<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1160">
  <Title>Computational Cognitive Linguistics</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 The Neural Theory of Language
</SectionTitle>
    <Paragraph position="0"> For some sixteen years, an interdisciplinary group at ICSI and UC Berkeley has been building a theory of language that respects all experimental findings and is consistent with biological and computational constraints. The intellectual base for the project is a synthesis of cognitive linguistics and structured connectionist modelling. This involves work from neurobiology through politics and philosophy (Lakoff 1999). An accessible overview of the goals and methodology of the project can be found in (Regier 1996) and further information is available at: WWW.icsi.berkeley.edu/NTL The focus of this talk is computational; how can an embodied theory of language support scalable systems for language learning and use. The key to scalability in any paradigm is compositionality; our goal in modeling language understanding is to systematically combine the heterogeneous structures posited in cognitive linguistics to yield overall interpretations. We have identified four conceptual primitives that appear to capture the suffice for building scalable language understanding systems: SCHEMA, MAP, (MENTAL) SPACE, and CONSTRUCTION.</Paragraph>
    <Paragraph position="1"> Schemas are language-independent representation of meanings and constructions are form-meaning pairs. Constructions are form-meaning mappings as depicted in Figure 1. Maps and mental spaces are discussed in (Mok 2004). The ECG formalism has, in addition to the usual inheritance hierarchy, an EVOKES relation that makes an outside structure accessible to a schema through a local name.</Paragraph>
    <Paragraph position="2"> As shown in Figure 1, langauge understanding is implemented as having distinct analysis and simulation phases. During analysis, a Semantic Specification (SemSpec) is created from the meaning poles of the constructions, and is essentially a network of schemas with the appropriate roles filled in.. Unification of constructions requires compatability of their embodied semantic scehemas as well as form matching. Crucially, within this network of schemas are executing schemas (or x-schemas Narayanan 1999), which are models of events.</Paragraph>
    <Paragraph position="3"> They are active structures for event-based asynchronous control that can capture both sequential flow and concurrency.</Paragraph>
    <Paragraph position="4"> Simulation is a dynamic process which includes executing the x-schemas specified in the SemSpec and propagating belief updates in a belief network (Jensen 1996, Narayanan 1999). This will be discussed further in Section 5.</Paragraph>
    <Paragraph position="5"> Figure1. Overview of the Comprehension Model</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Embodied Construction Grammar
</SectionTitle>
    <Paragraph position="0"> The cornerstone of the effort is the formalism called Embodied Construction Grammar (ECG). In traditional terms, ECG resembles a unification grammar like HPSG or LFG and many of the computational insights carry over. But the central task of grammar is taken to be accounting for the full range of language learning and use rather than the specification of acceptable forms.</Paragraph>
    <Paragraph position="1"> Grammars in ECG are deeply cognitive, with meaning being expressed in terms of cognitive primitives such as image schemas, force dynamics, etc. The hypothesis is that a modest number of universal primitives will suffice to provide the core meaning component for the grammar. Specific knowledge about specialized items, categories and relations will be captured in the external ontology. As a linguistic formalism, ECG combines the idea of Construction Grammar as form-meaning pairings (Croft 2001, Fillmore &amp; Kay 1999, Goldberg 1995, etc.) with the embodied semantics of the Cognitive Linguistics tradition (Fauconnier 1997, Lakoff 1999, Langacker 1991, etc.).</Paragraph>
    <Paragraph position="2"> Computationally, the central ideas involve probabilistic relational models (Pfeffer and Koller 2000, etc.) and active knowledge (Bailey 1998, Narayanan 1999, etc.) along with their reduction to structured connectionist form and thus to neural models (Shastri 2002).</Paragraph>
    <Paragraph position="3"> A grammar specification in ECG should simultaneously fill three distinct functions: capturing linguistic insights, specifying the analysis phase of a computational system, and serving as the high level description of the neural embodiment of the grammar. This has been achieved in some simple cases and serves as a major constraint in all ECG efforts.</Paragraph>
    <Paragraph position="4"> The deep semantic construction grammar of ECG also supports a novel style of general robust parsing. The first phase of analysis is an modified chunk parser (Abney 1996).</Paragraph>
    <Paragraph position="5"> The chunker generates a set of semantic chunks stored in a chart. The second phase of analysis extracts the smallest number of chunks that span the utterance from the chart, and performs semantic integration. Their common semantic structures are merged, and the resulting analyses are ranked according to the semantic density metric Without a complete analysis of an utterance, the system must infer exactly how a set of local, partial semantic structures fit together into a coherent, global analysis of the utterance. The approach taken is an abductive one in that it assumes compatible structures are the same and merges them. This has been shown to work quite well in a system for modeling the learning of new constructions (Bryant 2004, Chang 2004).</Paragraph>
  </Section>
class="xml-element"></Paper>