At the corpus preprocessing phase, we tokenized a corpus and then tagged it with parts-of-speech using Brill's Tagger(Brill, 1994). The tokenizer just sepa­rates symbols from a word. For example, a sentence "I'm straight, white, no longer middle class, anti-IRA, have ..." is tokenized to "J'm stright , white , no longer middle class , anti - IRA , have Un­like other symbols, an apostrophe is not separated from the following characters.

At the phrase filtering phase, we filtered senses using the satellite feature, which is marked with sat tag in training and test corpus given by the task organizer. For example, in a sentence This air of disengagement < head sats="carry-over. 067:0"> carried< /head> <sat id=" carry-over. 067:0">over</sat> to his apparent attitude toward his things, carried over is a phrase and also a satellite feature.

Phrase filtering is applied to sense disambiguation as in Table 1

Table 1: phrase filtering and sense disambiguation

if the number of filtered senses = 1 then determine sense

else if the number of filtered senses > 1 then execute sense-tagger with the filtered senses

else if the number of filtered senses — 0 then execute sense-tagger with all senses

There are satellite features in the English lexical sample, but not in the Korean lexical sample. Hence, phrase filtering was applied only in the English lex­ical sample task.

The CIM is a kind of classification model based on the entropy theory. Given an input instance, the CIM decides the proper class of the instance by con­sidering individual decisions made by each feature of the instance. In the model, the proper class of an instance,X, is determined by Equation 1.

where class j is the j-th class and Kel(classj, X) is the relevance between the j-th class and the instance X. Here, if we assume that features are independent of each other, the relevance can be defined as in Equation 2.

where m is the size of the feature set, X{ is the value of the 2-th feature and Wij is the weight of the i­th feature for the j-th class. In Equation 2, X{ has a binary value (1 if the feature occurs within the window, 0 otherwise) and is defined in terms of classification information.

The classification information of a feature is com­posed of two components. One is the discrimination score (DS), which represents the discrimination abil­ity of classifying instances. The other is the most probable class (MPC), which represents the most closely related class to the feature. Wij is defined by using these two components as follows:

def Win —

DS; if class j = MPC, 0 otherwise

In Equation 3, DSi and MPd represent the DS and MPC of the i-th. feature, respectively. In the CIM, DS and MPC are defined in terms of the con­ditional probability of a class given a feature, which is normalized by the corpus size. The normalized conditional probability is defined as follows:

def Pji =

p(classj\fi

N(classj

p(fi\classj)

In Equation 4, pji is a normalized conditional probability, N(classj) is the number of instances belonging to the j-th class in the training data, N(class) is the average number of instances for each class and n is the number of classes. Given the normalized conditional probability distribution, DSs and MPCs are defined as follows:

arg max pji

In Equation 5, H (pi) is the entropy of the i-th feature over the normalized conditional probability distribution.

The CIM has a problem caused by using MPCs, which is information loss. For example, let us con­sider the situation in Table 2 and Table 3. Table 2 shows the normalized conditional probability distri­bution, DSs and MPCs of features in an instance. Table 3 shows the weights and the relevance values at the CIM using Wij and at the modified CIM us­ing w^, for the instance of Table 2. The feature f\ co-occurred with class\ and class2 and the MPC of fi is classi at Table 2. In the CIM, this feature

Table 3: The weights and the relevance values at the CIM using and at the modified CIM using w^, for the instance of Table 2 contributes to only classi. Actually the feature }\ can contribute to distinguishing class2 from classs if it consults the normalized conditional probability distribution. In the CIM, however, the feature can not distinguish them because their weights have the same value.

Another aspect of the problem is that the CIM fails to capture the minor contribution of features, which is crucial in the case where the sum of the minor contribution of features to a non-MPC class dominates that of the major contribution of fea­tures to MPC classes. For example, at Table 2, all features, /i, f2, and fs, have different MPCs: classi, classs and class^, respectively, it is also ob­vious that they have some minor contribution to the class2. The CIM will classify the instance as classi because Rel(class\,X) — 1.1187 is the maximum number among the Rel(classj, X). However, if we consider the minor contribution of all the features, we prefer class2 to classi because class2 intuitively gains the total contribution more than clas$\.

A solution to the problem may be not to use MPCs, but to use a measure of contribution of a feature to a class which is proportional to the dis­crimination score of the feaure and the normalized conditional probability of the class given the feature. The modified CIM can be defined as follows:

Wij d= DSi x Pji (8) As shown in Table 3, the w\2 is larger than wis (0.3356 > 0) and the instance is classified not as classi but as class2 because Rel(class2,X) =

1.0028 > Rel(class!,X) = 0.7831, which is based on the modified CIM.

We used three types of features for word sense dis­ambiguation: local, topical and bigram context. In the preliminary experiment, we have observed that, when the CIM considered all these three types of features, it mostly achieved the best result.