Abstract
This paper introduces a Synthetic Synchronous Grammar (SSG) that merges the strengths of two popular machine translation approaches: the high-generalization capability of formal hierarchical models (SCFG) and the precise disambiguation of linguistically-motivated models (STSSG). By allowing a single translation process to draw from both types of grammar rules, the authors achieved significant performance gains over traditional single-grammar systems and standard post-decoding combination methods.
Introduction
In Statistical Machine Translation (SMT), researchers often face a trade-off between generative power (how well a model handles new, unseen sentences) and expressive power (how well it chooses the correct translation among many possibilities).
Formally-driven models, such as Synchronous Context-Free Grammars (SCFG), use abstract placeholders that generalize well to new text but often produce ambiguous candidates. Conversely, linguistically-motivated models, like Synchronous Tree Sequence Substitution Grammars (STSSG), use specific syntactic labels (e.g., NP for Noun Phrase) to improve disambiguation but often fail to "match" input text because their rules are too rigid. The authors propose the SSG framework to bridge this gap, creating a single grammar that utilizes both rule types simultaneously.
Methodology
The core innovation of the Synthetic Synchronous Grammar (SSG) is the "syncretization" of grammar types into a unified framework:
- Unified Rule Set: The model extracts both SCFG (abstract) and STSSG (syntactic) rules from the same parallel training data.
- Integrated Derivation: During translation, the decoder builds a single derivation using a mix of rules. For example, it might use a broad SCFG rule for sentence structure while plugging in a precise STSSG rule for a specific phrase.
- HITREE Decoder: The authors implemented a decoder that uses a beam search based on the CKY algorithm, treating translation as a search for the best derivation across the combined rule space.
Experimental Results
The researchers evaluated their model on a Chinese-to-English translation task using the FBIS parallel corpora and the NIST MT2005 evaluation set. The SSG-based system, HITREE, was compared against linguistic (LSTSSG) and formal (FSCFG) baselines, as well as standard system combination methods.
As shown in the results, HITREE outperformed all baselines, achieving a BLEU4 score of 0.2730. This improvement was statistically significant ($p < 0.001$) and notably surpassed Minimum Bayes Risk (MBR) decoding, a powerful post-decoding combination method.
Key Insights
- The Generalization Gap: The authors found that SCFG rules matched the test set 5% of the time, while STSSG rules only matched 3.5%. This confirms that SCFG is better at coverage, while STSSG provides better contextual understanding.
- Pre-decoding Combination: Unlike traditional "system combination" that happens after translation, SSG combines systems at the logic level, allowing for a much larger search space of possible translations.
- Rule Synergy: HITREE utilized significantly more rules (1,927k) than either individual system, suggesting it finds valid translation paths that neither individual grammar could complete on its own.
Conclusion and Future Work
The SSG model successfully leverages the complementary strengths of formal and linguistic translation models. Future research will focus on including "hybrid" rules that mix abstract and syntactic non-terminals and developing sophisticated features to reward grammar consistency and structural richness.
Summary updated