Lexical taxonomies, in particular WordNet (Fellbaum, 1998), are now widely used in NLP for applications including semantic tagging, text categorisation and parsing (Harabagiu and Chai, 1998). WordNet consists of sets of lex­ical items (words and phrases) with similar meanings called synsets which are organised into a hyponomy (IS_A) hierarchy. For exam­ple, "doctor" and "physician" are in the same synset which is directly subsumed by "medi­cal practitioner". However the coverage of lex­ical semantics in WordNet is not comprehen­sive, items such as discourse information are not included. For example, "tennis player" (a hyponym of person) is not closely related to "racket", "balls" or "net" (hyponyms of arti­fact). Motivated by this example, Fellbaum (1998) dubbed this the "tennis problem". Tax­onomies omitting this information are ignoring potentially valuable information which could be helpful for applications such as information re­trieval (IR), word sense disambiguation, infor­mation extraction and parsing.

This paper reports a step towards a solution for the tennis problem by adding thesaural re­lations to the noun taxonomy in WordNet (ver­sion 1.6). The aim is to produce groups of noun synsets which are related by topic or domain. Once identified these links can be added to WordNet to denote this new form of lexical information, which is in addition to the existing hyponomy and hypernymy relations.

An example fragment of the WordNet hier­archy with the types of links we aim to add is shown in Figure 1. The existing relations, shown as unbroken and dashed lines, demon­strate the relative distance between items such as "ball boy" and "tennis ball". This link would be made explicit by the addition of the thesaural links which are shown as dotted lines.

I game equipment |--

This new information is obtained from the related words classes found in the CIDE+ ma­chine readable dictionary (Procter, 1995). The main problem with using these existing classes is that it is not clear which of the possible synsets in WordNet each word refers to since they are not disambiguated against WordNet. Therefore the problem reduces to disambiguat­ing sets of related nouns against the senses in the taxonomy. This problem is addressed by

combining existing disambiguation techniques and applying this approach to WordNet and CIDE+ (Section 2). An evaluation of this ap­proach is presented in Section 3. Some related work is described in Section 4.

The machine readable version of the CIDE+ dictionary contains thesaural information in the form of 1924 related word classes. Example cat­egories are Board Games (dice-1-0, square-1-3, shake-1-12, backgammon-1-0 ...) and Colours (silver-1-1, redness-1-0, black-l-22y amber-1-0 ...). Each member of the category relates to a particular CIDE+ sense which allows access to the textual definitions contained in the dic­tionary. This extra information is the main ad­vantage of using CIDE+ related word categories compared to a traditional thesaurus which con­sists only of groups of related words untagged for sense.

The first stage is to disambiguate each noun in CIDE+ thesaurus classes against Word-Net. Formally, assume that CIDE+ thesaurus class, N, contains the word W with senses {si,S2—sn}. We wish to assign a value to each possible WordNet sense for W indicating the likelihood of that sense being relevant given that W is a member of the noun group N. This is achieved using a combination of existing lexical similarity metrics which are now described.

So far we have described three metrics which could be used to disambiguate the sense in CIDE+ noun groups against WordNet synsets. An obvious next step would be to combine them. This is done by computing the sum and product of various combinations shown in Table 1. The combinations of metrics are computed from the value of the relevant combination for a particular sense normalised by the total value for that combination across all possible senses. Thus, for example, the value of the IC + CD metric combination for sense i is given by:

To evaluate the disambiguation method we re­quire some resource which lists the appropriate

WordNet sense for each of the members of the CIDE+ categories. To our knowledge no such resource exists and we are forced to adapt an existing resource. The most widely used and re­liable text tagged with WordNet senses is SEM-COR (Landes et al, 1998), a 200,000 word por­tion of the Brown corpus semantically tagged as part of the WordNet project. This corpus consists of 103 files on a wide variety of top­ics, each concerned with a particular subject. Content words are tagged with WordNet synset numbers but there is no direct way of telling which of these are related to CIDE+ senses. However, when semantically related nouns oc­cur together they do so with the related mean­ings. For example, if "monitor", "drive", "zip" and "screen" occur in the same text it is highly likely that they are all used with their senses re­lated to computer hardware. In addition nouns appearing in a text on a given topic are likely to be related to that topic. So the word "Java" is more likely to mean 'programming language' than 'coffee' or 'island' in a text about software development. These two observations about the behaviour of noun meanings can be used to au­tomatically derive disambiguated noun groups suitable for our evaluation from the SEMCOR texts. It is likely that in SEMCOR texts which contain a large portion of the types from a par­ticular thesaurus class those words will be used with the WordNet sense appropriate to that class.

Mandala et. al. (1999) combined three the-saurii to expand queries for an IR system. It was found that the combination of all three pro­duced better results than no query expansion or when a single resource was used. The na­ture of their application meant that there was no need to produce an explicit mapping between the senses of the three resources.

Agirre et. al. (2000) constructed topic sig­natures constructed from web searches to add extra information to WordNet. A test set of 20 nouns which occur at least 100 times in

SEMCOR was chosen. For each possible Word-Net sense a query was sent to the AltaVista search engine and the results stored. These documents were used to construct a topic sig­nature for each concept which were evaluated within a sense disambiguation algorithm and found to outperform information extracted di­rectly from WordNet. They produced further improvements when used to cluster senses.

Knight and Luk (1994) provided a mapping between WordNet and LDOCE by combining textual definitions with information about the hierarchical structure of the resources reporting 96% mapping accuracy. Green et al. (2001) use a combination of similarity metrics, includ­ing Resnik's, to map entries in a verb database onto WordNet senses, reporting 72% precision and 58% recall. This suggests that the approach described here may be useful for other gram­matical categories.

We have presented a method for overcoming the "tennis problem" in taxonomies such as Word-Net by adding new relations to the hierarchy obtained by disambiguating the noun groups found in existing thesaural classes. It was found that this can be achieved using a combination of existing disambiguation techniques. The tech­niques were evaluated using gold standard tag-gings derived automatically from SEMCOR.

I am extremely grateful to Philip Resnik for ad­vice on the re-implementation of the algorithm described in Section 2. This paper benefited from comments on earlier drafts provided by Yorick Wilks, George Miller and Tim Baldwin.