• word category information comes from Chi­nese thesaurus

I2R system used a semi-supervised classification algorithm (label propagation algorithm) (Niu, et al., 2005). They used three types of features: PoS of neighboring words with position information, un­ordered single words in topical context, and local collocations.

In the label propagation algorithm (LP) (Zhu and Ghahramani, 2002), label information of any vertex in a graph is propagated to nearby vertices through weighted edges until a global stable stage is achieved. Larger edge weights allow labels to travel through easier. Thus the closer the examples, the more likely they have similar labels (the global consistency assumption). In label propagation process, the soft label of each initial labeled exam­ple is clamped in each iteration to replenish label sources from these labeled data. Thus the labeled data act like sources to push out labels through unlabeled data. With this push from labeled exam­ples, the class boundaries will be pushed through edges with large weights and settle in gaps along edges with small weights. If the data structure fits the classification goal, then LP algorithm can use these unlabeled data to help learning classification plane.

CITYU-HIF system was a fully supervised one based on a Naïve Bayes classifier with simple fea­ture selection for each target word. The features used are as follows:

• Local features at specified positions:

• Topical features within a given window:

Content words appearing within w-10 to w10• Syntactic features:

One characteristic of this system is the incorpo­ration of the intrinsic nature of each target word in disambiguation. It is assumed that WSD is highly lexically sensitive and each word is best character­ized by different lexical information. Human judged to consider for each target word the type of disambiguation information if they found useful. During disambiguation, they run two Naïve Bayes classifiers, one on all features above, and the other only on the type of information deemed useful by the human judges. When the probability of the best guess from the former is under a certain threshold, the best guess from the latter was used instead.

SWAT system uses a weighted vote from three different classifiers to make the prediction. The three systems are: a Naïve Bayes classifier that compares similarities based on Bayes' Rule, a clas­sifier that creates a decision list of context features, and a classifier that compares the angles between vectors of the features found most commonly with each sense. The features include bigrams, and tri-grams, and unigrams are weighted by distance from the ambiguous word.

TorMd used an unsupervised naive Bayes classi­fier. They combine Chinese text and an English thesaurus to create a 'Chinese word'--'English category' co-occurrence matrix. This system gener­ated the prior-probabilities and likelihoods of a Naïve Bayes word sense classifier not from sense-annotated (in this case English translation anno­tated) data, but from this word--category co­occurrence matrix. They used the Macquarie The­saurus as very coarse sense inventory.

They asked a native speaker of Chinese to map the English translations of the target words to ap­propriate thesaurus categories. Once the Naïve Bayes classifier identifies a particular category as the intended sense, the mapping file is used to label the target word with the corresponding English translation. They rely simply on the bag of words that co-occur with the target word (window size of 5 words on either side).

HIT is a fully unsupervised WSD system, which puts bag of words of Chinese sentences and the English translations of target ambiguous word to search engine (Google and Baidu). Then they could get all kinds of statistic data. The correct translation was found through comparing their cross entropy.

Target I Sen I Train I Test I Base-

Together six teams participate in this WSD task, four of them adopt supervised learning methods and two of them used unsupervised algorithms. All of the four supervised learning systems exceed ob­viously the baseline obtained by the most frequent sense. It is noted that the performances of the first three systems are very close. Two unsupervised methods' scores are below the baseline. More unlabeled data maybe improve their performance.

Although the SRCB-WSD system got the high­est scores among the six participants, it does not perform always better than other system from table 2 and table 3. But to each word, the four super­vised systems always predict correctly more in­stances than the two un-supervised systems.

Besides the corpus, we provide a specification of the PoS tag set. Only SRCB-WSD system utilized this knowledge in feature selection. We will pro­vide more instances in the next campaign.