Abstract
We describe how simple, commonly understood statistical models, such as statistical dependency parsers, probabilistic context-free grammars, and word-to-word translation models, can be effectively combined into a unified bilingual parser that jointly searches for the best English parse, Korean parse, and word alignment, where these hidden structures all constrain each other. The model used for parsing is completely factored into the two parsers and the TM, allowing separate parameter estimation. We evaluate our bilingual parser on the Penn Korean Treebank and against several baseline systems and show improvements parsing Korean with very limited labeled data.
Introduction
Syntactic parsing is a foundational task in natural language processing, yet its performance is heavily dependent on the availability of large-scale annotated treebanks. For many languages, such as Korean, these resources are scarce, leading to the "resource-poor" challenge. Previous research has attempted to project syntactic information from resource-rich languages like English onto resource-poor ones via word-aligned parallel text, typically through one-way pipelines. This paper proposes a joint inference approach, where the structural constraints of English and Korean mutually inform each other, alongside word alignment, to improve parsing accuracy for the resource-poor language.
Methods
The core of the proposed system is a stochastic bilingual multitext grammar (2-MTG) that utilizes a factored estimation approach. This allows for the independent training of four distinct components: an English log-linear bilexical dependency parser, a Korean log-linear trigram morphological model, an unlexicalized Korean PCFG, and an unsupervised translation model (IBM Models 1–4).
To maintain computational feasibility, the authors employ "half-bilexicalization," lexicalizing only the English side to reduce parsing complexity to $O(n^7)$. A critical innovation is the "English Head Span Filter," a heuristic that restricts Korean constituents to those corresponding to a maximal span of an English head word. This joint search maximizes the combined score of all four models, effectively using English syntax as a structural filter for Korean.
Results and Discussion
The bilingual parser was evaluated on the Penn Korean Treebank (KTB) against a standalone Korean PCFG baseline. The experiments demonstrate that English syntax provides a powerful disambiguation signal, particularly when Korean training data is minimal.
Key findings indicate that on very small datasets (e.g., 32 sentences), the bilingual parser with the English head span filter significantly improves Labeled Precision (from 32.10% to 45.65%) and reduces structural errors, as evidenced by a drop in the "Crossing Brackets" metric from 1.23 to 0.72. However, the aggressive nature of the filter leads to a decrease in Recall (e.g., from 33.78% to 20.82% at 32 sentences), as it discards Korean structures lacking strong English support. The benefits of the bilingual approach are most pronounced in low-resource settings; as more Korean labeled data becomes available, the relative contribution of the English structural advice diminishes.
Conclusion
This work demonstrates that a factored bilingual parser can effectively leverage resource-rich languages to improve parsing in resource-poor ones without requiring bilingual treebanks. While the current model is constrained by computational complexity and alignment sparsity, it establishes a framework for joint inference in translation and parsing. Future improvements may involve more sophisticated pruning techniques or $A^*$ search to handle longer sentences and more complex grammars.
Summary updated