<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0607">
  <Title>EBLA: A Perceptually Grounded Model of Language Acquisition</Title>
  <Section position="3" start_page="0" end_page="0" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> EBLA is based on research into language acquisition in children as well as existing computational models. This section highlights some of this related research. For a more detailed discussion of existing works on early language acquisition in children including works by Calvin and Bickerton (2001), Lakoff (1990), Locke (1993), Norris and Hoffman (2002), Pinker (2000), and Smith (1999), see chapter 2 of Pangburn (2002). For a more detailed discussion of existing computational models including Steels and Kaplan (2000) and Roy (1999; 2000), see chapter 3 of Pangburn (2002).</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Experiential Model of Child Development and
Language Acquisition
</SectionTitle>
      <Paragraph position="0"> Katherine Nelson (1998) has worked to bring together many of the domains involved in the cognitive development of children with special emphasis on the role played by language. She views language and cognition as heavily intertwined--language cannot develop without early, nonlinguistic cognitive function, and full cognitive development cannot occur without language.</Paragraph>
      <Paragraph position="1"> Nelson takes an experiential approach to her work, focusing on how children adapt to meet their current needs and how that adaptation then affects their future experiences.</Paragraph>
      <Paragraph position="2"> Nelson's Experiential Model is centered on events in the child's environment rather than objects. Nelson broadly defines an event as &amp;quot;an organized sequence of actions through time and space that has a perceived goal or end point.&amp;quot; (Nelson 1998, 93-94) Events place objects and actions on those objects in the context of their ultimate goal or purpose, adding temporal ordering with a beginning and an ending. A child's perception, processing, classification, and storage of events form his/her mental event representations (MERs). The MER becomes the cognitive building block for increasingly complex knowledge representation and, ultimately, natural language.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Cross-Situational Techniques for Lexical Ac-
</SectionTitle>
      <Paragraph position="0"> quisition Throughout the 1990's, Siskind (1992; 1997) has established algorithms to map words to symbolic representations of their meanings. For example, given the utterance, &amp;quot;John walked to school.&amp;quot; and a symbolic representation of the event, &amp;quot;GO(John, TO(school)),&amp;quot; his system would learn the mappings, &amp;quot;John-John, walked-GO(x, y), t -TO(x), and school-school.&amp;quot; To perform the word-to-meaning mappings, Siskind utilizes cross-situational learning. Basically, this means that the system resolves mappings only after being presented with multiple utterance/symbolic concept sets representing multiple situations. By drawing inferences about word mappings from multiple uses, the system is able to determine the correct symbolic mappings.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.3 Force Dynamics and Event Logic for
Grounded Event Recognition
</SectionTitle>
      <Paragraph position="0"> In distinct but related research, Siskind (1992; 2000; Siskind and Morris 1996) has developed several software systems to classify and describe dynamic events.</Paragraph>
      <Paragraph position="1"> In 1992, he described ABIGAIL, a system that constructs semantic descriptions of events occurring in computer-generated stick-figure animations. ABIGAIL perceives events by detecting support, contact, and attachment using counterfactual simulation.</Paragraph>
      <Paragraph position="2"> Using a subsequent system named HOWARD, Siskind and Morris built event representations based on real video. HOWARD produces hidden Markov models (HMMs) of the motion profiles of the objects involved in an event.</Paragraph>
      <Paragraph position="3"> Siskind's most recent approach has been to use event-logic to describe changes in support, contact, and attachment, which he now terms force-dynamics. His latest system, LEONARD, uses a camera to capture a sequence of images and then processes that sequence  using three subroutines: 1. Segmentation-and-Tracking - places a polygon around the objects in each frame 2. Model-Reconstruction - builds a force dynamic model of each polygon scene, determining grounding, attachment, and depth/layering 3. Event-Classification - determines over which in null tervals various primitive event types are true and from that data, over which intervals various compound event types are true</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.4 X-Schemas, F-Structs, and Model-Merging for
Verb Learning
</SectionTitle>
      <Paragraph position="0"> Bailey (1997) has developed a computational model of the role of motor control in verb acquisition. He argues that proprioception, which is knowledge of the body's own state, is linked to the acquisition of action verbs. In fact, he maintains that grounding action verbs in the motor-control system constrains the variety of lexical action categories and makes verb acquisition tractable.</Paragraph>
      <Paragraph position="1"> Bailey introduces the executing schema (x-schema) as a mechanism that can represent and carry out verbal commands, and feature structures (f-structs) as a mechanism for linking x-schema activities to related linguistic features.</Paragraph>
      <Paragraph position="2"> X-schemas are formal representations of sequences of motor control actions. In Bailey's model, x-schemas are modeled as Petri nets with extensions to handle the passing of parameters.</Paragraph>
      <Paragraph position="3"> In order to connect x-schemas to verbs, the linking feature structure (f-struct) is introduced. The f-struct is an intermediate set of features that allows a layer of abstraction between the individual motions of an action and the action verb that describes them. An f-struct is a list of feature-value pairs represented in a table with two rows. Each pair maps to a column with the feature located in the top row and the value in the bottom row.</Paragraph>
      <Paragraph position="4"> Bailey experientially determined a list of twelve features for his system comprised of eight motor control features and four perceived world state features.</Paragraph>
      <Paragraph position="5"> Bailey's system performs verb acquisition using an algorithm that develops a lexicon of word senses based on a training set of verbs and linking f-structs summarizing that verb. Verb learning becomes an optimization problem to find the best possible lexicon given the training examples. Bailey terms this approach for merging word senses, model-merging, and implements a solution using a hill-climbing algorithm.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>