The task of word sense disambiguation consists of assigning the most appropriate meaning to a poly-semous word within a given context. Applications such as machine translation, knowledge acquisition, common sense reasoning, and others, require knowl­edge about word meanings, and word sense disam­biguation is considered essential for all these tasks.

Most of the efforts in solving this problem were concentrated so far toward targeted supervised learning, where each sense tagged occurrence of a particular word is transformed into a feature vector, which is then used in an automatic learning process. The applicability of such supervised algorithms is however limited only to those few words for which sense tagged data is available, and their accuracy is strongly connected to the amount of labeled data available at hand.

Instead, methods that address all words in unre­stricted text have received significantly less atten­tion. While the performance of such methods is usu­ally exceeded by their supervised lexical-sample alternatives, they have however the advantage of pro­viding larger coverage.

In this paper, we describe SuperSense­Learner - a system for solving the semantic am­biguity of all words in unrestricted text. Super­SenseLearner brings together under one system the features previously used in the SenseLearner (Mihalcea and Csomai, 2005) and the SuperSense (Ciaramita and Altun, 2006) all-words word sense disambiguation systems. The system is using a rel­atively small pre-existing sense-annotated data set for training purposes, and it learns global semantic models for general word categories.

Our goal is to use as little annotated data as possi­ble, and at the same time make the algorithm gen­eral enough to be able to disambiguate as many content words as possible in a text, and efficient enough so that large amounts of text can be anno­tated in real time. SuperSenseLearner is at­tempting to learn general semantic models for var­ious word categories, starting with a relatively small sense-annotated corpus. We base our experiments on SemCor (Miller et al., 1993), a balanced, se-mantically annotated dataset, with all content words manually tagged by trained lexicographers.

The input to the disambiguation algorithm con­sists of raw text. The output is a text with word meaning annotations for all open-class words.

The algorithm starts with a preprocessing stage, where the text is tokenized and annotated with partof-speech tags; collocations are identified using a sliding window approach, where a collocation is de­fined as a sequence of words that forms a compound concept defined in WordNet (Miller, 1995).

Next, a semantic model is learned for all pre­defined word categories, where a word category is defined as a group of words that share some com­mon syntactic or semantic properties. Word cate­gories can be of various granularities. For instance, a model can be defined and trained to handle all the nouns in the test corpus. Similarly, using the same mechanism, a finer-grained model can be defined to handle all the verbs for which at least one of the meanings is of type e.g., "<move>". Finally, small coverage models that address one word at a time, for example a model for the adjective "small," can be also defined within the same framework. Once de­fined and trained, the models are used to annotate the ambiguous words in the test corpus with their corre­sponding meaning. Sections 3 and 4 below provide details on the features implemented by the various models.

Note that the semantic models are applicable only to: (1) words that are covered by the word category defined in the models; and (2) words that appeared at least once in the training corpus. The words that are not covered by these models (typically about 10­15% of the words in the test corpus) are assigned the most frequent sense in WordNet.

Different semantic models can be defined and trained for the disambiguation of different word cat­egories. Although more general than models that are built individually for each word in a test corpus (Decadt et al., 2004), the applicability of the seman­tic models built as part of SenseLearner is still limited to those words previously seen in the train­ing corpus, and therefore their overall coverage is not 100%.

Starting with an annotated corpus consisting of all the annotated files in SemCor, augmented with the Senseval-2 and Senseval-3 all-words data sets, a separate training data set is built for each model. There are seven models provided with the current SenseLearner distribution, implementing the following features:

A great deal of work has focused in recent years on shallow semantic annotation tasks such as named entity recognition and semantic role labeling. In the former task, systems analyze text to detect mentions of instances of coarse-grained semantic categories such as "person", "organization" and "location". It seems natural to ask if this type of shallow seman­tic information can be leveraged to improve lexical disambiguation. Particularly, since the best perform­ing taggers typically implement sequential decoding schemes, e.g., Viterbi decoding, which have linear complexity and can be performed quite efficiently. In practice thus, this type of pre-processing resem­bles pOS-tagging and could provide the WSD sys­tem with useful additional evidence.

For the final system we create a combined feature set for each target word, consisting of the lemma, the part of speech, the collocational SenseLearner features, and the three coarse grained semantic tags of the target word. Note that the semantic fea­tures are represented as lemma .TAG to avoid over-generalization.

In the training stage, a feature vector is con­structed for each sense-annotated word covered by a semantic model. The features are model-specific, and feature vectors are added to the training set pertaining to the corresponding model. The label of each such feature vector consists of the target word and the corresponding sense, represented as word#sense. Table 1 shows the number of feature vectors constructed in this learning stage for each semantic model. To annotate new text, similar vec­tors are created for all the content-words in the raw text. Similar to the training stage, feature vectors are created and stored separately for each semantic model.

Next, word sense predictions are made for all the test examples, with a separate learning process run for each semantic model. For learning, we are using the Timbl memory based learning algorithm (Daelemans et al., 2001), which was previously found use­ful for the task of word sense disambiguation (Hoste et al., 2002; Mihalcea, 2002).

Following the learning stage, each vector in the test data set is labeled with a predicted word and sense. If the word predicted by the learning algo­rithm coincides with the target word in the test fea­ture vector, then the predicted sense is used to an­notate the test instance. Otherwise, if the predicted word is different from the target word, no annota­tion is produced, and the word is left for annotation in a later stage (e.g., using the most frequent sense back-off method).

The SuperSenseLearner system participated in the SemEval all-words word sense disambigua­tion task. Table 1 shows the results obtained for each part-of-speech (nouns and verbs), as well as the overall results. We have also ran a separate evaluation excluding the u (unknown) tag, which is shown in Table 2. SuperSenseLearner was ranked the third among the fourteen participating systems, proving the validity of the approach.