Context-rule Model for Pos Tagging

Context-rule Model for Pos Tagging Yu-Fang Tsai Academia Sinica, Institute of Information Science 128 Sec. 2Academy Rd. Nankang, Taipei, Taiwan 115 eddie@iis.sinica.edu.tw Keh-Jiann Chen Academia Sinica, Institute of Information Science 128 Sec. 2Academy Rd. Nankang, Taipei, Taiwan 115 kchen@iis.sinica.edu.tw Abstract Part-of-speech tagging for a large corpus is a labour intensive and time-consuming task. In order to achieve fast and high quality tagging, algorithms should be high precision and in particular, its tagging results should require less manual proofreading. In this paper, we proposed a context-rule model to achieve both the above goals for pos tagging. We compared the tagging precisions between Markov bi-gram model and context-rule classifier. According to the experiments, context-rule classifier performs better than those two other algorithms. Also, it covers the data sparseness problem by utilizing more context features, and reduces the amount of corpus that is need to be manual proofread by introducing the confidence measure. 1� Introduction Part-of-speech tagging for a large corpus is a labour intensive and time-consuming task. In order to achieve fast and high quality tagging, algorithms should be high precision and in particular, its tagging results should require less manual proofreading. There is lots of work on part-of-speech tagging such as Hidden Markov Models (HMMs), Maximum Entropy Models (MEs), and Support Vector Machines (SVMs), etc. Most of works addressed on the high accuracy of tagging results only. In this paper, we proposed a context-rule model to achieve both the above goals for pos tagging. 2 Tagging Algorithms In this study, we are going to test two different tagging algorithms based on same training data and testing data. The two tagging algorithms are Markov bi-gram model, and context-rule classifier. For Markov bi-gram model, we propose a new form named word-dependent Markov bi-gram model, which will be described later. The training data and testing data are extracted from Sinica corpus, a 5 million word balanced Chinese corpus with pos tagging (Chen et al., 1996). The confidence measure will be defined for each algorithm and the best accuracy will be estimated at the constraint of only a fixed amount of testing data being proofread. It is easier to proofread and make more consistent tagging results, if proofreading processes were done by checking the keyword-in-context file for each ambivalence word and only the tagging results of ambivalence word need to be proofread. The words with single category need not be rechecked their pos tagging. For instance, in Table 1, the keyword-in-context file of the word '1111:F9: (research), which has pos-categories of verb type VE and noun type Nv, is sorted according to its left/right context. The proofreader can see the other examples as references to determine whether or not each tagging result is right. If all of the occurrences of ambivalence word have to be rechecked, it is still too much of the work. The common terms used in the following tagging algorithms were defined as follows: W k� The k-th word in a sequence Ck� The pos-category associated with k-th word wk 146 ilig(DE)� …

Content truncated to 500 words