<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1014">
  <Title>Automatic Learning for Semantic Collocation</Title>
  <Section position="3" start_page="104" end_page="104" type="intro">
    <SectionTitle>
2 Overview
</SectionTitle>
    <Paragraph position="0"> Before giving the algorithm formally, we illustrate an overview in this section, by using simple examples.</Paragraph>
    <Paragraph position="1"> Though the algorithm can be used to extract knowledge useful to resolve a wide range of syntactic ambiguities, we use here the prepositional phrase attachment problem as an illustrative example.</Paragraph>
    <Paragraph position="2"> We assume a syntactic parser which provides all possible analyses. It produces syntactic descriptions of sentences in the form of syntactic dependency structures. That is, the description to be produced is represented by a set oftuples like \[head word, syntactic relation, argument\], each of which expresses a dependency relation in the input. The syntactic relation in a tuple is either a grammatical relation like SUB J, OBJ, etc. (in case of a noun phrase) or a surface preposition like BY, WITH, etc. Following the normal convention of dependency representation, the argument is represented by the governor of the whole phrase which fills the argument position.</Paragraph>
    <Paragraph position="3"> When an input sentence has attachment ambiguities, two or more tuples share the same argument and the same syntactic-relation but have different head-words.</Paragraph>
    <Paragraph position="4"> For example, the description of the sentence &amp;quot;I saw a girl with a scarf.&amp;quot; contains two tuples like \[girl, WITH, scarf\] \[saw, WITH, scarf\] As repeatedly claimed in natural language understanding literature, in order to resolve this ambiguity, a system may have to be able to infer &amp;quot;a scarf cannot be used as an instrument to see&amp;quot;, based on extra-linguistic knowledge. A practical problem here is that there is no systematic way of accumulating such extra-linguistic knowledge for given subject fields. Furthermore, the ambiguity in a sentence like &amp;quot;I saw a girl with a telescope&amp;quot; cannot be resolved only by referring to knowledge about the world. It requires a full range of context understanding abilities, because the interpretation of &amp;quot;a girl with a telescope&amp;quot; is less likely in general but can be a correct one in certain contexts. That is, unless a system has a full range of contextual understanding abilities (which we think will be impossible in most application environments in the foreseeable future), it cannot reject either of the possible interpretations as &amp;quot;impossible&amp;quot;. The best a system can do, without full understanding abilities, is to select more plausible ones or reject less plausible ones. This implies that we have to introduce a measure by which we can judge plausibility of &amp;quot;interpretations&amp;quot;. The algorithm we propose computes such measures from a given sample corpus in a certain way. It gives a plausibility value to each possible tuple, based on the sample corpus. For example, the tuples (saw, WITH, scarf) and (girl, WITH, scarf) might be assigned 0.5 and 0.82 as their plausibility value, which would show (girl, WITH, scarf) to be more plausible than (saw, WITH, scarf) . This produced knowledge can be used to disambiguate interpretations of the sentence &amp;quot;I saw a girl with a scarf&amp;quot;.</Paragraph>
    <Paragraph position="5"> The algorithm is based on the assumption that the ontological characteristics of the objects and actions denoted by words (or linguistic expressions in general) and the nature of the ontological relations among them are exhibited, though implicitly, in sample texts.</Paragraph>
    <Paragraph position="6"> For example, nouns denoting objects which belong to the same ontological classes tend to appear in similar linguistic contexts (for example, in the same argument positions of the same or similar verbs). Or if an object (or an ontological class of objects) is &amp;quot;intrinsically&amp;quot; related to an action (like &amp;quot;telescope&amp;quot; to &amp;quot;see&amp;quot;), the word denoting the class of objects co-occurs frequently with the verb denoting the action. The co-occurrence would be more frequent than that of those whose ontological relations are rather fortuitous, like &amp;quot;girl&amp;quot; and &amp;quot;telescope&amp;quot;.</Paragraph>
    <Paragraph position="7"> Note that we talk about extra-linguistic &amp;quot;ontology&amp;quot; for the sake of explaining the basic idea behind the actual algorithm. However, as you will see, we do not represent such things as ontological entities in the actual algorithm. The algorithm counts frequencies of co-occurrences among words and calculates word distances which interpret such co-occurrences as contexts. Nor do we posit any dichotomy between &amp;quot;intrinsic&amp;quot; relations and &amp;quot;accidental&amp;quot; relations among actions and objects. Differences are quantitative, not qualitative. That is, co-occurrences of &amp;quot;girl&amp;quot; and &amp;quot;scarf&amp;quot; are more frequent than, for example, those of &amp;quot;pig&amp;quot; and &amp;quot;scarf&amp;quot;. The algorithm in this paper computes the plausibility value of hypothesis-tuples like (girl, WITH, scarf), (saw, WITH, scarf), etc., basically by counting frequencies of instance-tuples \[girl, WITH, scarf\], \[saw, WITH, scarf\], etc. generated from sample texts by a syntactic parser.</Paragraph>
  </Section>
class="xml-element"></Paper>