USYD: WSD and Lexical Substitution using the Web1T corpus

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 446–453, Prague, June 2007. c⃝2007 Association for Computational Linguistics USYD: WSD and Lexical Substitution using the Web1T Corpus Tobias Hawker School of Information Technologies University of Sydney NSW 2006, Australia toby@it.usyd.edu.au Abstract This paper describes the University of Syd- ney’s WSD and Lexical Substitution sys- tems for SemEval-2007. These systems are principally based on evaluating the substi- tutability of potential synonyms in the con- text of the target word. Substitutability is measured using Pointwise Mutual Informa- tion as obtained from the Web1T corpus. The WSD systems are supervised, while the Lexical Substitution system is unsuper- vised. The lexical sample sub-task also used syntactic category information given from a CCG-based parse to assist in verb disam- biguation, while both WSD tasks also make use of more traditional features. These related systems participated in the Coarse-Grained English All-Words WSD task (task 7), the Lexical Substitution Task (task 10) and the English Lexical Sample WSD sub-task (task 17). 1 Introduction This paper describes closely related systems that were applied to three tasks of the SemEval-2007 workshop. The unifying characteristic of these sys- tems is that they use the same measure of ‘substi- tutability’ for a given word and a surrounding con- text to perform the tasks. This measure is based on frequencies involving the word and the context from n-gram counts derived from one trillion words of Web text. These systems participated in the English Coarse- Grained All Words and English Lexical Sample Word Sense Disambiguation (WSD) tasks, and in the Lexical Substitution task. The Lexical Substitution system relies entirely on the substitutability measure to rank potential syn- onyms, and only uses manual sense inventories to preferentially select words which have been iden- tified by lexicographers as being synonyms for the original word in some contexts. It does not make use of any machine learning, and is thus unsupervised. The WSD systems are supervised, using a Sup- port Vector Machine (SVM) to learn from sense- tagged examples of ambiguous words and predict the class of the test instances. Classifiers for both systems use a small number of additional feature types beyond those derived from the n-gram counts, including Bag of Words (BOW) and local context features. A single separate model was trained for each ambiguous lemma. For verbs in the lexical sample, the classifier also uses the syntactic category assigned to the target verb by a parser as additional information for dis- ambiguation. The remainder of this paper is organised as fol- lows. Relevant background for the ideas employed is briefly discussed, as is the nature of the Web1T corpus. Descriptions of the particular systems used for each of the tasks are described in ascending or- der of task number. Details of particular sources of information and the methods used to capture them are introduced along with the task they are used in. A presentation of results and discussion follows the description of each system, and overall conclusions 446 are presented at …

Content truncated to 500 words