<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1007">
  <Title>Abstract Anaphora Resolution in Danish</Title>
  <Section position="8" start_page="61" end_page="62" type="evalu">
    <SectionTitle>
6 Evaluation of the Algorithm
</SectionTitle>
    <Paragraph position="0"> I have applied the modified ES99-algorithm to three randomly selected SL dialogues (6,305 words) and to one of the dialogues between native Danes recorded in the PID collection (5,367 words). It must be noted that in my test only one annotator (the author) identiffed dialogue acts, classified the anaphors in the dialogues, marked NPs and anaphor antecedents. In (Eckert and Strube, 1999a) these tasks have been accomplished by two annotators.</Paragraph>
    <Paragraph position="1"> In dividing the three SL dialogues into discourse segments I have mainly used a partition made by two researchers at the University of Copenhagen in an independent project. The discrimination criteria were topic shift and a few linguistic clues. I have then ap-IsI have followed the cache model described in (Walker, 1998). In the present test it was necessary to go back maximally seven SUs to find an antecedent to an individual pronominal anaphor.</Paragraph>
    <Paragraph position="2">  plied the same discrimination criteria to the dialogue from the PID collection.</Paragraph>
    <Paragraph position="3"> I have defined dialogue units syntactically following (Eckert and Strube, 1999a). 14 Because it is not always possible to distinguish between den, det, de used as personal or demonstrative pronouns without having access to stress information, I have classified them as personal pronouns unless they are topicalized, or occur in syntactic constructions where demonstratives are normally used. The manual classification of pronouns and demonstratives in the four dialogues can be found in table 4.</Paragraph>
    <Paragraph position="4"> The results of the individual anaphora resolution algorithm can be found in table 5, while the results of the discourse deictics resolution algorithm are given in table 6.</Paragraph>
    <Paragraph position="5"> The results obtained are better than those reposed in (Eckert and Strube, 1999a), but I have used more background information than ES99 and extended the scope of resolution for individual anaphors (without this extension the precision of the individual resolution algorithm was of 64.5). Furthermore the Danish deictic det occurs in more contexts than the English it, this and that, thus there are more I* predicates in the Danish version of the algorithm than in the original one. The fact that only one annotator divided the dialogues into SUs may also have influenced the results.</Paragraph>
    <Paragraph position="6"> The algorithm classifies anaphors and resolves some of them, thus there are two types of error, classification errors and resolution errors. Most of the instances of wrongly classified anaphors are due to the fact that the algorithm classifies vague anaphors as discourse deictics and then resolves the anaphor to a preceding predicate or clause. Few errors are due to the fact, already noticed by ES99, that the defined I* and A* predicates do not contain information about nominals referring to abstract objects. 15 These errors resulted in most cases in resolution errors.</Paragraph>
    <Paragraph position="7"> Some errors are due to the inability to find 14The dialogue collections have been tagged.</Paragraph>
    <Paragraph position="8"> 15The semantic lexicon I used did not contain the relevant nominals.</Paragraph>
    <Paragraph position="9"> an individual NP antecedent to the pronoun det, when this refers generally to an NP of different gender 16 and to wrongly resolved plural pronouns with complex NP antecedents or with no antecedent. Correctly classified, but wrongly resolved discourse deictics are, i.a., due to the fact that I did not mark in any particular way parenthetical utterances. The latter kind of errors are chaining errors. In table 7 the occurrences of each type of error are reported.</Paragraph>
  </Section>
class="xml-element"></Paper>