<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0608">
  <Title>Coreference in Knowledge Editing</Title>
  <Section position="3" start_page="57" end_page="58" type="metho">
    <SectionTitle>
2 WYSlWYM + indices
</SectionTitle>
    <Paragraph position="0"> How can a feedback text indicate coreference? One possibility is to let feedback texts make use of anaphoric Noun Phrases, such as St' or %his document'. In this paper we will assume that although output texts, which are meant to be read by a user who is not a domain expert, are expressions of some natural language, ,feedback texts may sometimes include artificial elements if this is necessary to avoid ambiguity. We will explore how the feedback language may be extended by the use of referential indices: If two NPs have the same index, they must have the same referent, whereas if they have different indices, they may or may not have the same referent. Brackets can be added to disambiguate the scope of indices in complex nominals. For example, the new feedback language can contain such expressions as (the name of (the document)3h and (the date of (the document)3)2, where the indices imply that the documents in the two expressions are the same. In the remainder of this section, we will explore in what different ways an author may want to control the indices in the feedback text when she inserts a new instance into the knowledge base, or when she cuts or copies an instance from the knowledge base.</Paragraph>
    <Section position="1" start_page="57" end_page="58" type="sub_section">
      <SectionTitle>
2.1 Which options for editing
</SectionTitle>
      <Paragraph position="0"> Imagine the author wants to expand an anchor of the form 'this object', specifying that it is a document. Imagine, furthermore, that two other documents have already appeared in the feedback text, namely (document)l and (document)2. Then the newly introduced object may corefer with (document)l or with (document)2 or with neither. In order to avoid asking potentially superfluous questions, the system can first present the user with a menu containing the two options I Existing object? New object? \] Only in case of the first choice will the system follow up with the menu</Paragraph>
      <Paragraph position="2"> In case of the second choice, the system will replace the anchor by some expression of the form (document)/, where i is a new index.</Paragraph>
      <Paragraph position="3">  In the current version of DRAFTER, cutting is a conceptually simple operation that does not allow variations: an attribute value is replaced by its anchor. When coreference enters the picture, this is no longer the case. In particular, there are two questions that we have to address when an indexed NP is cut (as we will say by an obvious extension of usage of the word 'cut'). Firstly, Does the author intend to cut this NP alone, or does she intend to cut all NPs with the same index as this NP? Secondly, if the author intends to cut all expressions with the same index, then Does the author intend the system to respect the indices? In other words, does she assume that the anchors that will result when the NP is de-selected must all be filled by NPs that have the same index? Depending on how these questions are answered, three different variants of the Cut operation arise. Let al be the NP on which the author has clicked. Then Cut-one only affects this occurrence of O~ i * Cut-all affects all occurrences of ai, also severing all coreference links between their anchors.</Paragraph>
      <Paragraph position="4"> Cut-all c affects all occurrences of ai, respecting all coreference links between their anchors.</Paragraph>
      <Paragraph position="5"> Consider the example in Section 1.2, adapted to the situation where docl is the value of two  To spell out the effect of the new operations, it will be convenient to use variables. If Cut-one is applied to the occurrence of docl in the owner attribute then the only effect on the content of the knowledge base is that the owner of namel is undefined, which may be represented by a variable, say x. If, instead, Cut-all is applied to the same instance of docl, then both the owner ofnamel and the actee of save1 are undefined, which may be represented by using two different variables, say x and y. If, finally, Cut-all c is applied then, once more, both are undefined, but their values must be equal. This may be represented by using the same variable, say x, in both cases: actee(savel, x), owner(namel, x).</Paragraph>
    </Section>
    <Section position="2" start_page="58" end_page="58" type="sub_section">
      <SectionTitle>
2.1.3 Copying
</SectionTitle>
      <Paragraph position="0"> Like cutting, copying allows variations now that the knowledge base can use one and the same instance (e.g., docl) as the value of different attributes (e.g. actee(savel, docl) and owner(namel, docl))). There are at least two options. One option amounts to a faithful copy of an instance. This option can be implemented simply by letting a buffer point to the instance that is copied. The other option (called Replicate) duplicates the original feedback text, while renumbering all the indices in such a way that all the indices in the replica are new, respecting equalities between indices in the origina/. The second option arises when an author wants to reuse a part of the feedback text to refer to a new instance. Both editing operations are relatively straightforward to understand and implement.</Paragraph>
    </Section>
    <Section position="3" start_page="58" end_page="58" type="sub_section">
      <SectionTitle>
2.2 Example
</SectionTitle>
      <Paragraph position="0"> Imagine an author wanting to create the procedure given in Section 2.1.2, starting out with the incomplete representation presented in Section 1.2, where it is not yet specified what namel is the name of. The content of the knowledge base can be reflected by the feedback text Save document/by entering the name of this object (further actions), where i is an arbitrary number. Suppose the author opens a pop-up menu on the anchor 'this object' and chooses the option 'Existing object'. Note that, owing to some obvious type constraints, the only instance in the existing knowledge base that can take the place of 'this object' is the object docl. Consequently, there is no need for the system to ask further questions.</Paragraph>
      <Paragraph position="1"> A slightly more round-about way in which the same potentially complete knowledge base can come about is as follows. Confronted with the just-presented menu, she decides to choose 'New object'. The system then adds the statements ovner (namel, doe-j).</Paragraph>
      <Paragraph position="2"> document (doc-j).</Paragraph>
      <Paragraph position="3"> to the knowledge base, where doc-j is a constant that has not occurred before. The following feedback text may be generated: Save document/by entering the name of documentj (further actions).</Paragraph>
      <Paragraph position="4"> Seeing her mistake, the author can then click on either 'document/' or 'document/and select Cut. This will cause the NP in question to be replaced by the anchor 'this object', which brings the author back to a situation where she can decide that the two documents should corefer after all.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="58" end_page="59" type="metho">
    <SectionTitle>
3 Some problems
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="58" end_page="59" type="sub_section">
      <SectionTitle>
3.1 Limitations of the use of indices
</SectionTitle>
      <Paragraph position="0"> Indices can be used to indicate, for two text spans, whether or not they are assumed, by the author, to refer to the same thing. This implies a number of problems familiar from the problem of tagging text corpora (e.g. Hirschman et al.</Paragraph>
      <Paragraph position="1"> 1997). Thus, it is unclear how indices should be used in relation to NPs occurring in intensional contexts. For example, it seems that the bracketed NPs in the following feedback text should  receive different indices, even though their extension is equal: Make sure that (the date of the program) equals (the current date).</Paragraph>
      <Paragraph position="2"> Furthermore, it can be tempting to view an NP as anaphoric, even if it is impossible to name any one text span as its antecedent. For example, this happens when the role of an antecedent is played by a combination of several NPs (e.g., the NP 'a file' and the NP 'a program' can together form the antecedent of the plural pronoun 'they') or when the antecedent is implied, rather than directly mentioned, by the text (e.g., in the case of bridging anaphora). In these cases, it is difficult to see how indices can be used. Note, however, that the problematic phenomena mentioned here may not pose a serious problem in the context of the present work, since we axe dealing with the construction of feedback texts (as opposed, for example, to the output texts generated by the system) which tend to make a simplified use of language.</Paragraph>
      <Paragraph position="3"> (cf. Section 2).</Paragraph>
    </Section>
    <Section position="2" start_page="59" end_page="59" type="sub_section">
      <SectionTitle>
3.2 Embedded NPs
</SectionTitle>
      <Paragraph position="0"> Suppose that the author has reached the following feedback text Save (the document)l by entering (the name of (the document)l)2 and entering (the name of (the document)l)2 (.further actions).</Paragraph>
      <Paragraph position="1"> with the intention of changing the second occurrence of 'document' into 'directory'. To avoid removing both occurrences of 'the document' she chooses the Cut-one option rather than Cutall. null Save (the document)l by entering (the name of (the document)l)2 and entering (the name of this object)2 (further actions).</Paragraph>
      <Paragraph position="2"> Now she clicks on the anchor this object and chooses 'directory'. What effect should this operation have on the noun phrase in which the anchor is embedded? There are two possibilities. Either the index of the embedding NP should remain the same: (the name of (the directory)3)2 or it should change, implying that the two names are distinct: (the name of (the directory)3)4 In this example, it seems clear that a name can belong only to one object, so that the latter result should be preferred. However, in some cases the index on the embedding phrase might plausibly remain the same after the embedded phrase has been changed. For instance, if we copy 'the menu containing the Save option', then cut 'the Save option' from the second occurrence of the phrase, replacing it by 'the Print option', there is no reason why the Save and Print options should not belong to the same menu. Thus to react appropriately to the Cut-one operation, the system may have to apply domain knowledge or seek guidance from the author.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>