Introduction
Vietnamese Natural Language Processing (NLP) faces a unique challenge compared to languages like English: word segmentation. In Vietnamese, spaces are used to separate syllables rather than words. A single word can consist of multiple syllables (e.g., cá nhân meaning "individual"). Consequently, word segmentation is a critical prerequisite for downstream tasks like Part-of-Speech (POS) tagging. This paper presents a rigorous empirical evaluation of two primary strategies for addressing this: a two-step pipeline approach versus a joint segmentation-tagging model.
Methodology
The study investigates two architectural paradigms:
- The Pipeline Strategy: Syllables are first grouped into words using a dedicated segmenter (RDRsegmenter), and the resulting tokens are then passed to a POS tagger.
- The Joint Strategy: A single model predicts both word boundaries and POS tags simultaneously. This is achieved by transforming the data into a syllable-based format using a B-I-Tag scheme (e.g.,
B-Vfor the beginning of a Verb orI-Vfor the inside of a Verb).
Evaluated Models
The researchers compared five distinct models:
- Feature-based models: RDRPOSTagger (a transformation rule-based learner) and MarMoT (a high-performance CRF-based tagger).
- Neural models: Three variants of BiLSTM-CRF, including architectures incorporating CNN-based or LSTM-based character-level feature extraction.
Experimental Setup
Experiments were conducted using the Vietnamese Treebank, the largest available annotated corpus for the language, comprising approximately 28,000 manually annotated sentences for training and development, and 2,120 sentences for testing. Neural models utilized 100-dimensional Word2Vec embeddings pre-trained on a large Vietnamese news corpus.
Results and Discussion
Performance under Gold Segmentation
When evaluated on perfectly segmented text ("Gold" segmentation), traditional feature-based models demonstrated superior accuracy. MarMoT achieved the highest accuracy at 95.88%, outperforming the neural BiLSTM-CRF variants.
Pipeline vs. Joint Strategies
A key finding of this research is that the pipeline strategy consistently outperforms the joint strategy for Vietnamese POS tagging. As shown in the comparative results, the MarMoT pipeline achieved a PTag F1 score of 93.96%, significantly higher than the best joint neural model (92.95%).
Efficiency and Speed
In terms of computational efficiency, RDRPOSTagger is the clear leader, processing 180,000 words per second. This is several orders of magnitude faster than neural models, which ranged between 1.5k and 3k words per second.
Conclusion
The empirical evidence suggests that for Vietnamese POS tagging, the traditional pipeline approach using feature-based models remains the state-of-the-art. The underperformance of neural models is likely attributable to the relatively small size of the Vietnamese Treebank compared to datasets available for English or Chinese. Furthermore, feature-based models appear better equipped to capture the word-level context necessary for accurate tagging in the Vietnamese linguistic structure. Future work should focus on integrating Vietnamese-specific linguistic features into feature-based sequence taggers to further advance performance.
Summary updated