Automatic Cross-Lingual Similarization of Dependency Grammars for Tree-based Machine Translation

Wenbin Jiang, Wen Zhang, Jinan Xu, Rangjia CaiEMNLP 2016 · DOI

Introduction

A significant challenge in tree-based Machine Translation (MT) is the structural divergence between the dependency grammars of different languages. Even when expressing identical concepts, linguistic annotation guidelines often vary significantly across languages—for instance, whether a preposition or its object is considered the "head" of a phrase. This lack of structural isomorphism (matching shapes) hinders the ability of MT models to learn effective mappings between source and target trees.

Factorized Cooperative Similarization

To address this friction, this research introduces an automatic algorithm designed to "nudge" the dependency grammars of two languages toward a more isomorphic state. Unlike previous methods that rely on complex, hand-crafted tree transformation rules, this approach operates at the level of individual classification decisions—specifically, the scores assigned to potential edges between words.

The Blending Mechanism

The core of the method is Factorized Cooperative Similarization. For any potential edge between words, the algorithm calculates a "similarized" score by blending the local grammar's score with a score projected from the reference language via word alignments. This process is iterative: both source (e.g., Chinese) and target (e.g., English) grammars are updated simultaneously to converge on a shared structural middle ground.

Balancing Similarity and Accuracy

A critical component is the weight parameter $\lambda$, which controls the trade-off between structural similarity and linguistic fidelity. Without this constraint, grammars might converge on trivial, "flat" structures that match perfectly but lose all syntactic meaning. To ensure the resulting trees remain linguistically valid, the authors utilize a transfer classifier. This pipeline verifies that the similarized trees retain enough information to be mapped back to the original linguistic style.

The training pipeline for the transfer classifier, which ensures that similarized trees maintain their underlying syntactic integrity.

Experimental Results

The method was evaluated using the Penn Chinese Treebank (CTB) and the Wall Street Journal (WSJ) portion of the Penn Treebank, with the FBIS Chinese-English dataset driving the similarization process.

Impact on Structural Isomorphism

The algorithm successfully increased cross-lingual consistency. As shown in the performance analysis across different configurations, structural similarity (represented by the rising blue curves) can be significantly improved as the model iterates. In primary tests, similarity between Chinese and English dependency trees rose from a baseline of 34.2% to over 60%.

Performance metrics across different configurations, illustrating the trade-off between increasing similarity (blue line) and maintaining accuracy as parameters vary.

Machine Translation Performance

When integrated into a tree-to-string MT system, these similarized grammars yielded substantial improvements in BLEU scores on NIST evaluation sets:

  • NIST 04: Increased from 35.44 to 36.78.
  • NIST 05: Increased from 33.08 to 35.12.

These results outperform previous state-of-the-art dependency-based MT models that utilize standard, non-similarized parsers.

Conclusion and Future Work

This work demonstrates that "soft" similarization—increasing the probability of isomorphic structures during ambiguous parsing decisions—is a powerful tool for cross-lingual tasks. While the current study focused on tree-to-string models, the authors suggest that tree-to-tree translation models would benefit even more from such aligned structural representations. Future work may also address "hard" isomorphism, where structures are forced to be identical through the use of super-nodes.

Summary updated

BibTeX
@inproceedings{jiang2016automatic,
  author = {Wenbin Jiang and Wen Zhang and Jinan Xu and Rangjia Cai},
  title = {{Automatic Cross-Lingual Similarization of Dependency Grammars for Tree-based Machine Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2016},
  doi = {10.18653/v1/d16-1048},
}
Figure