The Swarthmore College Senseval-3 system

The Swarthmore College SENSEVAL3 System Richard Wicentowski, Emily Thomforde and Adrian Packel Computer Science Department Swarthmore College Swarthmore, Pennsylvania, 19081, USA {richardw,ethomfo1,packel}@cs.swarthmore.edu Abstract This paper presents the Swarthmore College word- sense disambiguation system which was designed for the 2004 SENSEVAL3 competition. Our system participated in five tasks: the lexical sample tasks in Basque, Catalan, Italian, Romanian, and Spanish. For each task, a suite of supervised algorithms were combined using voting to form the final system. 1 Introduction The Swarthmore College system consisted of three supervised classifiers which were used to perform lexical ambiguity resolution in five languages. A nearest-neighbor clustering classifier, a na¨ıve Bayes classifier, and a decision list classifier were each trained on several permutations of the extracted fea- ture set, then the answers were joined using voting. The training data was limited to the labeled data provided by the organizers; no outside or unlabeled data was used. The systems presented in this paper were devel- oped by undergraduates as part of a class project at Swarthmore College. 2 Features Each of the supervised algorithms made use of the same set of features, extracted from only the labeled data provided to us by the task organizers. We used no unlabeled data. We used the tagged and lemma- tized data to extract the following features, which were the only features used in our system: • Bag-of-words and bag-of-lemmas • Bigrams and trigrams of words, lemmas, part- of-speech, and case (Basque-only) around the target word • Topic or code (Basque, Catalan and Spanish) In order to prevent individual features from dom- inating any individual system, we used up to eight permutations of the above mentioned features (de- pending on the language) for each of our classifiers. Catalan and Spanish provided fine-grained part- of-speech tags which we felt would lead to sparse data problems. To reduce this problem, for some feature sets we made the part-of-speech tags more coarse by simplifying the tags to include only the first or first two letters of the tag. 3 Systems The following systems were used to complete the Basque, Catalan, Italian and Romanian lexical sam- ple tasks. The Spanish lexical sample task was com- pleted before the other four tasks were begun and used only a subset of the systems presented below. Full details on the systems and methods used for the Spanish lexical sample task can be found in Sec- tion 7.3. See Section 4 for details on the classifier com- bination, and Section 5.2 for information about our use of bagging. 3.1 Cosine-based Clustering The first system developed was a nearest-neighbor clustering method which used the cosine similarity of feature vectors as the distance metric. A centroid was created for each attested sense in the training data, and each test sample was assigned to a cluster based on its similarity to the existing centroid. Cen- troids were not recalculated after each added test in- stance. 3.2 Na¨ıve Bayes The second system used was a na¨ıve Bayes classifier where the similarity between an instance, …

Content truncated to 500 words