Text Chunking by Combining Hand-Crafted Rules and Memory-Based Learning

Text Chunking by Combining Hand-Crafted Rules and Memory-Based Learning Seong-Bae Park Byoung-Tak Zhang School of Computer Science and Engineering Seoul National University Seoul 151-744, Korea {sbpark,btzhang}@bi.snu.ac.kr Abstract This paper proposes a hybrid of hand- crafted rules and a machine learning method for chunking Korean. In the par- tially free word-order languages such as Korean and Japanese, a small number of rules dominate the performance due to their well-developed postpositions and endings. Thus, the proposed method is primarily based on the rules, and then the residual errors are corrected by adopting a memory-based machine learning method. Since the memory-based learning is an efficient method to handle exceptions in natural language processing, it is good at checking whether the estimates are excep- tional cases of the rules and revising them. An evaluation of the method yields the im- provement in F-score over the rules or var- ious machine learning methods alone. 1 Introduction Text chunking has been one of the most interest- ing problems in natural language learning commu- nity since the first work of (Ramshaw and Marcus, 1995) using a machine learning method. The main purpose of the machine learning methods applied to this task is to capture the hypothesis that best deter- mine the chunk type of a word, and such methods have shown relatively high performance in English (Kudo and Matsumoto, 2000; Zhang et. al, 2001). In order to do it, various kinds of information, such as lexical information, part-of-speech and grammat- ical relation, of the neighboring words is used. Since the position of a word plays an important role as a syntactic constraint in English, the methods are suc- cessful even with local information. However, these methods are not appropriate for chunking Korean and Japanese, because such lan- guages have a characteristic of partially free word- order. That is, there is a very weak positional con- straint in these languages. Instead of positional con- straints, they have overt postpositions that restrict the syntactic relation and composition of phrases. Thus, unless we concentrate on the postpositions, we must enlarge the neighboring window to get a good hypothesis. However, enlarging the win- dow size will cause the curse of dimensionality (Cherkassky and Mulier, 1998), which results in the deficiency in the generalization performance. Especially in Korean, the postpositions and the endings provide important information for noun phrase and verb phrase chunking respectively. With only a few simple rules using such information, the performance of chunking Korean is as good as the rivaling other inference models such as ma- chine learning algorithms and statistics-based meth- ods (Shin, 1999). Though the rules are approxi- mately correct for most cases drawn from the do- main on which the rules are based, the knowledge in the rules is not necessarily well-represented for any given set of cases. Since chunking is usually processed in the earlier step of natural language pro- cessing, the errors made in this step have a fatal in- fluence on the following steps. Therefore, the ex- ceptions that are ignored by …

Content truncated to 500 words