Abstract: Hwee Tou Ng, Jin Kiat Low. Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing.
2004.
TL;DR
This paper investigates the optimal architecture for Chinese part-of-speech (POS) tagging, comparing pipeline ("one-at-a-time") versus joint ("all-at-once") models, as well as word-based versus character-based features. The authors find that while an all-at-once character-based approach yields the highest accuracy, a one-at-a-time character-based pipeline offers a highly efficient compromise with nearly identical performance and drastically lower computational costs.
What Problem Does This Paper Address?
In English NLP, words are neatly separated by spaces, making POS tagging a relatively straightforward task. Chinese, however, lacks explicit word boundaries. Before you can assign a POS tag to a Chinese word, you generally have to figure out where the word begins and ends—a task known as word segmentation.
This creates two major architectural questions for building a Chinese POS tagger:
- Processing Architecture: Should the system perform word segmentation first and POS tagging second (a "one-at-a-time" pipeline), or should it do both simultaneously in a single step (an "all-at-once" joint model)?
- Feature Representation: Should the POS tagger look at word-level features in the surrounding context (the standard approach for English), or should it look at character-level features?
The authors systematically evaluate these combinations using a Maximum Entropy framework to determine the most effective and practical approach for Chinese text.
Methods
The authors built and compared several Maximum Entropy (MaxEnt) models:
- A Standalone Word Segmenter: They built a baseline segmenter that assigns boundary tags to characters:
b(begin),m(middle),e(end), ands(single-character word). They improved upon existing models by adding novel features like punctuation checks, character type classes, and a dynamic programming algorithm to ensure valid tag sequences. - One-at-a-Time, Word-Based POS Tagger: A direct port of a classic English MaxEnt POS tagger to Chinese, using pre-segmented words and surrounding word contexts.
- One-at-a-Time, Character-Based POS Tagger: This model assigns POS tags character-by-character, leveraging the surrounding characters as predictive features.
- All-at-Once, Character-Based POS Tagger: A joint model that performs segmentation and tagging simultaneously by assigning combined boundary and POS tags to each character (e.g.,
b_NN).
Datasets
- Penn Chinese Treebank (CTB) version 3.0: A 250K-word corpus used for 10-fold cross-validation.
- 2003 SIGHAN Bakeoff Corpora: Datasets from Academia Sinica (AS), Hong Kong CityU (HK), Peking University (PK), and CTB, used to benchmark the standalone word segmenter.
Key Insights and Ideas
- English NLP techniques don't always port cleanly to Chinese: The word-based MaxEnt tagger that works for English performed poorly on Chinese text.
- Characters carry heavy semantic weight: Unlike English letters, individual Chinese characters have inherent meanings. Character-based models can look at component characters to make educated guesses about Out-Of-Vocabulary (OOV) words, whereas word-based models lack this granular insight.
- The Efficiency vs. Accuracy Tradeoff: Jointly modeling segmentation and tagging is theoretically superior because it prevents cascading errors. However, the computational cost of predicting from a massive combined tag space (e.g., 128 classes per character) makes the pipeline approach much more practical.
Key Results
Word Segmentation
The authors' standalone segmenter achieved a 95.1% F-measure on the CTB dataset. It outperformed the best 2003 SIGHAN bakeoff systems on 3 out of 4 closed-track corpora.
POS Tagging Performance
- Word-Based Pipeline POS: Yielded a poor accuracy of 84.1%.
- Character-Based Pipeline POS: Accuracy jumped to 91.7%. This model was highly efficient, with a testing time of only 50 seconds.
- Character-Based Joint POS: Achieved the highest accuracy at 91.9%, but at a massive computational cost (20 minutes for testing, a 24x slowdown compared to the pipeline).
Limitations and Future Work
The primary limitation is the severe computational penalty of the "all-at-once" approach. While slightly more accurate, its high training and testing times make it less viable for large-scale processing. The authors conclude that the "one-at-a-time" character-based approach is the optimal compromise for Chinese POS tagging, offering near-peak accuracy with vastly superior efficiency.
Summary updated