We use the method described in (McCarthy et al., 2004) for finding predominant senses from raw text. The method uses a thesaurus obtained from the text by parsing, extracting grammatical relations and then listing each word (w) with its top k nearest neighbours, where k is a constant. Like (McCarthy et al., 2004) we use k = 50 and obtain our thesaurus using the distributional similarity metric described by (Lin, 1998) and we use WordNet (WN) as our sense inventory. The senses of a word w are each assigned a ranking score which sums over the dis­tributional similarity scores of the neighbours and weights each neighbour's score by a WN Similarity score (Patwardhan and Pedersen, 2003) between the sense of w and the sense of the neighbour that max­imises the WN Similarity score. This weight is nor­malised by the sum of such WN similarity scores be­tween all senses of w and the senses of the neighbour that maximises this score. We use the WN Similarity jcn score (Jiang and Conrath, 1997) since this gave reasonable results for (McCarthy et al., 2004) and it is efficient at run time given precompilation of fre­quency information. The jcn measure needs word frequency information, which we obtained from the British National Corpus (BNC) (Leech, 1992). The distributional thesaurus was constructed using sub­ject, direct object adjective modifier and noun mod­ifier relations.

We contrasted ranking of the original WordNet senses with ranking produced using the coarse grained mapping between WordNet senses and the clusters provided for this task. In the first, which we refer to as fine-grained training (SUSSX-FR), we use the original method as described in section 2 using WordNet 2.1 as our sense inventory. For the second method which we refer to as coarse-grained train­ing (sussx-cr), we use the clusters of the target word as our senses. The distributional similarity of each neighbour is apportioned to these clusters us­ing the maximum WordNet similarity between any of the WordNet senses in the cluster and any of the senses of the neighbour. This WordNet similarity is normalised as in the original method, but for the de­nominator we use the sum of the WordNet similarity scores between this neighbour and all clusters of the target word.

The topic domain of a document has a strong influ­ence on the sense distribution of words. Unfortu­nately, it is not feasible to produce large manually sense-annotated corpora for every domain of inter­est. Since the method described in section 2 works with raw text, we can specialize our sense rank­ings for a particular topic domain, simply by feed­ing a domain specific corpus to the algorithm. Pre­vious experiments have shown that unsupervised es­timation of the predominant sense of certain words using corpora whose domain has been determined by hand outperforms estimates based on domain-independent text for a subset of words and even outperforms the estimates based on counting oc­currences in an annotated corpus (Koeling et al., 2005). A later experiment (using SENSEVAL2 and 3 data) showed that using domain specific predomi­nant senses can slightly improve the results for some domains (Koeling et al., 2007). However, a firm idea of when domain specilisation should be considered could not (yet) be given.

Coarse Disambiguation of coarse-grained senses is obviously an easier task than ine grained training. We had hoped that the coarse-grained training might show superior performance by removing the noise created by related but less frequent senses. Since the mapping between ine-grained senses and clus­ters is used anyway in the scorer the noise from related senses does not seem to be an issue. Re­lated senses are scored correctly. Indeed the per­formance of the ine-grained training is superior to that of the coarse-grained training. We believe this is because predominant meanings have more related senses. There are therefore more chances that the distributional similarity of the neighbours will get apportioned to one of the related senses when there are more related senses. The coarse grained rank­ing would have an advantage on occasions when in the ine-grained ranking the credit between related senses is split and an unrelated sense ends up with a higher ranking score. Since the coarse-grained ranker lumps the credit for related sense together it would be at an advantage. Clearly this doesn't hap­pen enough in the data to outweigh the beneicial effect of the number of related senses compensating for other noise in the data.

Domain Unfortuately, the system specialised for domain (SUSSX-C-WD) did not improve the results over the 5 documents significantly. However, if we look at the contributions made by each document, we might learn something about the relation beteen the output of the classiier and the impact on the WSDresults. Table 2 shows the per-document results for the systems SUSSX-FR and SUSSX-C-WD. The irst two documents show very little difference with the domain independent results. The documents 'd004' and 'd005' show a small but clear improved performance for the domain results. Unfortunately, document 'd003' displays a very disappointing drop of more than 3% in performance, and cancels out all the gains made by the last two documents.

The output of the classiier seems to be indica­tive of the results for all documents except 'd003'. The classifier doesn't seem to find enough evidence for a marked preference for a particular domain for documents 'd001' and 'd002'. This could be an indication that there is no strong domain ef­fect to be expected. The strong preference for the 'computer_science' domain for 'd004' is reflected in good performance of SUSSX-C-WD and even though the conidence scores for the irst 2 alternatives of 'd005' are fairly close, there is a clear drop in con-idence score for the third alternative, which might indicate that the topic of this document is related to both first choices of the classifier. It will be interest­ing to evaluate the results for 'd005' using the 'Art' sense rankings. One would expect those results to be similar to the results found here. Finally, the results for 'd003' are hard to explain. We will need to do an extensive error analysis as soon as the gold-standard is available.

In this paper we investigated two directions where we expect potential for improving the performance of our method for acquiring predominant senses. In order to fully appreciate what the effects of the coarse grained sense inventory are (i.e. whether some of the more unfortunate errors are resolved), we will have to do an extensive error analysis as soon as the gold standard becomes available. Con­sidering the fairly low number of attempted tokens (only 72.8% of the tokens are attempted), we are at a disadvantage compared to systems that back-off to (for example) the first sense in WN. However, we are well pleased with the high precision (71.7%) of the method SUSSX-FR, considering this is a completely unsupervised method. There seems to be potential gains for domain adaptation, but applying it to each document does not seem to be advisable. More re­search needs to be done to identify in which cases a performance boost can be expected. Five documents is not enough to fully investigate the matter. At the moment we are performing a larger scale experiment with the documents in SemCor. These documents seem to cover a fairly wide range of domains (ac­cording to our text classiier) and many domains are represented by several documents.

This work was funded by UK EPSRC project EP/C537262 "Ranking Word Senses for Disambiguation: Models and Applications", and by a UK Royal Society Dorothy Hodgkin Fellowship to the second author. We would also like to thank Piek Vossen for giving us access to the Irion Technolo­gies text categoriser.