Discriminative Slot Detection Using Kernel Methods

1 Discriminative Slot Detection Using Kernel Methods Shubin Zhao, Adam Meyers, Ralph Grishman Department of Computer Science New York University 715 Broadway, New York, NY 10003 shubinz, meyers, grishman@cs.nyu.edu Abstract Most traditional information extraction approaches are generative models that assume events exist in text in certain patterns and these patterns can be regenerated in various ways. These assumptions limited the syntactic clues being considered for finding an event and confined these approaches to a particular syntactic level. This paper presents discriminative framework based on kernel SVMs a that takes into account different levels of syntactic information and automatically identifies the appropriate clues. Kernels are used to represent certain levels of syntactic structure and can be combined in principled ways as input for an SVM. We will show that by combining a low level sequence kernel with a high level kernel on a GLARF dependency graph, the new approach outperformed a good rule-based system on slot filler detection for MUC-6. 1 Introduction The goal of Information Extraction (IE) is to extract structured facts of interest from text and present them in databases or templates. Much of the IE research was promoted by the US Government-sponsored MUCs (Message Understanding Conferences). The techniques used by Information Extraction depend greatly on the sublanguage used in a domain, such as financial news or medical records. The training data for an IE system is often sparse since the target domain changes quickly. Traditional IE approaches try to generate patterns for events by various means using training data. For example, the FASTUS (Appelt et al., 1996) and Proteus (Grishman, 1996) systems, which performed well for MUC-6, used hand-written rules for event patterns. The symbolic learning systems, like AutoSlog (Riloff, 1993) and CRYSTAL (Fisher et al., 1996), generated patterns automatically from specific examples (text segments) using generalization and predefined pattern templates. There are also statistical approaches (Miller et al., 1998) (Collins et al., 1998) trying to encode event patterns in statistical CFG grammars. All of these approaches assume events occur in text in certain patterns. However this assumption may not be completely correct and it limits the syntactic information considered by these approaches for finding events, such as information on global features from levels other than deep processing. This paper will show that a simple bag-of-words model can give us reliable information about event occurrence. When training data is limited, these other approaches may also be less effective in their ability to generate reliable patterns. The idea for overcoming these problems is to avoid making any prior assumption about the syntactic structure an event may assume; instead, we should consider all syntactic features in the target text and use a discriminative classifier to decide that automatically. Discriminative classifiers make no attempt to resolve the structure of the target classes. They only care about the decision boundary to separate the classes. In our case, we only need criteria to predict event elements from text using the syntactic features provided. This seems a more suitable solution for IE where training …

Content truncated to 500 words