Investigating Features for Classifying Noun Relations

Investigating Features for Classifying Noun Relations Dominick Ng, David J. Kedziora, Terry T. W. Miu and James R. Curran School of Information Technologies University of Sydney NSW 2006 Australia {dong7223,dked9370,tmiu9098,james}@it.usyd.edu.au Abstract Automated recognition of the semantic re- lationship between two nouns in a sentence is useful for a wide variety of tasks in NLP. Previous approaches have used ker- nel methods with semantic and lexical evi- dence for classification. We present a sys- tem based on a maximum entropy classifier which also considers both the grammatical dependencies in a sentence and significance information based on the Google Web 1T dataset. We report results comparable with state of the art performance using limited data based on the SemEval 2007 shared task on nominal classification. 1 Introduction Analysis of the semantics of natural language is an area of research undergoing renewed interest, driven by the many applications which can di- rectly benefit from such information, including question answering and text summarization. In particular, recognition of the relationship between words in a sentence is useful for clarifying ambi- guities in tools that attempt to interpret and re- spond to natural language. Recent developments in the semantic web – a vision where information is comprehensible to machines as well as people – have also accelerated the need for tools which can automatically analyse nominal relationships. We approach the task of relation classification using the relation set and training data provided by the SemEval 2007 shared task on nominal clas- sification (Girju et al., 2007). The problem as defined by the task description is to discover the underlying relationship between the concepts ex- pressed by two nominals, excluding named enti- ties. The relationship is informed by the context of an English sentence, e.g. it is clear that the relationship between door and car differs in the fragments the car door and the car scraped the door of the garage. Resolving the relation be- tween nominals in cases where ambiguities exist is useful for generalisation in NLP systems. We created a system based upon a maximum entropy (ME) classifier developed by Clark and Curran (2004). A separate binary classifier for each of the relations was trained over the cor- responding training data, and the additional fea- tures used for each relation were selected by performing a seven-fold cross-validation over all combinations of features developed for this task. We report an overall accuracy of 71.9% macro- averaged over the seven relations and an overall F-measure of 70.7%, comparable with state of the art performance. 2 Background The task of relation classification is complicated by the lack of consensus on relation sets and al- gorithms. Previous research has studied areas as diverse as noun compound classification in the medical domain (Rosario and Hearst, 2001), gene relations (Stephens et al., 2001), verb-verb se- mantic relations (Chklovski and Pantel, 2004), and noun-modifier relations (Nastase and Sz- pakowicz, 2003). Many independent class hier- archies have been developed to suit each appli- cation domain, and it is difficult to transfer one Relation …

Content truncated to 500 words