<?xml version="1.0" standalone="yes"?>
<Paper uid="M93-1025">
  <Title>USC : DESCRIPTION OF THE SNAP SYSTEM USED FOR MUC- 5</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
INTRODUCTION
Background
</SectionTitle>
    <Paragraph position="0"> The SNAP information extraction system has been developed as a part of a three-year SNAP projec t sponsored by the National Science Foundation. The main goal of the SNAP project is to build a massively parallel computer capable of fast and accurate natural language processing [5] . Throughout the project, a parallel computer was built in the Parallel Knowledge Processing Laboratory at USC, and various softwar e was developed to operate the machine [3] . The approach in designing SNAP was to find a knowledge repre sentation and a reasoning paradigm useful for natural language processing which exhibits massive parallelism.</Paragraph>
    <Paragraph position="1"> We have selected marker-passing on semantic networks as a way to represent and process linguistic knowledge .</Paragraph>
    <Paragraph position="2"> The work for MUC-5 started at the end of January 1993. Prior to this we had implemented on SNA P a memory-based parsing system which was also used for MUC-4 . Since the domain has been changed fro m MUC-4, we improved the dictionary with semantic tags necessary for EJV domain, developed a new templat e generation module, and constructed a new knowledge base including concept hierarchy and concept sequenc e patterns for parsing. Our group consisting of one faculty and five graduate students spent approximately 6 months to engineer a large system .</Paragraph>
    <Paragraph position="3"> Approach The underlying ideas of the SNAP natural language processing system are : (1) memory based parsing , and (2) marker-passing on semantic networks [5] . In SNAP, parsing becomes a guided search over the knowledge base which stores linguistic information . Input words activate and predict concepts in the knowledge base. While the semantic network represents the static part of the knowledge base, marker passing is th e mechanism which changes the state of the knowledge base with each new word . Based on the MUC-4 experience, we emphasized efficiency and practicality in developing the system for MUC-5 . Full analysis of input texts is avoided. The system picks up only necessary and relevant information from input texts . The knowledge base of domain specific phrasal patterns for TIE-UP relationship has been constructed automatically by using a lexical acquisition system called PALKA, and couple of back-up routes have been established to extract partial information when parsing fails.</Paragraph>
  </Section>
class="xml-element"></Paper>