Complex, Corpus-Driven, Syntactic Features for Word Sense Disambiguation

Complex, Corpus-Driven, Syntactic Features for Word Sense Disambiguation Ari Chanen and Jon Patrick Sydney Language Technology Research Group School of Information Technologies University of Sydney Sydney, Australia, 2006 � ari,jonpat � @it.usyd.edu.au Abstract Although syntactic features offer more specific information about the context surrounding a target word in a Word Sense Disambiguation (WSD) task, in general, they have not distinguished themselves much above positional fea- tures such as bag-of-words. In this pa- per we offer two methods for increas- ing the recall rate when using syntac- tic features on the WSD task by: 1) using an algorithm for discovering in the corpus every possible syntactic fea- ture involving a target word, and 2) us- ing wildcards in place of the lemmas in the templates of the syntactic features. In the best experimental results on the SENSEVAL-2 data we achieved an F- measure of 53.1% which is well above the mean F-measure performance of of- ficial SENSEVAL-2 entries, of 44.2%. These results are encouraging consider- ing that only one kind of feature is used and only a simple Support Vector Ma- chine (SVM) running with the defaults is used for the machine learning. 1 Introduction: Syntactic Features The best features for machine learning classifi- cation are the ones that have the most discrimi- natory power, for the task at hand. This paper will be discussing the use of syntactic features (SF’s) in word sense disambiguation (WSD.) In WSD, the task is to choose (or classify) the cor- rect sense of the target word (the word whose sense is to be disambiguated) given the surround- ing text. One type of feature that is commonly used in WSD classification systems is called bag- of-words. Bag-of-words features are rather infor- mation poor, only specifying the presence or ab- sence of words in the target word’s context. SF’s, on the other hand, are much richer in information. Not only do syntactic features have information on the presence of words in the context but they also include information about the syntactic relation- ships that hold between the target word and con- text words in the same sentence as the target. In order to use SF’s, a syntactic parser is needed that produces a parse tree for every training corpus sentence. The parse tree gives the syntactic rela- tionships between words in each sentence. Con- nexor parser (J¨arvinen and Tapanainen, 1997) was used to annotate the data with syntactic relation- ships in the research presented here. In this re- search, a SF is defined as a connected group of words from a parse tree that must include the tar- get word. The SF includes information on each of its words, the syntactic relationships between them, and information on how each word relates to the others in the tree hierarchy. The SF word information includes the word, lemma, and part- of-speech (POS) for each word. The use of syntactic features in WSD might seem to be a more effective discriminatory feature compared to a information poor feature like …

Content truncated to 500 words