<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2011">
  <Title>Knowledge Extraction Using Dynamical Updating of Representation</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> From an historical perspective, four types of knowledge representation schemas are worth to be considered (W.Kintsch, 1998).</Paragraph>
    <Paragraph position="1"> &amp;quot;Feature systems&amp;quot; (J.J. Katz, J.A. Fodor, 1963) have been developed in philosophy and linguistics and became very popular especially in psychology.</Paragraph>
    <Paragraph position="2"> This representation aimed at finding a limited set of basic semantic characteristics that, combined by means of particular composition rules, could express complex concepts. It was a very simple representation system but conceptual relations were not considered. Furthermore the defined features did not change with the context and the goals that had to be achieved.</Paragraph>
    <Paragraph position="3"> &amp;quot;Associative networks&amp;quot; consider also semantic relations between concepts. Knowledge is represented by a network of concepts bounded by more or less strong associations. This formalism is bolstered by a lot of experimental data, for example by word priming experiments (D.E.</Paragraph>
    <Paragraph position="4"> Meyer, R.W. Schvaneveldt, 1971). But networks whose links are not labelled are not very expressive.</Paragraph>
    <Paragraph position="5"> &amp;quot;Semantic networks&amp;quot; (A.M. Collins, M.R. Quillian, 1969) are an evolution of associative networks. Concepts continue to be symbolized by nodes, but these are linked by labeled arcs (IS-A, PART-OF etc.). In this way well ordered concept hierarchies can be defined and the hereditariness of properties is allowed.</Paragraph>
    <Paragraph position="6"> &amp;quot;Schemas&amp;quot;, &amp;quot;frames&amp;quot; and &amp;quot;scripts&amp;quot; are structures for coordinating concepts that belong to the same event or superstructure. Classical examples of these formalisms are the &amp;quot;room frame&amp;quot; of Minsky (M. Minsky, 1975) and the restaurant script of Schank and Abelson (R.C. Schank, R.P. Abelson, 1977).</Paragraph>
    <Paragraph position="7"> The problem with these representation forms is that they are static. In fact human mind generates contextualized structures, that are adapted to the particular context of use.</Paragraph>
    <Paragraph position="8"> &amp;quot;Networks of propositions&amp;quot; (or &amp;quot;knowledge nets&amp;quot;, W.Kintsch, 1998) are an alternative formalism that combines and extends the advantages of the representation forms that have been introduced so far.</Paragraph>
    <Paragraph position="9"> The predicate-argument schema can be considered as the fundamental linguistic unit especially in the representation of textual content. Atomic propositions consist of a relational term (the predicate) and one or more arguments.</Paragraph>
    <Paragraph position="10"> Networks of propositions link these atomic propositions through weighted and not labeled arcs. According to this formalism the meaning of a node is given by its position in the net.</Paragraph>
    <Paragraph position="11"> From a psychologic point of view only the nodes that are active (i.e. that are maintained in the working memory) contribute to specify the sense of a node. Hence the meaning of a concept is not permanent and fixed but is built every time in the working memory by the activation of a certain subset of propositions in the neighbour of the node that represents the concept. The context of use (objectives, accumulated experiences, emotional and situational state etc.) determines which nodes have to be activated.</Paragraph>
    <Paragraph position="12"> For the definition of retrieval modalities Ericsson and Kintsch has introduced the concept of long term working memory (LTWM) (W.Kintsch, V.L. Patel, K.A.Ericsson, 1999). They noticed that some cognitive tasks, as textual comprehension, cannot be explained only using the concept of working memory. Given the strict limits of capacity of the short term memory (STM) and of the working memory (WM), tasks that require an enormous employment of resources cannot be carried out.</Paragraph>
    <Paragraph position="13"> The theory of long term working memory specifies under which conditions the capacity of WM can be extended. The LTWM is involved only in the execution of well known tasks and actions, that belong to a particular cognitive domain that has been well experienced. In these cases the working memory can be subdivided in a short term part (STWM) that has a limited capacity and a LTWM that is a part of the long term memory represented by the network of propositions. The content of STWM automatically generates the LTWM. In particular objects present in the STWM are linked to other objects in the LTM by fixed and stable memory structures (retrieval cues).</Paragraph>
  </Section>
class="xml-element"></Paper>