<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2214">
  <Title>Revising the WORDNET DOMAINS Hierarchy: semantics, coverage and balancing</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The continuous expansion of the multilingual information society with a growing number of new languages present on the Web has led in recent years to a pressing demand for multilingual applications. To support such applications, multilingual language resources are needed, which however require a lot of human effort to be built.</Paragraph>
    <Paragraph position="1"> For this reason, the development of language-independent resources which factorize what is common to many languages, and are possibly linked to the language-specific resources, could bring great advantages to the development of the multilingual information society.</Paragraph>
    <Paragraph position="2"> A language-independent resource, usable in many automatic and human applications, is represented by domain hierarchies. The notion of domain is related to similar notions such as semantic field, subject matter, broad topic, subject code, subject domain, category. These notions are used, sometimes interchangeably, sometimes with significant distinctions, in various fields such as linguistics, lexicography, cataloguing, text categorization. As far as this work is concerned, we define a domain as an area of knowledge which is somehow recognized as unitary. A domain can be characterized by the name of a discipline where a certain knowledge area is developed (e.g.</Paragraph>
    <Paragraph position="3"> chemistry) or by the specific object of the knowledge area (e.g. food). Although objects of knowledge and disciplines that study them are clearly related, the relation between these two points of view on domains is sometimes blurred and may be a source of uncertainty on their exact definition.</Paragraph>
    <Paragraph position="4"> Another interesting duality when speaking about domains is related to the fact that knowledge manifests itself in both words and texts. So the notion of domain can be applied both to the study of words, where a domain is the area of knowledge to which a certain lexical concept belongs, or to the study of texts, where the domain of a text is its broad topic. In this work we will assume that also these two points of view on domains are strictly intertwined.</Paragraph>
    <Paragraph position="5"> By their nature, domains can be organized in hierarchies based on a relation of specificity. For instance we can say that TENNIS is a more specific domain than SPORT, or that ARCHITECTURE is more general than TOWN PLANNING.</Paragraph>
    <Paragraph position="6"> Domain hierarchies can be usefully integrated into other linguistic resources and are also profitably used in many Natural Language Processing (NLP) tasks such as Word Sense Disambiguation (Magnini et al. 2002), Text Categorization (Schutze, 1998), Information Retrieval (Walker and Amsler, 1986).</Paragraph>
    <Paragraph position="7"> As regards the usage of Domain hierarchies in the field of multilingual lexicography, an example is given by the EuroWordNet Domain-ontology, a language independent domain hierarchy to which interlingual concepts (ILI-records) can be assigned (Vossen, 1998). In the same line, see also the SIMPLE domain hierarchy (SIMPLE, 2000).</Paragraph>
    <Paragraph position="8"> Large domain hierarchies are also available on the Internet, mainly meant for classifying web documents. See for instance the Google and Yahoo directories.</Paragraph>
    <Paragraph position="9"> A large-scale application of a domain hierarchy to a lexicon is represented by WORDNET DOMAINS (Magnini and Cavaglia, 2000). WORDNET DOMAINS is a lexical resource developed at ITC-irst where each WordNet synset (Fellbaum, 1998) is annotated with one or more domain labels selected from a domain hierarchy which was specifically created to this purpose. As the WORDNET DOMAINS Hierarchy (WDH) is language-independent, it has been possible to exploit it in the framework of MultiWordNet (Pianta et al., 2002), a multilingual lexical database developed at ITC-irst in which the Italian component is strictly aligned with the English WordNet. In MultiWordNet, the domain information has been automatically transferred from English to Italian, resulting in a Italian version of WORDNET DOMAINS. For instance, as the English synset {court, tribunal, judicature} was annotated with the domain LAW, also the Italian synset {corte, tribunale}, which is aligned with the corresponding English synset, results automatically annotated with the LAW domain. This procedure can be applied to any other WordNet (or part of it) aligned with Princeton WordNet (see for instance the Spanish WordNet).</Paragraph>
    <Paragraph position="10"> It is worth noticing that two of the main on-going projects addressing the construction of multilingual resources, that is MEANING (Rigau et al. 2002) and BALKANET (see web site), make use of WORDNET DOMAINS. Finally, WORDNET DOMAINS is being profitably used by the NLP community mainly for Word Sense Disambiguation tasks in various languages.</Paragraph>
    <Paragraph position="11"> Another application of domain hierarchies can be found in the field of corpus creation. In many existing corpora (see for instance the BNC, the ANC, the Brown and LOB Corpora) domain is one of the most used criteria for text selection and/or classification. Given that a domain hierarchy is language independent, if the same domain hierarchy is used to build reference corpora for different languages, then it would be easy to create (a first approximation of) comparable corpora by putting in correspondence corpora sections belonging to the same domain.</Paragraph>
    <Paragraph position="12"> An example of a corpus in which the complete representation of domains is pursued in a systematic way is represented by the MEANING Italian corpus, a large size corpus of written contemporary Italian in which a subset of the WDH labels has been chosen as the fundamental criterion for the selection of the texts to be included in the corpus (Bentivogli et al., 2003). Given the relevance of language-independent domain hierarchies for multilingual applications, it is of primary importance that these resources have a well-defined semantics and structure in order to be useful in various application fields. This paper reports the work done to improve the WDH so that it complies with such requirements. In particular, the WDH revision has been carried out with reference to the Dewey Decimal Classification.</Paragraph>
    <Paragraph position="13"> The paper is organized as follows. Section 2 briefly introduces the WORDNET DOMAINS Hierarchy and its main characteristics, with a short overview of the Dewey Decimal Classification system. Section 3 describes features and properties of the revision. Finally, in section 4, conclusions are reported.</Paragraph>
  </Section>
class="xml-element"></Paper>