Abstract
Traditional syntax-based machine translation models often rely on strict isomorphism between source and target language trees. However, real-world translations frequently exhibit syntactic divergences. This paper introduces Quasi-Synchronous Grammars (QG), a flexible framework that allows target trees to be "inspired" by source trees without requiring identical structures. By modeling "soft" alignments, QG effectively captures natural linguistic variations, outperforming standard word-alignment baselines in German-English translation tasks.
Introduction
In the mid-2000s, the field of machine translation (MT) sought to incorporate syntactic information to improve translation quality. Most approaches utilized Synchronous Context-Free Grammars (SCFGs), which assume that source and target sentences are generated simultaneously with isomorphic tree structures. This assumption is problematic; empirical studies of human translations show that dependency relationships are preserved only 40–60% of the time. Languages differ in word order (scrambling), head-attachment preferences, and how concepts are lexicalized (e.g., a single verb vs. a light verb construction). Rigid models struggle with these divergences, leading to either poor coverage or excessive computational complexity.
The Quasi-Synchronous Framework
To address these limitations, Smith and Eisner (2006) proposed Quasi-Synchronous Grammars (QG). Unlike synchronous models, QG defines a monolingual grammar for the target language that is conditioned on the source language tree.
The core innovation lies in the use of "senses" as pointers. Each node in the target grammar is associated with a sense that points to a specific node in the source tree, establishing a node-to-node alignment. Crucially, QG employs "soft projection." While it prefers target parent-child pairs to align with source parent-child pairs, it does not mandate it. This allows the model to accommodate various "sloppy" configurations such as head-swapping, sibling alignments, and alignments to more distant relatives or even NULL nodes.
Modeling Syntactic Divergence
By relaxing the isomorphism constraint, QG can learn a taxonomy of structural divergences. The model combines a monolingual dependency model for fluency with bilingual parameters for adequacy. This dual approach ensures that the generated target sentence is both grammatically correct in its own right and a faithful translation of the source.
As shown in Figure 6, the model can handle complex mappings where a single German verb ("bekommen") might align with both an auxiliary ("did") and a main verb ("receive") in English, or where negation is expressed differently across languages.
Evaluation and Results
The authors evaluated QG on the German-English portion of the Europarl corpus. Using dependency trees generated by the Stanford parser, they compared QG against GIZA++ (Model 4), the prevailing standard for word alignment.
The experiments demonstrated that increasing the permissiveness of the grammar—allowing for more types of "breakages" in the alignment—led to a significant drop in cross-entropy on held-out data. This suggests that "sloppy" syntax is a more accurate representation of actual translation data than strict isomorphism. In terms of Alignment Error Rate (AER), the QG model consistently outperformed GIZA++ Model 4. For instance, with a training set of 100,000 sentences, the best QG configuration achieved an AER of 25.99%.
Conclusion
Quasi-Synchronous Grammars provide a robust middle ground between purely formal synchronous grammars and unstructured word-alignment models. By treating the source tree as a flexible "inspiration" for the target tree rather than a rigid template, QG successfully captures the syntactic nuances of human translation. While the computational cost of searching the alignment space remains a challenge, the framework's ability to handle structural divergences marks a significant advancement in syntax-aware machine translation.
Summary updated