We show an example of Iwanami Kokugo Jiten (Nishio et al, 1994), which is a dictionary used as a sense inventory. As shown in Figure 1, each en­try has POS information and definition sentences including example sentences.

We show an example of the given training data in (1). The given data are morphologically ana­lyzed and partly tagged with Iwanami's sense IDs, such as "37713-0-0-1-1" in(l).

This task includes 50 target words that were split into 219 senses in Iwanami; among them, 143 senses including two Xs that were not defined in Iwanami, appear in the training data. In the test data, 150 senses including eight Xs appear. The training and test data share 135 senses including two Xs; that is, 15 senses including six Xs in the test data are unseen in the training data.

We performed two preliminary pre-processing steps. First, we restored the base forms because the given training and test data have no informa­tion about the base forms. (1) shows an example of the original morphological data, and then we added the base form (lemma), as shown in (2).

Secondly, we extracted example sentences from Iwanami, which is used as a sense inventory. To compensate for the lack of training data, we an­alyzed examples with a morphological analyzer, Mecab UniDic version, because the training and test data were tagged with POS based on UniDic.

In preliminary experiments, we used both SVMand MEM (Nigam et al., 1999), with optimization method L-BFGS (Liu and Nocedal, 1989) to train the WSD model.

First, we investigated the effect between do­mains (pn, pb, and ow). For training data, we se­lected words that occur in more than 50 sentences, separated the training data by domain, and tested different domain combinations.

Table 1 shows the SVM results of the domain combinations. For Table 1, we did a 5-fold cross validation for the self domain and for comparison with the results after adding the other domain data. In Table 1, Diff. shows the differences to the self domain.

As shown in Table 1, for PN and OW, using other domains improved the results, but for PB, other do­mains degraded the results. So we decided to se­lect the domains for each target word.

In the formal run, for each pair of domain and target words, we selected the combination of do­main and dictionary examples that got the best cross-validation result in the training data. Note that in the case of no training data for the test data domain, for example, since no OCs have training data, we used all training data and dictionary ex­amples.

We show the number of selected domain combi­nations for each target domain in Table 2. Because the distribution of target words is very unbalanced in domains, not all types of target words appear in every domain, as shown in Table 2.

We also tried to predict new senses (x) that didn't appear in the training data by calculating the en­tropy for each target given in the MEM. We as­sumed that high entropy (when the probabilities of classes are uniformly dispersed) was indicative of X; i.e., if [entropy > threshold] => predict X; else => predict with MEM's output sense tag.

Note that we used the words that were tagged with Xs in the training data, except for the target words. We compared the entropies of x and not x of the words and heuristically tuned the thresh­old based on the differences among entropies. Our three official submissions correspond to different thresholds.