<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1160">
  <Title>Computational Cognitive Linguistics</Title>
  <Section position="5" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Applications
</SectionTitle>
    <Paragraph position="0"> Narayanan (Narayan 1999) has built a biologically plausible model of how such metaphorical uses can be understood by mapping to their underlying embodied meaning. We assume that people can execute x-schemas with respect to structures that are not linked to the body, the here and the now. In this case, x-schema actions are not carried out directly, but instead trigger simulations of what they would do in the imagined situation.</Paragraph>
    <Paragraph position="1"> This ability to simulate or imagine situations is a core component of human intelligence and is central to our model of language. The system models the physical world as other x-schemas that have input/output links to the x-schema representing the planned action.</Paragraph>
    <Paragraph position="2"> In the computational implementation, the spatial motion domain (source domain) is encoded as connected x-schemas. Our model of the source domain is a dynamic system based on inter-xschema activation, inhibition and interruption. In the simulation framework, whenever an executing x-schema makes a control transition, it potentially modifies state, leading to asynchronous and parallel triggering or inhibition of other x-schemas. The notion of system state as a graph marking is inherently distributed over the network, so the working memory of an x-schema-based inference system is distributed over the entire set of x-schemas and source domain feature structures. The KARMA system has been tested on narratives from the abstract domain of international economics. The implemented model has about 100 linked x-schemas, and about 50 metaphor maps from the domains of health and spatial motion.</Paragraph>
    <Paragraph position="3"> These were developed using a database of 30 2-3 phrase fragments from newspaper stories all of which have been successfully interpreted by the program. Results of testing the system have shown that a surprising variety of fairly subtle and informative inferences arise from the interaction of the metaphoric projection of embodied terms with the topic specific target domain structure (Narayanan, 1999). Among the inferences made were those related to goals (their accomplishment, modification, subsumption, concordance, or thwarting), resources (consumption, production, depletion, level), aspect (temporal structure of events) frame-based inferences, perspectival inferences, and inferences about communicative intent.</Paragraph>
    <Paragraph position="4"> The ECG formalisms as well as the analyzer described above play a crucial role in a computational model of how language comprehension may drive the acquisition of early phrasal and clausal constructions (Chang, 2004).</Paragraph>
    <Paragraph position="5"> This model takes ECG as the target representation to be learned from a sequence of utterances in context. Learning is usage-based in that utterances are first analyzed using the existing set of constructions, typically resulting in only a partial analysis that neither provides complete coverage of the richer background context nor exploits all the potential input forms and relations in the utterance. This incomplete analysis prompts the formation of new constructions that take up the slack.</Paragraph>
    <Paragraph position="6"> Constructions can also be formed on the basis of similarity (two constructions can merge into a more general construction) and co-occurrence (two constructions can compose into a larger construction).</Paragraph>
    <Paragraph position="7"> Besides specifying the means for forming new ECG constructions, the acquisition model provides an overarching computational framework for converging on an optimal set of constructions, based on a minimum description length principle ( Rissanen 1978) that favors compactness in describing both the grammar and the statistical properties of the data. This framework extends previous work in Bayesian model merging for lexical acquisition (Bailey, 1997) and the induction of context-free grammars (Stolcke 1994) to handle the relational structures and usage-based considerations madepossible with ECG.</Paragraph>
    <Paragraph position="8"> Specifically, the criteria employed favor constructions that have simple descriptions (relative to the available meaning representations and current set of constructions) and are frequently employed.</Paragraph>
    <Paragraph position="9"> The model has been applied to learn simple English motion constructions from a corpus of child-directed utterances, paired with situation representations. The resulting learning trends reflect cross-linguistic acquisition patterns, including the incremental growth of the constructional inventory based on experience, the prevalence of early grammatical markers for conceptually basic scenes (Slobin, 1985) and the learning of lexically specific verb island constructions before more abstract grammatical patterns (Tomasello, 1992). For current purposes, the systems described demonstrate the utility of the ECG formalism for supporting computational modeling and offers a precisely specified instantiation of the simulation-based approach to language.</Paragraph>
    <Paragraph position="10"> Conclusion For current purposes, the systems described above demonstrate the utility of the ECG formalism for supporting computational modeling and offer a precisely specified instantiation of the simulation-based approach to language.</Paragraph>
    <Paragraph position="11"> Cognitive Linguistics has developed many profound insights, but these had not been formalized and made computationally tractable.</Paragraph>
    <Paragraph position="12"> Recent results like these suggest that a Computational Cognitive Linguistics is both scientifically productive and a semantic basis for a wide range of natural language understanding applications.</Paragraph>
  </Section>
class="xml-element"></Paper>