Abstract: Jun Sun, Min Zhang, Chew Lim Tan. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics.
2010.
Introduction
In syntax-based Statistical Machine Translation (SMT), systems rely on grammar rules to translate and reorder text. Traditionally, these systems build syntactic translation rules upon a foundational layer of word alignment. However, this pipeline approach introduces two significant challenges. First, it is susceptible to error propagation: inaccuracies in word alignment directly degrade the quality of subsequent syntactic rules. Second, word alignment inherently fails to leverage the rich structural information present in parse trees.
To address these limitations, this research proposes a method to directly align syntactic sub-trees—chunks of the parse tree—between source and target languages. By moving beyond rigid heuristic rules and simple lexical similarities, the authors introduce a mathematical framework to capture complex structural correspondences across languages.
Methodology: Bilingual Tree Kernels
The authors frame sub-tree alignment as a binary classification problem (aligned vs. unaligned) solved via a Support Vector Machine (SVM). The core innovation is the introduction of Bilingual Tree Kernels (BTKs), designed to capture structural relationships between source and target trees. Two primary variations were developed:
- Independent BTK (iBTK): This kernel measures the structural similarity of source and target sub-trees independently before combining them. While computationally efficient, it may overlook some direct cross-lingual correspondences.
- Dependent BTK (dBTK): This kernel evaluates the joint similarity of a source-target sub-tree pair as a single unit. It captures richer cross-lingual relationships but requires a significantly higher-dimensional feature space.
To enhance these kernels, the researchers experimented with various sub-structure definitions, such as Subset Trees and Root-directed Subset Trees, to help the model distinguish between nodes sharing similar grammatical tags. Finally, a composite kernel was constructed by combining structural BTKs with a standard polynomial kernel. This polynomial component incorporates "plain" features, including lexical translation probabilities and span differences, to provide a holistic alignment model.
Experimental Setup
The approach was evaluated on two primary datasets:
- HIT Corpus: A gold-standard English-Chinese parallel treebank with 5,000 manually parsed and aligned sentence pairs.
- FBIS (Selected): 300 Chinese-English sentence pairs featuring automatic parsing to test robustness against real-world parsing noise.
For downstream evaluation, the model was integrated into a translation pipeline trained on 240k sentences from the FBIS corpus, using NIST MT-2002 for development and NIST MT-2005 for testing.
Results and Discussion
The composite kernel approach significantly outperformed heuristic baselines in alignment accuracy. On the gold-standard HIT corpus, the F-measure reached 85.32%, a substantial increase from the 65.53% baseline. Even on the noisy, automatically parsed FBIS data, the F-measure improved from 74.36% to 82.70%.
Key insights from the experiments include:
- Lexical-Structural Synergy: Structural features alone are insufficient for disambiguation because common syntactic structures (e.g., Noun Phrases) repeat frequently. BTKs must be paired with lexical features to be effective.
- Kernel Robustness: The complex dBTK performed exceptionally well on the larger HIT corpus but struggled with data sparsity on the smaller FBIS dataset. In contrast, the simpler iBTK proved more robust when training data was limited.
- Hybrid Alignment for MT: In machine translation tasks, the best results were achieved using a hybrid method (EWoS: Either Word or Sub-tree alignment). This approach maintains the statistical power of traditional word alignment while injecting high-quality, linguistically sound syntactic rules, leading to BLEU score improvements for both phrase-based (Moses) and syntax-based (STSSG) systems.
Conclusion
Bilingual Tree Kernels provide a powerful mechanism for direct sub-tree alignment, bypassing the errors inherent in word-alignment-only pipelines. While the performance is currently influenced by the accuracy of automatic parsers and the availability of training data for complex kernels, the results demonstrate that integrating structural syntactic features significantly enhances both alignment quality and downstream translation performance.
Summary updated