Chinese Part-of-Speech Tagging: One-at-a-Time or All-at-Once? Word-Based or Character-Based?

Chinese Part-of-Speech Tagging: One-at-a-Time or All-at-Once? Word-Based or Character-Based? Hwee Tou Ng and Jin Kiat Low Department of Computer Science National University of Singapore 3 Science Drive 2, Singapore 117543 {nght, lowjinki}@comp.nus.edu.sg Abstract Chinese part-of-speech (POS) tagging assigns one POS tag to each word in a Chinese sentence. However, since words are not demarcated in a Chinese sentence, Chinese POS tagging requires word segmentation as a prerequisite. We could perform Chinese POS tagging strictly after word segmentation (one-at-a-time approach), or perform both word segmentation and POS tagging in combined, single step simultaneously (all-at-a once approach). Also, we could choose to assign POS tags on a word-by-word basis, making use of word features in the surrounding context (word-based), or on a character-by-character basis with character features (character-based). This paper presents an in-depth study on such issues of processing architecture and feature representation for Chinese POS tagging, within a maximum entropy framework. We found that while the all-at-once, character- based approach is the best, the one-at-a-time, character-based approach is a worthwhile compromise, performing only slightly worse in terms of accuracy, but taking shorter time to train and run. As part of our investigation, we also built a state-of-the-art Chinese word segmenter, which outperforms the best SIGHAN 2003 word segmenters in the closed track on 3 out of 4 test corpora. 1 Introduction Most corpus-based language processing research has focused on the English language. Theoretically, we should be able to just port corpus-based, machine learning techniques across different languages since the techniques are largely language independent. However, in practice, the special characteristics of different languages introduce complications. For Chinese in particular, words are not demarcated in a Chinese sentence. As such, we need to perform word segmentation before we can proceed with other tasks such as part-of-speech (POS) tagging and parsing, since one POS tag is assigned to each Chinese word (i.e., all characters in a Chinese word have the same POS tag), and the leaves of a parse tree for a Chinese sentence are words. To build a Chinese POS tagger, the following questions naturally arise: (1) Should we perform Chinese POS tagging strictly after word segmentation in two separate phases (one-at-a-time approach), or perform both word segmentation and POS tagging in a combined, single step simultaneously (all-at-once approach)? (2) Should we assign POS tags on a word-by- word basis (like in English), making use of word features in the surrounding context (word-based), or on a character-by-character basis with character features (character-based)? This paper presents an in-depth study on such issues of processing architecture and feature representation for Chinese POS tagging, within a maximum entropy framework. We analyze the performance of the different approaches in our attempt to find the best approach. To our knowledge, our work is the first to systematically investigate such issues in Chinese POS tagging. 2 Word Segmentation As a first step in our investigation, we built a Chinese word segmenter capable of performing word segmentation without using POS tag information. Since errors in word segmentation will propagate …

Content truncated to 500 words