Chunking Using Conditional Random Fields in Korean Texts

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 155 – 164, 2005. © Springer-Verlag Berlin Heidelberg 2005 Chunking Using Conditional Random Fields in Korean Texts Yong-Hun Lee, Mi-Young Kim, and Jong-Hyeok Lee Div. of Electrical and Computer Engineering POSTECH and AITrc, San 31, Hyoja-dong, Nam-gu, Pohang, 790-784, R. of Korea {yhlee95, colorful, jhlee}@postech.ac.kr Abstract. We present a method of chunking in Korean texts using conditional random fields (CRFs), a recently introduced probabilistic model for labeling and segmenting sequence of data. In agglutinative languages such as Korean and Japanese, a rule-based chunking method is predominantly used for its sim- plicity and efficiency. A hybrid of a rule-based and machine learning method was also proposed to handle exceptional cases of the rules. In this paper, we present how CRFs can be applied to the task of chunking in Korean texts. Ex- periments using the STEP 2000 dataset show that the proposed method signifi- cantly improves the performance as well as outperforms previous systems. 1 Introduction Text chunking is a process to identify non-recursive cores of various phrase types without conducting deep parsing of text [3]. Abney first proposed it as an intermedi- ate step toward full parsing [1]. Since Ramshaw and Marcus approached NP chunking using a machine learning method, many researchers have used various machine learn- ing techniques [2,4,5,6,10,11,13,14]. The chunking task was extended to the CoNLL- 2000 shared task with standard datasets and evaluation metrics, which is now a stan- dard evaluation task for text chunking [3]. Most previous works with relatively high performance in English used machine learning methods for chunking [4,13]. Machine learning methods are mainly divided into the generative approach and conditional approach. The generative approach relies on generative probabilistic models that assign a joint probability p(X,Y) of paired input sequence and label sequence, X and Y respectively. It provides straightforward understanding of underlying distribution. However, this approach is intractable in most domains without strong independence assumptions that each input element is independent from the other elements in input sequence, and is also difficult to use multiple interacting features and long-range dependencies between input elements. The conditional approach views the chunking task as a sequence of classification problems, and defines a conditional probability p(Y|X) over label sequence given input sequence. A number of conditional models recently have been developed for use. They showed better performance than generative models as they can handle many arbitrary and overlapping features of input sequence [12]. A number of methods are applied to chunking in Korean texts. Unlike English, a rule-based chunking method [7,8] is predominantly used in Korean because of its well-developed function words, which contain information such as grammatical 156 Y.-H. Lee, M.-Y. Kim, and J.-H. Lee relation, case, tense, modal, etc. Chunking in Korean texts with only simple heuristic rules obtained through observation on the text shows a good performance similar to other machine learning methods [6]. Park et al. proposed a hybrid of rule-based and machine learning method to handle exceptional cases of the rules, to …

Content truncated to 500 words