Convolution Kernels on Constituent, Dependency and Sequential Structures for Relation Extraction

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1378–1387, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Convolution Kernels on Constituent, Dependency and Sequential Structures for Relation Extraction Truc-Vien T. Nguyen and Alessandro Moschitti and Giuseppe Riccardi nguyenthi,moschitti,riccardi@disi.unitn.it Department of Information Engineering and Computer Science University of Trento 38050 Povo (TN), Italy Abstract This paper explores the use of innovative kernels based on syntactic and semantic structures for a target relation extraction task. Syntax is derived from constituent and dependency parse trees whereas se- mantics concerns to entity types and lex- ical sequences. We investigate the effec- tiveness of such representations in the au- tomated relation extraction from texts. We process the above data by means of Sup- port Vector Machines along with the syn- tactic tree, the partial tree and the word sequence kernels. Our study on the ACE 2004 corpus illustrates that the combina- tion of the above kernels achieves high ef- fectiveness and significantly improves the current state-of-the-art. 1 Introduction Relation Extraction (RE) is defined in ACE as the task of finding relevant semantic relations between pairs of entities in texts. Figure 1 shows part of a document from ACE 2004 corpus, a collec- tion of news articles. In the text, the relation be- tween president and NBC’s entertainment division describes the relationship between the first entity (person) and the second (organization) where the person holds a managerial position. Several approaches have been proposed for au- tomatically learning semantic relations from texts. Among others, there has been increased interest in the application of kernel methods (Zelenko et al., 2002; Culotta and Sorensen, 2004; Bunescu and Mooney, 2005a; Bunescu and Mooney, 2005b; Zhang et al., 2005; Wang, 2008). Their main prop- erty is the ability of exploiting a huge amount of This work has been partially funded by the LiveMemo- ries project (http://www.livememories.org/) and Expert Sys- tem (http://www.expertsystem.net/) research grant. Jeff Zucker, the longtime executive producer of NBC’s ”Today” program, will be named Friday as the new president of NBC’s entertainment division, replacing Garth Ancier, NBC execu- tives said. Figure 1: A document from ACE 2004 with all entity mentions in bold. features without an explicit feature representation. This can be done by computing a kernel function between a pair of linguistic objects, where such function is a kind of similarity measure satisfy- ing certain properties. An example is the sequence kernel (Lodhi et al., 2002), where the objects are strings of characters and the kernel function com- putes the number of common subsequences of characters in the two strings. Such substrings are then weighted according to a decaying factor pe- nalizing longer ones. In the same line, Tree Ker- nels count the number of subtree shared by two in- put trees. An example is that of syntactic (or sub- set) tree kernel (SST) (Collins and Duffy, 2001), where trees encode grammatical derivations. Previous work on the use of kernels for RE has exploited some similarity measures over di- verse features (Zelenko et …

Content truncated to 500 words