<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-1301">
  <Title>Resolving Bridging References in Unrestricted Text</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Bridging Descriptions
</SectionTitle>
    <Paragraph position="0"> Linguistic and computational theories of bridging descriptions identify two main subtasks involved in their resolution: first, finding the element in the text to which the bridging description is related (ANCHOR) and second, finding the relation (LINK) holding between the bridging description and its anchor (Clark, 1977; Sidner, 1979; Helm, 1982; Carter, 1987; Fraurud, 1990; Strand, 1997).</Paragraph>
    <Paragraph position="1"> A speaker is licensed to use a bridging DD when he/she can assume that the common-sense knowledge required to identify the relation is shared by the listener (Hawkins, 1978; Clark and Marshall, 1981; Prince, 1981). This reliance on commonsense knowledge means that, in general, a system could only resolve bridging references when supplied with an adequate knowledge base; for this reason, the typical way of implementing a system for resolving bridging references has been to restrict the domain and feeding the system with hand-tailored world knowledge. (This approach is discussed in detail in (Carter, 1987)). In order to get a system capable of performing on unrestricted text, we decided to use WordNet (WN) (Miller, 1993) as an approximation of a knowledge base containing generic information, and to supplement it with heuristics to handle those cases which WN couldn't handle.</Paragraph>
    <Paragraph position="2"> Vieira and Teufel (1997) analyzed the corpus to identify the cases of bridging descriptions that could be resolved using WordNet, those for which we could use heuristics, and those that couldn't be interpreted at the moment. Six classes of bridging descriptions were identified. 2 Synonymy/Hyponymy/Meronymy This class (henceforth, Syn/Hyp/Mer) includes those DDs which are in a synonymy/hyponymy/meronymy relation with their anchors, i.e., the kind of semantic relation that is currently encoded in WordNet. Examples 3 are: (1) a. Synonymy: new album -- the record; three bills -- the legislation.</Paragraph>
    <Paragraph position="3"> b. Hypernymy-Hyponymy: rice -- the plant; the daily television show -- the program.</Paragraph>
    <Paragraph position="4"> c. Meronymy (part-of relation): plants -- the pollen; house -- the chimney.</Paragraph>
    <Paragraph position="5"> Names This class includes definite descriptions that refer back to proper names such as people's and company names, as in: (2) Bach -- the composer; Pinkerton's Inc -- the company.</Paragraph>
    <Paragraph position="6"> Compound Nouns This class includes bridging descriptions whose LINGUISTIC ANCHOa (i.e., the element in the text to which they are related) is a noun occurring as part of a compound noun other than the head. Examples include: (3) stock market crash -- the markets; discount packages -- the discounts.</Paragraph>
    <Paragraph position="7"> Events These are cases where the linguistic anchors of DDs are not NPs but VPs or sentences.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Examples are:
</SectionTitle>
      <Paragraph position="0"> as bridging even though the relation is of co-reference.</Paragraph>
      <Paragraph position="1">  (4) Individual investors and professional money managers contend. -- They make the argument ...; Kadane Oil Co. is currently drilling two wells and putting money into three others. -- The activity ...</Paragraph>
      <Paragraph position="2"> Discourse Topic There are some cases of DDs which are related to the (often implicit) discourse topic (in the sense of (Reinhart, 1981)) of a text, rather than to some specific NP or VP. For instance, (5) the industry (in a text whose discourse topic is oil companies); the first half(in a text whose  discourse topic is a concert).</Paragraph>
      <Paragraph position="3"> Inference We collect in this class all the cases of bridging descriptions whose relation with their NP anchor was based on more complex inferential relations: for example, cases in which the relation between the anchor and the DD was of reason, cause, consequence, or set-membership:  (6) last week's earthquake -- the suffering people are going through; Democratics/Republicans -- the two sides.</Paragraph>
      <Paragraph position="4">  The relative importance of these classes in our corpus is shown in Table 1. This classification is based on what we took to be the main linking relation for each of the 204 bridging DDs in the corpus 4.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Resolution of Bridging
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Descriptions
</SectionTitle>
      <Paragraph position="0"> We used Vieira and Teufel's analysis as the basis for the implementation of a second prototype. In this section we discuss how this prototype handles the different types of bridging descriptions.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Bridging DDs and WordNet
</SectionTitle>
      <Paragraph position="0"> We implemented a WordNet interface that reports a possible semantic link between two nouns when one 4One problem with bridging references is that they are often related to more than one antecedent in the discourse (Poesio and Vieira, 1997).</Paragraph>
      <Paragraph position="1"> of the following is true: * The nouns are in the same synset (= synonyms of each other), as in suit -- lawsuit.</Paragraph>
      <Paragraph position="2"> * The nouns are in direct hyponymy relation with each other, for instance, dollar -- currency.</Paragraph>
      <Paragraph position="3"> * There is a direct or indirect meronymy between them. Indirect meronymy holds when a concept inherits parts from its hypernyms, like car inherits the part wheel from its hypernym wheeled_vehicle.</Paragraph>
      <Paragraph position="4"> * Due to WordNet's idiosyncratic encoding, it is often necessary to look for a semantic relation between coordinate sisters, i.e. hyponyms of the same hypernym, such as home -- house which are hyponyms of housing, lodging.</Paragraph>
      <Paragraph position="5"> Sometimes, ifa relation between two head nouns is not encoded in WN directly, the semantic closeness might be found through the compound nouns made up of them. Thus, for a pair such as record, album we find synonymy between record_album and album. We ran a test in which WordNet was used to identify the DD's anchors. For each of the 204 bridging DDs in our corpus, we considered the NPs in the previous five sentences as a potential anchor 5, and queried WN with the DD and potential anchor pair. WordNet reported a possible relation for 107 of the 204 DDs. Often, more than one link was found between a DD and a potential anchor, 6 and sometimes links could be found with more than one potential anchor in the previous five sentences. Only in 34 of these 107 cases we found at least one appropriate anchor linking relation, and only in 21 cases we found only appropriate anchors (for 13 there was a mixture of suitable and unsuitable anchors).</Paragraph>
      <Paragraph position="6"> Of these 34 DDs for which a correct anchor was found, only 18 were among those we had classified as Syn/Hyp/Mer. In 8 of these 18 cases, WordNet found a link with an anchor that was not the one we had identified manually, but which was still valid; for instance, we identified the link pound -- the currency which was not found in WordNet, whereas our automatic search found sterling -- the currency).</Paragraph>
      <Paragraph position="7"> 5Our system does not currently include a proper segmentation algorithm. Instead, we use a simple recency heuristic--we only consider the antecedents in the n previous sentences, where n is a constant determined empirically. In our previous work (Vieira and Poesio, 1997). we observed that 5 was the value of n which gave the best tradeoff between precision and recall.</Paragraph>
      <Paragraph position="8">  be in a hypernym relation and were also reported as coordinate sisters having as common hypernymy commerce. The 16 remaining relations were found for DDs that we had not classified as Syn/Hyp/Mer: for instance, whereas we had classified the DD the population as belonging to the class of DDs based on names having as anchor Koreans, our implementation found a WordNet relation for the pair nation -- the population; the system also found a few relations with proper names, such as Bach -- the composer.</Paragraph>
      <Paragraph position="9"> In the following tests we have considered only direct meronymy, as indirect meronymy presented extreme low recall and precision at a very expensive computational cost.</Paragraph>
      <Paragraph position="10"> In order to reduce the number of false positives (86 out of 107) 7, we tried using a stack-based approach towards finding potential anchors in the previous sentences, as suggested in (Sidner, 1979); i.e., the system would go back one sentence at a time, and stop as soon as a relation with a potential anchor was found, rather than trying to find all possible links. As a result, we found exactly one correct anchor for 30 DDs, slightly improving our results (76 false positives).</Paragraph>
      <Paragraph position="11">  mantic relations between DDs and the anchors found by our stack-based search. It presents precision figures (P) related to each type of relation s. Sister relations are the least satisfactory ones.</Paragraph>
      <Paragraph position="12"> We tested, in particular, whether WordNet encoded a semantic link between the 38 syn/hyp/mer relations in our corpus (just described) plus other 32 relations extracted from a second corpus study (Poesio and Vieira, 1997). The results for the 70 DDs are summarised in Table 3. Overall recall (R) was 46%.</Paragraph>
      <Paragraph position="13"> We could have expected 100% precision, since we had manually identified the anchors, but the precision figures (P) report an error when a sister relation is found instead of the expected (syn/hyp/mer) re7The anchors found for 73 of the DDs were incorrect, for the remaining 13 DDs the resulting anchors were mixed (some right/some wrong).</Paragraph>
      <Paragraph position="14"> SWe cannot estimate recall since we do not have a precise number of syn/hyp/mer anchors that should be found (as different types of anchors may allow resolution of bridging descriptions).</Paragraph>
      <Paragraph position="15"> lation. The poorest recall was surprisingly obtained for synonymy relations, followed by meronymy relations, as expected, since these are only partially implemented in WordNet.</Paragraph>
      <Paragraph position="16">  The low recall for synonymy relations may be due to the context dependent, specialized senses of sublanguage terminology (for instance, crash, bust and slump in Economics terminology). Some nouns were not even encoded in WN (such as newsweekly, spino3~). Other relations were missed due to the unexpected way in which knowledge is organised in WordNet. For example, no association was found between house and walls, because house is not encoded in WordNet as a hyponym of building but of housing, and housing does not have a meronymy link to wall whereas building does. Another example of counter-intuitive position in the hierarchy is that of catastrophe, not listed as a hypernym of earthquake, but as its coordinate sister. This example demonstrates the problems that WordNet lexicographers faced when they had to coerce real-world concepts into a hierarchy with single, no-default inheritance: catastrophes clearly subclassify into earthquakes and floods, but minor earthquakes don't have to be catastrophic.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Proper names
</SectionTitle>
      <Paragraph position="0"> Definite descriptions which refer back to proper names are very common in Wall Street Journal articles. Processing such DDs involves, first, determining an entity type for each name in the text, then searching for semantic relations. If we get the entity type person for the a name such as Mrs. Y.J. Park we could, ideally, resolve the subsequent DD the housewife using WordNet. A few names are available in WordNet: typically, famous people, countries, states, cities and languages. Other entity types can be identified using appositive constructions and abbreviations like Mr., Co., Inc. etc. as cues.</Paragraph>
      <Paragraph position="1"> The algorithm we developed, based on a mixture of access to WordNet and heuristics such as those we described, found the correct type for 66% of the names in our corpus (535/814). Including a back-tracking mechanism which re-processes a text looking for missing name types (with this mechanism we identify the type for the name Morishita in a textual sequence like Morishita -- Mr. Morishita) increases our recall to 69% (562/814). We then used WordNet to match the types found with previous references in the text. This resulted in the resolution of 53% (26/49) of the cases based on names. We missed relations which are not found in WordNet (for instance, Mr. Morishita -- the 57 year-old). But again we also found a large number of false positives.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 Compound Nouns
</SectionTitle>
      <Paragraph position="0"> Sometimes, a bridging description is linked to a non-head noun in a compound noun: (7) stock market crash -- the markets; rule changes -- the rules; discount packages ~ the discounts.</Paragraph>
      <Paragraph position="1"> One way of processing these definite descriptions would be to update the discourse model with discourse referents not only for the NP as a whole, but also for the embedded nouns: for example, after processing stock market crash, we could introduce a discourse referent for stock market, and another discourse referent for stock market crash. The DD the markets would be co-referring with the first of these referents (with identical head noun), and we could simply use our anaphora resolution algorithms. This solution, however, makes available discourse referents that are generally unaccessible for pronominal anaphora. We therefore followed a different route: our algorithm for identifying antecedents attempts to match not only heads with heads, but also the head of a DD with the pre-modifiers of a previous NP, the pre-modifiers of a DD with the pre-modifiers of its antecedents, and the pre-modifiers of the DD with the head of a previous NP. With this, we obtained a total of 54 resolutions of which 36 were correct, although we did not always find a main linking relation licensing the use of a DD.</Paragraph>
      <Paragraph position="2"> Examples of correct resolutions are: (8) a. Head of DD with pre-modifier of antecedent: the stock market crash -- the markets; rule changes -- the rules; b. Pre-modifiers of DD with pre-modifiers of antecedent: most oil companies -- the oil fields; his art business -- the art gallery; c. Pre-modifiers of DD with head of antecedent: New ~brk City -- the city council district lines; a 15-acre plot and main home - the home site.</Paragraph>
      <Paragraph position="3"> There are also cases in which the pre-modifiers plus the head noun of a DD may indicate a bridging reference: we may find an antecedent with the same head noun for a DD but referring to a different entity, this being signalled by the pre-modification. Some examples: (9) the company's abrasive segment -- the engineering materials segment; Italy's unemployment rate -- the southern unemployment rate; Pinkerton -- the new Pinkerton; increases of 3.9 ~ -- the actual wage increases may have been bigger.</Paragraph>
      <Paragraph position="4"> Our previous heuristics for treatment of pre-modifiers in anaphoric resolution handled the first two examples correctly (Vieira and Poesio, 1997): as they present different pre-modifiers we did not treat them as anaphoric in the first version of our system. Such cases, as well as DDs modified by new and actual (last two examples), may now be treated as bridging references 9.</Paragraph>
    </Section>
    <Section position="5" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.4 Events
</SectionTitle>
      <Paragraph position="0"> To process DDs based on events (situations or propositions), we are trying, as a first approach, to transform verbs into their nominalizations, and then looking for a relation in WordNet. Some nominalizations can be generated by general procedures or learned by means of a stochastic method: e.g., we could use WordNet's morphology component as a stemmer, and augment the verbal stems with the most common suffixes for nominalizations which could be kept in a list, like -ment, -ion. In our corpus, 17% (7/40) of the bridging references based on events are direct nominalizations of this type (for instance, changes were proposed - the proposals).</Paragraph>
      <Paragraph position="1"> Another 17% are cases in which knowledge of the semantics of the verb is necessary (as in borrowed the loan). The remaining 66% (26 cases) of bridging DDs based on events require inference reasoning based on the compositional meaning of the phrases (as in It went looking for a partner - pitching the prospect); these cases are out of reach just now, as well as the cases listed under &amp;quot;discourse topic&amp;quot; and &amp;quot;inference&amp;quot;.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Restrictive Post-modification as
Anchors
</SectionTitle>
    <Paragraph position="0"> Whereas the problem of finding the appropriate textual anchor for bridging descriptions requires knowledge inference and reasoning, DDs with restrictive post-modification give the reader both anchor and description in the same expression. As Hawkins points out (Hawkins, 1978):  sociative clauses are not essentially different from the other uses mentioned in the last section (Hawkin's associative uses/our bridging uses). The only difference is that.</Paragraph>
    <Paragraph position="1"> in the latter uses set identification and tile locatability of the referent were possible on account of previous triggers, whereas it is now function of the modifier itself to provide the information which makes set identification and location possible(parentheses ours).</Paragraph>
    <Paragraph position="2"> This could be seen as an advantage, as we directly find the anchors of these DDs. They are quite common uses of DDs, but not much attention has been devoted to them as a special case of anchor linking relations. The main problem for these cases is to find out their links, which is also a remaining problem for our proposals of anchor identification throughout t he paper.</Paragraph>
    <Paragraph position="3"> The head noun of a DD and its modifier may be related in different ways, as shown by the examples: (10) the number of job seekers; the anthers of the plant; the ideal of a level playing field; the flip side of the Stoltzman personality.</Paragraph>
    <Paragraph position="4"> In some cases, the relation between noun and complement seem to be looser than the relations for bridging descriptions. Sequences such as the laws of heredity; the cost of the plan are acceptable, whereas heredity -- the laws; the plan -- the cost are unlikely to occur. On the other hand, bridging such as the house -- the kitchen; the firm -- the owners are as acceptable as the kitchen of the house; the owners of the firm. Some proposals of a systematic treatment for the identification of anchor linking relations for bridging DDs are (Heim, 1982; Barker, 1991; Poesio, 1994; Strand, 1997).</Paragraph>
  </Section>
class="xml-element"></Paper>