SWAT-MP:The SemEval-2007 Systems for Task 5 and Task 14

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 308–313, Prague, June 2007. c⃝2007 Association for Computational Linguistics SWAT-MP: The SemEval-2007 Systems for Task 5 and Task 14 Phil Katz, Matthew Singleton, Richard Wicentowski Department of Computer Science Swarthmore College Swarthmore, PA {katz,msingle1,richardw}@cs.swarthmore.edu Abstract In this paper, we describe our two SemEval- 2007 entries. Our first entry, for Task 5: Multilingual Chinese-English Lexical Sam- ple Task, is a supervised system that decides the most appropriate English translation of a Chinese target word. This system uses a combination of Na¨ıve Bayes, nearest neigh- bor cosine, decision lists, and latent seman- tic analysis. Our second entry, for Task 14: Affective Text, is a supervised system that annotates headlines using a predefined list of emotions. This system uses synonym expan- sion and matches lemmatized unigrams in the test headlines against a corpus of hand- annotated headlines. 1 Introduction This paper describes our two entries in SemEval- 2007. The first entry, a supervised system used in the Multilingual Chinese-English Lexical Sample task (Task 5), is an extension of the system described in (Wicentowski et al., 2004). We implement five dif- ferent classifiers: a Na¨ıve Bayes classifier, a decision list classifier, two different nearest neighbor cosine classifiers, and a classifier based on Latent Seman- tic Analysis. Section 2.2 describes each of the in- dividual classifiers, Section 2.3 describes our clas- sifier combination system, and Section 2.4 presents our results. The second entry, a supervised system used in the Affective Text task (Task 14), uses a corpus of head- lines hand-annotated by non-experts. It also uses an online thesaurus to match synonyms and antonyms of the sense labels (Thesaurus.com, 2007). Section 3.1 describes the creation of the annotated training corpus, Section 3.2 describes our method for assign- ing scores to the headlines, and Section 3.3 presents our results. 2 Task 5: Multilingual Chinese-English LS This task presents a single Chinese word in context which must be disambiguated. Rather than asking participants to provide a sense label corresponding to a pre-defined sense inventory, the goal here is to label each ambiguous word with its correct English translation. Since the task is quite similar to more traditional lexical sample tasks, we extend an ap- proach used successfully in multiple Senseval-3 lex- ical sample tasks (Wicentowski et al., 2004). 2.1 Features Each of our classifiers uses the same set of context features, taken directly from the data provided by the task organizers. The features we used included: • Bag-of-words (unigrams) • Bigrams and trigrams around the target word • Weighted unigrams surrounding the target word The weighted unigram features increased the fre- quencies of the ten words before and after the tar- get word by inserting them multiple times into the bag-of-words. 308 Many words in the Chinese data were broken up into “subwords”: since we were unsure how to han- dle these and since their appearance seemed incon- sistent, we decided to simply treat each subword as a word for the purposes of creating bigrams, trigrams, …

Content truncated to 500 words