Extracting Relations with Integrated Information Using Kernel Methods

Proceedings of the 43rd Annual Meeting of the ACL, pages 419–426, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Extracting Relations with Integrated Information Using Kernel Methods Shubin Zhao Ralph Grishman Department of Computer Science New York University 715 Broadway, 7th Floor, New York, NY 10003 shubinz@cs.nyu.edu grishman@cs.nyu.edu Abstract Entity relation detection is a form of in- formation extraction that finds predefined relations between pairs of entities in text. This paper describes a relation detection approach that combines clues from differ- ent levels of syntactic processing using kernel methods. Information from three different levels of processing is consid- ered: tokenization, sentence parsing and deep dependency analysis. Each source of information is represented by kernel func- tions. Then composite kernels are devel- oped to integrate and extend individual kernels so that processing errors occurring at one level can be overcome by informa- tion from other levels. We present an evaluation of these methods on the 2004 ACE relation detection task, using Sup- port Vector Machines, and show that each level of syntactic processing contributes useful information for this task. When evaluated on the official test data, our ap- proach produced very competitive ACE value scores. We also compare the SVM with KNN on different kernels. 1 Introduction Information extraction subsumes a broad range of tasks, including the extraction of entities, relations and events from various text sources, such as newswire documents and broadcast transcripts. One such task, relation detection, finds instances of predefined relations between pairs of entities, such as a Located-In relation between the entities Centre College and Danville, KY in the phrase Centre College in Danville, KY. The ‘entities’ are the individuals of selected semantic types (such as people, organizations, countries, …) which are re- ferred to in the text. Prior approaches to this task (Miller et al., 2000; Zelenko et al., 2003) have relied on partial or full syntactic analysis. Syntactic analysis can find rela- tions not readily identified based on sequences of tokens alone. Even ‘deeper’ representations, such as logical syntactic relations or predicate-argument structure, can in principle capture additional gener- alizations and thus lead to the identification of ad- ditional instances of relations. However, a general problem in Natural Language Processing is that as the processing gets deeper, it becomes less accu- rate. For instance, the current accuracy of tokeniza- tion, chunking and sentence parsing for English is about 99%, 92%, and 90% respectively. Algo- rithms based solely on deeper representations in- evitably suffer from the errors in computing these representations. On the other hand, low level proc- essing such as tokenization will be more accurate, and may also contain useful information missed by deep processing of text. Systems based on a single level of representation are forced to choose be- tween shallower representations, which will have fewer errors, and deeper representations, which may be more general. Based on these observations, Zhao et al. (2004) proposed a discriminative model to combine in- formation from different syntactic sources using a kernel SVM (Support Vector Machine). We …

Content truncated to 500 words