Fast Methods for Kernel-based Text Analysis - ChaSen.org.pdfVIP

Fast Methods for Kernel-based Text Analysis Taku Kudo and Yuji Matsumoto Graduate School of Information Science, Nara Institute of Science and Technology taku-ku,matsu@is.aist-nara.ac.jp Abstract 2002) Text Chunking (Kudo and Matsumoto, 2001), Named Entity Recognition (Isozaki and Kazawa, Kernel-based learning (e.g., Support Vec- 2002), and Japanese Dependency Parsing (Kudo and tor Machines) has been successfully ap- Matsumoto, 2000; Kudo and Matsumoto, 2002). plied to many hard problems in Natural It is known in NLP that combination of features Language Processing (NLP). In NLP, al- contributes to a significant improvement in accuracy. though feature combinations are crucial to For instance, in the task of dependency parsing, it improving performance, they are heuris- would be hard to confirm a correct dependency re- tically selected. Kernel methods change lation with only a single set of features from either this situation. The merit of the kernel a head or its modifier. Rather, dependency relations methods is that effective feature combina- should be determined by at least information from tion is implicitly expanded without loss both of two phrases. In previous research, feature of generality and increasing the compu- combination has been selected manually, and the tational costs. Kernel-based text analysis performance significantly depended on these selec- shows an excellent performance in terms