paper we propose that recursion plays a central role in the differentiation of “exceptional” domains from truly productive grammar. As argued by Hauser, Chomsky and Fitch (2002), recursion is the central characteristic of core grammar. Our proposal is that recursion tells the child when a productive, grammatical operation has applied.

Abstract: Context-Driven Construction Learning Nancy Chang (nchang@icsi.berkeley.edu) UC Berkeley, Department of Computer Science and International Computer Science Institute 1947 Center St., Suite 600, Berkeley, CA 94704 Olya Gurevich (olya@socrates.berkeley.edu) UC Berkeley, Department of Linguistics 1203 Dwinelle Hall, University of California at Berkeley Berkeley, CA 94720-2650 communicated in context. We assume along with many the- ories of language that the basic unit of linguistic knowledge, for both lexical items and larger phrasal and clausal units, is a symbolic pairing of form and meaning, or construction (Langacker, 1987; Goldberg, 1995; Fillmore and Kay, 1999). Since the target of learning is rooted in both form and mean- ing domains, the learner should exploit information from both domains during learning. Most importantly, we view linguistic constructions as in- herently dependent on and supportive of dynamic processes of language use, anchored in a communicative context. A crucial but often neglected source of bias in learning con- structions must therefore be how much they help the child meet her communicative goals. This paper presents a computational model of construction learning consistent with these principles, focusing on how language understanding drives language learning. We de- scribe a statistically driven machine learning framework that takes as input a sequence of child-directed utterances paired with their associated situational context, along with the cur- rent grammar, or set of constructions; this grammar is ini- tially restricted to lexical items. The utterances are passed to a language understanding system (Bryant, 2003) that pro- duces a partial interpretation, which provides the basis for the learning model to form new constructions. We present re- sults showing how the model acquires simple English “verb island” constructions (Tomasello, 1992), and discuss how the same mechanisms handle the more complex constructions in- volved in Russian nominal case marking. These studies lend support for the larger program of integrating cognitive and constructional approaches to linguistics, crosslinguistic de- velopmental evidence, and machine learning techniques to address the puzzles of language acquisition. Abstract We present a computational model of how partial comprehen- sion of utterances in context may drive the acquisition of chil- dren’s earliest grammatical constructions. The model aims to satisfy convergent constraints from cognitive linguistics and crosslinguistic developmental evidence within a statistically driven computational framework. We examine how the tight coupling between contextually grounded language comprehen- sion and learning processes can be exploited to improve the model’s ability to search the space of possible constructions. In particular, previously learned constructions may not fully ac- count for all contextually perceived mappings between forms and meanings. In the model, these incomplete analyses di- rectly prompt the formation of new relational mappings that bridge the gap. We describe an experiiment applying the model to the acquisition of English verb island constructions and dis- cuss how the model handles more complex examples involving Russian morphological constructions. Together these demon- strate the viability of the overall approach and representational potential of the model. Beyond single words How do children make the leap from single words to complex combinations? The simple act of putting one word in front of another to indicate some relation between their meanings is widely considered the defining characteristic of linguistic competence and the key to unlocking the combinatorial and expressive power of language. A viable account of the acqui- sition of these combinatorial patterns, or grammatical con- structions, would thus have significant implications for any theory of language that aspires to cognitive plausibility. As with most issues impinging on the nature of gram- mar, linguistic and developmental inquiries into the source of combinatorial constructions have bifurcated along theoret- ical lines. These reflect divergent assumptions about, among other things, what kind of learning bias children bring to the task, how the target linguistic knowledge should be repre- sented, what kind of data should be considered part of the training input, and how (if at all) language learning interacts with other linguistic and cognitive processes. Theoreticians within the formalist “learnability” paradigm, for example, have generally restricted their attention to the form domain, taking the input for learning to be a set of surface strings (each a sequence of surface forms) and positing relatively abstract structures that govern the combination of linguistic units. This paper takes as starting point the hypothesis that the learning problem at hand may encompass a broader subset of the child’s experience, centrally including meaning as it is The Construction Learning model We briefly describe the construction learning model in terms of (1) the target representation of learning, (2) assumptions about the child language learning scenario, and (3) the com- putational learning framework; see (Chang, 2004; Chang and Maia, 2001) for more details. Target representation: embodied constructions Embodied Construction Grammar (Bergen and Chang, in press; Chang et al., 2002) is a computationally explicit for- malism for capturing insights from the construction gram- mar and cognitive linguistics literature. ECG supports an approach to language understanding based on two linked

Abstract: To avoid human errors in complex surgical work systems, it is mandatory to assess usability and reliability of new planning and navigation systems prior to its introduction into clinical routine. Clinical usability analysis includes guideline and user-based evaluation. In a user-based study, interaction of 25 expert and novice surgeons with a commercial CAS system was observed and assessed based on a FMEA under consideration of human errors. In total 152 incidents were observed. Ninety-nine (65%) were rated to be critical for the process as such, whereas 46 (30%) were found to be critical for the patient. Comparing interaction times of novices and expert surgeons have been the basis for an assessment of systems learnability. For novices preoperative planning revealed to be more difficult to learn than the intraoperative navigation. Moreover, a multidimensional questionnaire was used to document subjective user assessment of systems error tolerance, learnability as well as user satisfaction. Although novices obviously encountered severe problems during the tests, they rated learnability significantly better than experts. In contrast, experts' rating matched the findings of the test observers.

Abstract: This paper deals with two well discussed, but largely open problems on E-pattern languages, also known as extended or erasing pattern languages: primarily, the learnability in Gold’s learning model and, secondarily, the decidability of the equivalence. As the main result, we show that the full class of E-pattern languages is not inferrable from positive data if the corresponding terminal alphabet consists of exactly three or of exactly four letters – an insight that remarkably contrasts with the recent positive finding on the learnability of the subclass of terminal-free E-pattern languages for these alphabets. As a side-effect of our reasoning thereon, we reveal some particular example patterns that disprove a conjecture of Ohlebusch and Ukkonen (Theoretical Computer Science 186, 1997) on the decidability of the equivalence of E-pattern languages.

Abstract: This paper examines the learnability of a major subclass of E-pattern languages – also known as erasing or extended pattern languages – in Gold’s learning model: We show that the class of terminal-free E-pattern languages is inferrable from positive data if the corresponding terminal alphabet consists of three or more letters. Consequently, the recently presented negative result for binary alphabets is unique.

Abstract: This thesis investigates language acquisition and evolution, using the methodologies of Bayesian inference and expression-induction modelling, making specific reference to colour term typology, and syntactic acquisition. In order to test Berlin and Kay’s (1969) hypothesis that the typological patterns observed in basic colour term systems are produced by a process of cultural evolution under the influence of universal aspects of human neurophysiology, an expression-induction model was created. Ten artificial people were simulated, each of which was a computational agent. These people could learn colour term denotations by generalizing from examples using Bayesian inference, and the resulting denotations had the prototype properties characteristic of basic colour terms. Conversations between these people, in which they learned from one-another, were simulated over several generations, and the languages emerging at the end of each simulation were investigated. The proportion of colour terms of each type correlated closely with the equivalent frequencies found in the World Colour Survey, and most of the emergent languages could be placed on one of the evolutionary trajectories proposed by Kay and Maffi (1999). The simulation therefore demonstrates how typological patterns can emerge as a result of learning biases acting over a period of time. Further work applied the minimum description length form of Bayesian inference to modelling syntactic acquisition. The particular problem investigated was the acquisition of the dative alternation in English. This alternation presents a learnability paradox, because only some verbs alternate, but children typically do not receive reliable evidence indicating which verbs do not participate in the alternation (Pinker, 1989). The model presented in this thesis took note of the frequency with which each verb occurred in each subcategorization, and so was able to infer which subcategorizations were conspicuously absent, and so presumably ungrammatical. Crucially, it also incorporated a measure of grammar complexity, and a preference for simpler grammars, so that more general grammars would be learned unless there was sufficient evidence to support the incorporation of some restriction. The model was able to learn the correct subcategorizations for both alternating and non-alternating verbs, and could generalise to allow novel verbs to appear in both constructions. When less data was observed, it also overgeneralized the alternation, which is a behaviour characteristic of children when they are learning verb subcategorizations. These results demonstrate that the dative alternation is learnable, and therefore that universal grammar may not be necessary to account for syntactic acquisition. Overall, these results suggest that the forms of languages may be determined to a much greater extent by learning, and by cumulative historical changes, than would be expected if the universal grammar hypothesis were correct.

Abstract: One argument for parametric models of language has been learnability in the context of first language acquisition. The claim is made that “logical” arguments from learnability theory require non-trivial constraints on the class of languages. Initial formalisations of the problem (Gold, 1967) are however inapplicable to this particular situation. In this paper we construct an appropriate formalisation of the problem using a modern vocabulary drawn from statistical learning theory and grammatical inference and looking in detail at the relevant empirical facts. We claim that a variant of the Probably Approximately Correct (PAC) learning framework (Valiant, 1984) with positive samples only, modified so it is not completely distribution free is the appropriate choice. Some negative results derived from cryptographic problems (Kearns et al., 1994) appear to apply in this situation but the existence of algorithms with provably good performance (Ron et al., 1995) and subsequent work, shows how these negative results are not as strong as they initially appear, and that recent algorithms for learning regular languages partially satisfy our criteria. We then discuss the applicability of these results to parametric and non-

Abstract: This paper investigates the learnability of Pregroup Grammars, a context-free grammar formalism recently defined in the field of computational linguistics. In a first theoretical approach, we provide learnability and non-learnability results in the sense of Gold for subclasses of Pregroup Grammars. In a second more practical approach, we propose an acquisition algorithm from a special kind of input called Feature-tagged Examples, that is based on sets of constraints.

Abstract: "It is a common position in generative acquisition studies to accept Chomsky's view that first language acquisition is determined by a set of innate grammatical a priories. The development of the child would be more a matter of biological maturation than a matter of input-control. Because language universals are innate in the human mind, they cause grammar to grow into the mind almost automatically under the slightest provocations. Early child language would already show the relevance the grammatical a priories. Generative grammarians guided by this view have often drawn far-reaching conclusions about the structure of early child language. The present paper will present an alternative view, the derivation of UG principles from structural acquisition steps. It acknowledges that it is indeed a sentence-generating system that is acquired, but contends that generative systems are learned from the language-specific input material. The basic argument for this approach is that all eventual ‘UG’ properties are identified due to local relations with language specific shapes. One might see the language specific shapes as an entrance to the UG distinctions. Unless a grammar offers a way to identify UG properties, it will not be learnable. This suggests that UG properties may be seen as the outcome of an acquisition procedure rather than being its source"

Abstract: This research proposes methods and instruments for assessing usability of academic digital libraries. Criteria in this study are effectiveness, efficiency, satisfaction, and learnability. It is found that there exist interlocking relationships among effectiveness, efficiency, and satisfaction.

Abstract: Language theorists argue that the reason why spoken language is acquired so rapidly is that we have an innate predisposition for understanding linguistic structures. Theories of perception also hold that there may be deeply seated mechanisms for decomposing visual objects and analyzing them into both component parts and the structural interrelationships of those parts. We propose the theory that diagrams that activate the mechanisms for structural object perception should be similarly easy to learn. This builds on previous work in which we have developed diagramming principles based on the theory of structural object perception. We call these geon diagrams. We have previously shown that such diagrams are easy to remember and to analyze. To evaluate our hypothesis that geon diagrams should also be easy to understand we carried out an empirical study to evaluate the learnability of geon diagram semantics in comparison with the well-established UML convention. The results support our theory of learnability. Both "novices" and "experts" found the geon diagram syntax easier to apply in a diagram-to-textual description matching task than the equivalent UML syntax.

Abstract: The article is exploring the learnabilty issues of decision tables acquired from data within the frameworks of rough set and of variable precision rough set models. Measures of learning problem complexity and of learned table domain coverage are proposed. Several methods for enhancing the learnabilty of decision tables are discussed, including a new technique based on value reducts.

Abstract: We propose in this article a new practical algorithm for inferring μ-distinguishable stochastic deterministic regular languages. We prove that this algorithm will infer, with high probability, an automaton isomorphic to the target when given a polynomial number of examples. We discuss the links between the error function used to evaluate the inferred model and the learnability of the model class in a PAC like framework.

Abstract: This paper is concerned with learning in the model of Gold the Categorial Dependency Grammars (CDG), which express discontinuous (non-projective) dependencies. We show that rigid and k-valued CDG (without optional and iterative types) are learnable from strings. In fact, we prove that the languages of dependency nets coding rigid CDGs have finite elasticity, and we show a learning algorithm. As a standard corollary, this result leads to the learnability of rigid or kvalued CDGs (without optional and iterative types) from strings.

Abstract: Minimalism in ubiquitous interface design allows computational augmentations to seamlessly coexist with existing artifacts and the constellations of task behaviors surrounding them. Specifically, parsimony and transparency contribute to improved learnability and user acceptance of novel interfaces.