<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0613">
  <Title>Nominal Metonymy Processing</Title>
  <Section position="5" start_page="95" end_page="97" type="metho">
    <SectionTitle>
4. Metonymy Processing: An Example
</SectionTitle>
    <Paragraph position="0"> For the sentence Lynn drives a Saab, the semandc constraint for the appropriate slot of the appropriate sense of the verb drive would be *ENGINE-PROPELLED-VEHICLE. Yet the potential filler Saab is of type (or a subtype of) *FOR-PROFIT-MANUFACTURING-CORPORATION which is a violation of the constraint. The ontological concept *FOR-PROFIT-MANUFACTURING-CORPORATION has a slot PRODUC-ER-OF, which has an &amp;quot;inverse&amp;quot; relation called PRODUCED-BY. The path which is found by the ontological  search process is (expressed in the \[SOURCE-NODE OUTGOING-ARC --&gt; DESTINATION</Paragraph>
    <Paragraph position="2"> If FOR-PROFIT-MANUFACTURING-CORPORA~ON417 were a concept in the maned endty inventory (with knowledge about Saab Scania AB), i.e., with slot/fillers such as (NAME $SAAB ), (PRODUCER-OF *AUTOMOBILE *JET-AIRCRAFt, the above path would be found. But even if that world knowledge tidbit (about Saab's products) were not available, the path that the ontological search process finds is:</Paragraph>
    <Paragraph position="4"> The latter path has a lesser preference (i.e., a greater cost) than the former, because of the more expensive traversed arcs (SUBCLASSES is always more expensive than IS-A), but illustrates that the mechanism is still able to identify the metonymy in the absence of the specific product information.</Paragraph>
    <Paragraph position="5"> Once a path is found (let's assume the latter no-named-entity-inventory case), it is inspected for the appearance of a metonymic relation arc. If such an arc is found, the inverse of that arc is available in constructing the final meaning representation of the sentence. For the above example, the most specific information that is available from the path (identifiable by following SUBCLASSES arcs after the metonymie arc) is utilized in making an inference about the replaced metonym and instantiating an appropriate concept %ENGINE-PROPELLED-VEHICLE460 (the TMR is our interlingua or meaning representation language): THR:</Paragraph>
    <Paragraph position="7"> The inference notation used in this example is more generally available to represent inferences made by a variety of specialized mechanisms or microtheories during the course of semantic analysis. This notation preserves the original literal interpretation, while making available the replaced entity that was inferred to exist by the metonymy processing mechanism; this inferred information (in this case, the existence of a produced vehicle) satisfies the goals of metonymy resolution mentioned above.</Paragraph>
    <Paragraph position="8">  The real challenge to this approach is when the system has no information about the word Saab at all. As a system heuristic, one of the most likely possibilities for an unrecognized word in noun position (particularly if we utilize the English capitalization information) is that it is a name for some named entity (i.e.,  This path, albeit expensive, is found by the search algorithm; the challenge of this approach is to adjust all of the arc weights to return these weights with fairly low cost relative to other returned paths.</Paragraph>
  </Section>
  <Section position="6" start_page="97" end_page="99" type="metho">
    <SectionTitle>
5. Inventory of Metonymie Relations
</SectionTitle>
    <Paragraph position="0"> Although not receiving nearly as much attention in the literature as metaphor, there have been a few attempts in the various literatures to categorize metonyrny into types. None of the inventories are comprehensive enough to support the population of a working ontology for use in the analysis of real-world texts.</Paragraph>
    <Paragraph position="1"> Thus the strategy used by us to build such an inventory consisted of combining multiple sources in the literature, experiments and analysis of corpora, and carefully filtering inventories of other kinds of semantic relations (e.g., syntagrnatic and paradigmatic lexical relations, meaning change, cognitive meronyrnic classification) for relations that do reflect metonymic use of language in English.</Paragraph>
    <Paragraph position="2"> As mentioned above, it is not possible to build an exhaustive inventory of metonyrny. So although this inventory was compiled for the purpose of seeding the metonymy processing mechanism, it is augmented with the mechanism for handling novel or unexpected (i.e., uninventoried) metonymic relations and combinations (chains) of metonymic relations.</Paragraph>
    <Paragraph position="3"> We built an inventory of metonymy types based on various sources, spanning theoretical linguistics, lexicography, cognitive science, philosophy of language, and computational linguistics, not necessarily dealing explicitly with metonymy: Apresjan (1974), Fass (1986), Kamei and Wakao (1992), Lakoff and Johnson (1980), Mel'chuk and Zholkovsky (1984), Nunberg (1978), Stem (1965), Winston et al. (1987), Yamanashi (1987). Our inventory consists of about 20 major categories, with another 20 subtypes.</Paragraph>
    <Paragraph position="4"> We encountered (in various corpora) some examples which seem to fall into multiple categories: The White House announced that.., could be either Symbol-for-Thing-symbolized or Place-for-Occupants.</Paragraph>
    <Paragraph position="5"> There is also group of alternations that reflect a semantic relation that could be arguably treated as either metonymy, regular polysemy (i.e., handled by Lexical Rules in our format or by generative processes in Pustejovsky (1995)), or derivational processes, such as Product-for-Plant or Music-for-Dance.</Paragraph>
    <Paragraph position="6"> We need to ensure that the metonymies in the inventory mentioned above are representable by relations in the ontology, with certain metonymies weeded out for lack of productivity (often because there is only a limited possibility of examples of the metonymy, and those are diachronically lexicalized). For each metonymic relation, we identify a relation that is used in the ontology to represent the relation between the referent and the metonym (i.e., from the thing being replaced to the thing that replaces it), along with an inverse relation (which is what actually appears in the path in a filler-to-constraint search).</Paragraph>
    <Paragraph position="7"> A potential problem with this approach is that triggering conditions may differ from the canonical metonymy, where a selectional restriction violation is a clear indicator that some kind of relaxation is neces- null 1. Numerous such Name Tagging systems, with accuracy very near human, have been evaluated in the scope of the Message Understanding Conferences (MUCs) and are described in Sundheim (1995).</Paragraph>
    <Paragraph position="8">  sary. In particular, there might not be any selectional restriction violation for some &amp;quot;pragmatic&amp;quot; metonymies, such as I'm going to spend money this a.~ernoon (which, arguably, are actually metaphors).</Paragraph>
    <Paragraph position="9"> 6. Knowledge Base for Metonymy Processing The knowledge required for processing metonymy is not specifically differentiated from the constraint satisfaction data requirements of the overall processing mechanism. Those static knowledge resources do, however, reflect ontology arcs and weights that are used for identifying and resolving metonymy. The knowledge acquisition consisted first of identifying the arcs that needed special treatment because they are used in resolving frequently-occurring metonymies, then second by setting weights for those arcs by the automated training mechanism (using simulated annealing). The latter part of the task, however, required manually building a training data set.</Paragraph>
    <Paragraph position="10"> The example below illnstrates the training data. The example from the corpus is quoted, followed by an enumeration of the metonymy categories in effect in the example. The matrix verb is the source of constralnts on the metonym in this case, so the concept is listed, along with the constraint that it places on the AGENT role. The path given in this example needs to be matched by the ontological graph search exactly.</Paragraph>
    <Paragraph position="12"> The training process for the weight assignment mechanism simply produces a weight for each of the arcs represented in a manually-produced inventory of arcs, mostly reflecting the arcs (actually, the second of each pair) identified in the inventory mentioned above. In our experiment, the arc types that receive special weights are manually specified, and the training mechanism assigns weights. It would have been possible for the training mechanism to assemble the list of arcs, as well, by examining the arcs reflected in the training data; one drawback of the latter approach would have been the inability to call out specific arcs that aren't used in the training data, in expectation of their occurrence in other corpora.</Paragraph>
    <Paragraph position="13"> First we constructed a data set which essentially reflects an opportunistic collection of metonymies, and is in no way exhaustive or reflective of the distribution of metonymies over a corpus. Weights were produced by a simulated annealing training process; the training was able to produce a set of weights that accounted for 100% of the training data. A typical set of such weights is abbreviated below:  The last line reflects the weight used for all arcs not explicitly inventoried.</Paragraph>
    <Paragraph position="14"> A second training set was produced more systematically from English-language newswire, specifically  the February 9-11 1997 edition of USA Today (bardcopy) and the February 11 1997 edition of the on-line edition of the Mercury News.</Paragraph>
    <Paragraph position="15"> After the ontology was augmented as required, new weights were produced by simulated annealing. The annealing run used the same annealing schedule and Cauchy cooling rate, and began by initially &amp;quot;heating&amp;quot; the temperature (by 10 complete randomizing annealing iterations) to an energy of 0.97 (in the interval \[0.0, 1.0\]). The simulated annealing run resulted in final energy of 0.0575, or 94.25% example accuracy (percentage of example sentences correctly analyzed, as compared to metonymic llnk accuracy, where examples with a chain of multiple metonymies count multiple times). Of the remaining errors (i.e., metonymic relations not found by the ontological search program), one is unsolvable by the current approach. The example, Eddie Jones had a hot hand in today's game has no selecfional constraint violations (and is, in fact, understandable and incorrectly acceptable literally). 1 Handling this type of non-literal expression is beyond the scope of the work described here, and would require a substantially different approach. null Of the other four examples that were not solvable after training, one is actually ambiguous, and the ontological search mechanism suggested a reading not supported by context: Fufimori tom Peruvian radio that.., appeared in a context which suggested that he talked to the nation via radio, vs. talking to the people in charge of the Peruvian radio service, as the ontological search program suggested. Two of the other examples, Other dinners brought in more money and The dinner is adding to the questions being asked about fund-raising activities, were incorrectly analyzed as using &amp;quot;dinner&amp;quot; to refer to the people who prepared the dinner, not the people who attended the dinner (in the former case); in is unclear how to analyze the latter of these, which is complicated by ellipsis, so there is no correct answer given in the training data, resulting in an automatic failure. 3 The last of the incorrect examples .... will move people from welfare rolls into jobs, also involves some metaphorical or elliptical mechanism. 4</Paragraph>
  </Section>
  <Section position="7" start_page="99" end_page="100" type="metho">
    <SectionTitle>
7. Results
</SectionTitle>
    <Paragraph position="0"> A test set was produced in exactly the same way as the training set described above, from USA Today and Mercury News articles (7 March 1997 editions). The test data in Table lreflect the first fifty metony- null Errors due to Errors due to Errors due to # Correct missing arc representation gap bad path 47/50- 94% 0/3 1/3 213 mies found in the two sources (actually, many repeat metonymies of the form X announced... X also announced.., were omitted; the inclusion of all these (easy) metonymies would have resulted in a ratio of about 95/100 for the test set). The table shows results on this test set using weights produced by training on both the training sets described above.</Paragraph>
    <Paragraph position="1"> The texts used for training and testing for the Spanish WSD experiments (see Mahesh et al. 1997) were also examined for metonymies produced as part of the semantic analysis process. The results there showed, realistically, how metonymy resolution, WSD, and building semantic dependency structure (to</Paragraph>
  </Section>
class="xml-element"></Paper>