Structured Lexical Similarity via Convolution Kernels on Dependency Trees

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1034–1046, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Structured Lexical Similarity via Convolution Kernels on Dependency Trees Danilo Croce DII University of Tor Vergata 00133 Roma, Italy croce@info.uniroma2.it Alessandro Moschitti DISI University of Trento 38123 Povo (TN), Italy moschitti@disi.unitn.it Roberto Basili DII University of Tor Vergata 00133 Roma, Italy basili@info.uniroma2.it Abstract A central topic in natural language process- ing is the design of lexical and syntactic fea- tures suitable for the target application. In this paper, we study convolution dependency tree kernels for automatic engineering of syntactic and semantic patterns exploiting lexical simi- larities. We define efficient and powerful ker- nels for measuring the similarity between de- pendency structures, whose surface forms of the lexical nodes are in part or completely dif- ferent. The experiments with such kernels for question classification show an unprecedented results, e.g. 41% of error reduction of the for- mer state-of-the-art. Additionally, semantic role classification confirms the benefit of se- mantic smoothing for dependency kernels. 1 Introduction A central topic in Natural Language Processing is the design of lexical and syntactic features suitable for the target application. The selection of effective patterns composed of syntactic dependencies and lexical constraints is typically a complex task. Additionally, the availability of training data is usually scarce. This requires the development of generalized features or the definition of seman- tic similarities between them, e.g. as proposed in (Resnik, 1995; Jiang and Conrath, 1997; Schtze, 1998; Pedersen et al., 2004a; Bloehdorn and Mos- chitti, 2007b; Davis et al., 2007) or in semi- supervised settings, e.g. (Chapelle et al., 2006). A semantic similarity can be defined at structural level over a graph, e.g. (Freeman, 1977; Bunke and Shearer, 1998; Brandes, 2001; Zhao et al., 2009), as well as combining structural and lexical similarity over semantic networks, e.g. (Cowie et al., 1992; Wu and Palmer, 1994; Resnik, 1995; Jiang and Conrath, 1997; Schtze, 1998; Leacock and Chodorow, 1998; Pedersen et al., 2004a; Budanitsky and Hirst, 2006). More recent research also focuses on mechanisms to define if two structures, e.g. graphs, are enough similar, as explored in (Mihalcea, 2005; Zhao et al., 2009; F¨urstenau and Lapata, 2009; Navigli and La- pata, 2010). On one hand, previous work shows that there is a substantial lack of automatic methods for engi- neering lexical/syntactic features (or more in gen- eral syntactic/semantic similarity). On the other hand, automatic feature engineering of syntactic or shallow semantic structures has been carried out by means of structural kernels, e.g. (Collins and Duffy, 2002; Kudo and Matsumoto, 2003; Cumby and Roth, 2003; Cancedda et al., 2003; Daum´e III and Marcu, 2004; Toutanova et al., 2004; Shen et al., 2003; Gliozzo et al., 2005; Kudo et al., 2005; Titov and Henderson, 2006; Zelenko et al., 2002; Bunescu and Mooney, 2005; Zhang et al., 2006). The main idea of structural kernels is to generate structures that in turn represent syntactic or shallow semantic features. Most notably, the …

Content truncated to 500 words