<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2125">
  <Title>ALGORITHM FOR AUTOMATIC INTERPRETATION OF NOUN SEQUENCES</Title>
  <Section position="6" start_page="786" end_page="787" type="evalu">
    <SectionTitle>
5. TEST, RESULTS AND I)ISCUSSION
</SectionTitle>
    <Paragraph position="0"> The results that arc under discussion were obtained on the basis of semantic information which was automatically extracted from Longman I)ictionary of Contemporary English (I~I)OCE) as described in Montemagni and Vanderwende (1992) 1 . The semantic information has not been altered in any way fl'onl its automatically derived form, and so there are still errors: for the 94,()00 attribute clusters extracted fl'om nearly 75,000 single noun and verb definitions in L1)()CE, we estimate the accuracy to be 78%, with a margin of error of +/- 5% (see Richardson et al., 1993).</Paragraph>
    <Paragraph position="1"> A training corpus of 100 NSs was collected \[rein lhc examples of NSs in lhe 1)revious literature, to ensure that all known classes of NSs are handled in this system. These results were expected to be good because these NSs were used to develop the rules and their weights. The system successfully identified the most likely interpretation for 79 of the 100 NSs (79%). Of the remaining 21 NSs, tile most plausible interpretation was alnong the possible interpretations 8 times, (8 %), and no interpretation at all was given for 4 NSs (4 %).</Paragraph>
    <Paragraph position="2"> The test corpus consisted of 97 NSs from the tagged version of the Brown corpus (I;rancis and Kucera, 1989), to ensure the adequacy of applying this approach to unrestricted test; the results for an expanded test corpus will be reported in Vanderwende (in preparation). Tbe system currently identified successfully II1e most likely inlerpretation lor 51 of the 97 NSs (52%).</Paragraph>
    <Paragraph position="3"> ()1' the remaining 46 NSs, the most likely interpretation was presented second for 21 NSs IAlthough 1,1)OCE includes somc semantic information in the form of box codes and subject codes, these were not used in this system. This approach is designed t() work with semantic information from any dictionary.</Paragraph>
    <Paragraph position="4"> (22 %); when first and second interpretations were considered, the system was successful approximately 74% of the time. A wrong or no interpretation was given for 25 NSs (26 %1). Upon examination of these results, several areas for improvement arc suggested. First is to improve lhe semantic information: Dolan et al. (1993) describes a network of semantic information, given not only the definition of the Icxical entry but also all of the other definitions which have a labeled relation to that entry.</Paragraph>
    <Paragraph position="5"> Secondly, while the NS classification proposed in Vanderwende (1993) proves to be adequate for analyzing NSs in tmrestrictcd text, an additional 'What about?' class, suggested in Levi (1978), may be justified. In the current classification schema, NSs such as cigare.tte war and history confi, rence have been considered 'Whom/what?' NSs given tim verbs that are associated with the head, fight and eonJer/talk about respectively. In unrestricted text, similar NSs are quite fi'equent, for example university policy, prevention program, care plan, but the definitions of the heads do not always specify a related verb. The bead definitions for policy, program and plan, however, do allow a IIAS-TOPIC semantic feature to be identified, and this IIAS-TOPIC can be used to establish a 'What about?' interpretation.</Paragraph>
    <Paragraph position="6"> Applying this algorithm to previously tmseen text also produced a very promising result: the verbs that are associated with nouns in their definitions (i.e., role nominals in lqnin, 1980) are being used often and correctly to produce NS interpretations. While some rules had been developed to handle obvious cases in the training corpus, how often the conditions on these rules would be met could not be predicted. In fact, such NS interpretations are frequent. For example, the NS wine cellar is analyzed as a 'What for?' relation with a high score, and the system provides as a paraphrase: cellar w,~ich is for storing wine, given the definition of cellar (1, n, 1): 'an undergrotmd room, ttsu. used for storing goods; basement'. This result is promising for two reasons: first, by analyzing the definitions (and later also the example sentences) in an on-line dictionary, we now have access to a nonhandcoded source of semantic information which includes the verbs and their lelation to the notms, essential for determining role nominals. Second, the related verbs are used to construct the paraphrases of a NS, and doing so makes a general interpretation such as 'What lot?' more  specific, e.g., a service office is not an office for service, but an office for ~ service, and a vegetable market is not a market for vegetables, but a market for buying and selling vegetables.</Paragraph>
    <Paragraph position="7"> Enhancing the general interpretations with the related verb(s) approximates at least in part Downing's observation that the types of relations that can hold between the nouns in a NS are possibly infinite (Downing, 1977).</Paragraph>
  </Section>
class="xml-element"></Paper>