TL;DR
This paper investigates whether rule-based syntactic preprocessing, proven effective for Phrase-Based Machine Translation (PBMT), can benefit Syntax-Based Machine Translation (SMT). They adapt a successful English-Japanese preprocessing method, "Head Finalization," and find that while gains are smaller than in PBMT, significant accuracy improvements are still achievable, particularly through lexical processing.
What Problem Does This Paper Address?
Statistical Machine Translation (SMT) systems, especially those based on phrases (PBMT), often struggle with languages that have significantly different word orders. While PBMT relies on reordering models, these can have difficulty with long-distance reorderings. To address this, researchers have developed preprocessing techniques that rearrange source sentences to better match the target language's word order. "Head Finalization" is a particularly successful rule-based method for English-to-Japanese PBMT, which significantly improves translation quality by applying simple syntactic transformations.
However, the application of such rule-based syntactic preprocessing to syntax-based SMT systems, which inherently incorporate syntactic information, has been less explored. These systems, like Tree-to-String (T2S) models, already use syntactic structures. The question is whether pre-existing, manually crafted syntactic rules can further enhance these systems, or if their inherent syntactic awareness makes such preprocessing redundant. This paper aims to bridge this gap by examining the effectiveness of Head Finalization's components on a T2S system.
Methods
The authors adapt the "Head Finalization" preprocessing method, originally designed for English-Japanese PBMT, for use with a Tree-to-String (T2S) syntax-based SMT system. Head Finalization consists of two main parts:
- Reordering: This process transforms English syntax trees to better match the head-final structure of Japanese. It involves moving head words to the end of their constituent nodes.
- Lexical Processing: This involves three steps that don't alter word order but aim to make the sentence structure more Japanese-like:
- Pseudo-particle insertion: Adding particles (like "wa," "ga," "wo") to mark grammatical roles, as Japanese uses particles where English relies on word order.
- Determiner elimination: Removing English determiners ("a," "an," "the").
- Singularization: Converting plural nouns to their singular form.
The paper explores two ways of incorporating these methods into T2S:
- Direct Preprocessing: Applying the Head Finalization transformations to the input trees before they are used by the T2S system.
- Feature Addition: Incorporating information from the Head Finalization reordering rules as a "soft constraint" (an HF-feature) within the T2S decoder. This allows the system to learn the utility of head-final ordering without enforcing it strictly.
Datasets
The experiments were conducted using the NTCIR7 PATENT-MT's Patent corpus for English-to-Japanese translation. This corpus was divided into training, development, and test sets. Specific details on the size of these sets are provided in Table 2, showing millions of words for training and thousands for development and testing, with varying average sentence lengths.
Key Insights and Ideas
- Lexical Processing is Key for Syntax-Based SMT: Unlike reordering, the lexical processing components of Head Finalization (especially pseudo-particle insertion) significantly improve T2S translation quality. This suggests that even systems with inherent syntactic awareness can benefit from linguistic adjustments that bridge structural differences.
- Reordering's Limited Impact on T2S: The reordering component of Head Finalization showed little to no benefit, and sometimes even slightly degraded performance, for T2S. This implies that T2S systems, due to their reliance on parse trees and potentially better alignment handling, are already adept at managing word order differences.
- HF-Feature Addition Didn't Work as Expected: Attempting to incorporate reordering information as a soft constraint (HF-feature) did not improve translation quality. The authors hypothesize this might be due to confusion during optimization or the inherent strength of T2S's reordering capabilities.
- PBMT vs. T2S: Different Strengths: The results highlight the distinct advantages of each translation paradigm. Reordering is crucial for PBMT, while T2S benefits more from lexical adjustments that align grammatical structures.
Key Results
- PBMT Improvements: For PBMT, reordering provided substantial gains (e.g., BLEU from 33.16 to 37.77), and lexical processing offered a slight, though not always statistically significant, improvement.
- T2S Baseline: The baseline T2S system (without preprocessing) already outperformed all PBMT conditions.
- T2S with Lexical Processing: Applying lexical processing to T2S resulted in the highest translation quality, achieving a BLEU score of 39.38 and RIBES of 79.21 (line 8 in Table 3). This represents a significant improvement over the T2S baseline (38.94 BLEU, 78.48 RIBES).
- Reordering and HF-Feature on T2S: Reordering alone or combined with lexical processing did not improve T2S scores. The HF-feature addition also showed no positive effect.
Limitations and Future Work
- HF-Feature Optimization: The authors acknowledge that the HF-feature addition did not yield expected results, possibly due to issues with the optimization method (MERT) or the way the feature was integrated. They suggest exploring other optimization techniques or more sophisticated ways of incorporating features.
- Generalization: The study focused on English-Japanese translation. Future work should investigate whether these findings generalize to other language pairs and explore other rule-based preprocessing techniques.
- Tree Transformations: The authors propose exploring other tree transformations beyond simple reordering, such as restructuring the tree by moving children, which might help in cases of highly divergent syntactic structures between languages.
- Soft Constraints: Investigating alternative methods for incorporating preprocessing rules as soft constraints, such as using n-best lists or forests, is also suggested.
Summary updated