Classification of Semantic Relationships between Nominals Using Pattern Clusters

Proceedings of ACL-08: HLT, pages 227–235, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Classification of Semantic Relationships between Nominals Using Pattern Clusters Dmitry Davidov ICNC Hebrew University of Jerusalem dmitry@alice.nc.huji.ac.il Ari Rappoport Institute of Computer Science Hebrew University of Jerusalem arir@cs.huji.ac.il Abstract There are many possible different semantic re- lationships between nominals. Classification of such relationships is an important and dif- ficult task (for example, the well known noun compound classification task is a special case of this problem). We propose a novel pat- tern clusters method for nominal relationship (NR) classification. Pattern clusters are dis- covered in a large corpus independently of any particular training set, in an unsupervised manner. Each of the extracted clusters cor- responds to some unspecified semantic rela- tionship. The pattern clusters are then used to construct features for training and classifi- cation of specific inter-nominal relationships. Our NR classification evaluation strictly fol- lows the ACL SemEval-07 Task 4 datasets and protocol, obtaining an f-score of 70.6, as op- posed to 64.8 of the best previous work that did not use the manually provided WordNet sense disambiguation tags. 1 Introduction Automatic extraction and classification of seman- tic relationships is a major field of activity, of both practical and theoretical interest. A prominent type of semantic relationships is that holding between nominals1. For example, in noun compounds many different semantic relationships are encoded by the same simple form (Girju et al., 2005): ‘dog food’ de- notes food consumed by dogs, while ‘summer morn- 1Our use of the term ‘nominal’ follows (Girju et al., 2007), and includes simple nouns, noun compounds and multiword ex- pressions serving as nouns. ing’ denotes a morning that happens in the summer. These two relationships are completely different se- mantically but are similar syntactically, and distin- guishing between them could be essential for NLP applications such as question answering and ma- chine translation. Relation classification usually relies on a train- ing set in the form of tagged data. To improve re- sults, some systems utilize additional manually con- structed semantic resources such as WordNet (WN) (Beamer et al., 2007). However, in many domains and languages such resources are not available. Fur- thermore, usage of such resources frequently re- quires disambiguation and connection of the data to the resource (word sense disambiguation in the case of WordNet). Manual disambiguation is unfeasible in many practical tasks, and an automatic one may introduce errors and greatly degrade performance. It thus makes sense to try to minimize the usage of such resources, and utilize only corpus contexts in which the relevant words appear. A leading method for utilizing context informa- tion for classification and extraction of relationships is that of patterns (Hearst, 1992; Pantel and Pen- nacchiotti, 2006). The standard classification pro- cess is to find in an auxiliary corpus a set of patterns in which a given training word pair co-appears, and use pattern-word pair co-appearance statistics as fea- tures for machine learning algorithms. In this paper we introduce a novel approach, …

Content truncated to 500 words