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

Summary

Chinese text has no explicit word boundaries, so POS tagging depends on segmentation. Ng and Low ask two design questions. First, should segmentation and tagging be done in sequence ("one-at-a-time") or jointly ("all-at-once")? Second, should the tagger rely on word-level or character-level features? They compare the options within a single maximum entropy framework, using 10-fold cross-validation on the 250K-word Penn Chinese Treebank (CTB) 3.0. To make the pipeline comparison fair, they first build a strong character-level segmenter. It labels each character with a boundary tag and decodes with dynamic programming. This segmenter beats the best SIGHAN 2003 closed-track systems on three of four corpora. Porting Ratnaparkhi's English word-based tagger directly to Chinese performs poorly. Character-based tagging works much better. The joint character-based model is the most accurate, but its advantage over the pipelined character-based variant is small, and it costs far more time to train and run. The practical lesson is that character information matters most for Chinese, and joint modelling buys only a marginal gain at a high computational price.

Key points

• Segmenter design:

  • Each character gets a b/m/e/s boundary tag from a maximum entropy model (Java opennlp maxent, feature cutoff 2, 100 iterations).
  • On CTB 10-fold CV it averages 95.1% F-measure, versus Luo (2003) at 94.6% (and roughly 94.9% with a 58K-entry word list).

• SIGHAN 2003 bakeoff:

  • The segmenter beats the best closed-track score on AS, HK and PK.
  • It ranks third on CTB, whose test set has an 18.1% OOV rate.
  • With extra AS training data (open track), it reaches 92.2% on CTB, above the best open-track result.

• Ablation: The new features add about 0.4% F-measure on average, and the consistency post-processing adds another 0.3%.

• POS accuracy on CTB:

  Approach	POS accuracy	Segmentation F-measure	Testing time (10% of CTB)
  One-at-a-time, word-based	84.1%	95.1%	about 1 min 20 s
  One-at-a-time, character-based	91.7%	95.1%	about 1 min 50 s
  All-at-once, character-based	91.9%	95.2%	about 20 min

• Cost and significance:

  • The joint model needs about 3 hours of training, compared with 12 minutes for the segmenter.
  • In a paired t-test (p < 0.01), the joint model's POS gain is significant, but its segmentation gain is not.
  • The all-at-once character-based setup is essentially Luo's (2003) approach.

Interesting ideas

• I1. Constrained decoding over word candidates: Instead of picking the best label for each character, the decoder enumerates candidate last words up to 20 characters long. This guarantees only valid b/m/e/s sequences, and it lets word-level features such as W0C0 be used at test time. The trick could carry over to any BIO-style labelling task where invalid label sequences are a problem.
• I2. Coarse character-type n-gram feature: Characters are mapped to four classes (digits, date characters, Latin letters, other), and the class pattern over a five-character window is used as a feature. This lets the model generalise to unseen numbers and dates. The same idea suits other languages or noisy text with many rare tokens.
• I3. Document-level consistency post-processing: If a sequence of 2–6 adjacent output words, joined together, matches a 3+ character word segmented elsewhere in the same document, the sequence is merged into that word. This cheap step improves consistency for names and terms.
• I4. Character-level evidence for unknown words: Chinese characters carry meaning, and unseen words are far more frequent than unseen characters. Tagging from character features therefore helps on OOV words. The same logic could apply to subword or morpheme features in morphologically rich languages such as Ukrainian.
• I5. Weighing accuracy against cost: The comparison reports training and testing time and significance tests alongside accuracy. This framing showed that a cheaper pipeline can be a sensible choice when the joint model's gain is small.