TL;DR
This paper introduces a novel preprocessing technique for statistical machine translation (SMT) that leverages dependency parsing to reorder English (SVO) sentences into a more SOV-like structure before translation. This approach significantly improves translation quality for SOV languages by simplifying the reordering task for the SMT decoder, leading to better BLEU scores without increasing decoding complexity.
What Problem Does This Paper Address?
Statistical Machine Translation (SMT), particularly phrase-based SMT, has achieved state-of-the-art results. However, a significant challenge arises when translating between languages with vastly different word orders, such as Subject-Verb-Object (SVO) languages like English and Subject-Object-Verb (SOV) languages like Korean, Japanese, Hindi, Urdu, and Turkish. Standard phrase-based SMT systems struggle with long-distance reordering, leading to degraded translation quality. Existing reordering methods either explicitly model reordering distances (which can be computationally expensive or limited) or rely on complex syntactic parsing integrated into decoding, increasing complexity. This paper aims to address the long-distance reordering problem for SVO-to-SOV translation by performing reordering before the SMT decoding process, making it more efficient and effective.
Methods
The core innovation is a preprocessing reordering approach based on dependency parsing. The authors analyze English sentences using a dependency parser to understand the grammatical relationships between words (e.g., subject, verb, object). Based on this syntactic structure, they apply precedence reordering rules to transform the English sentence into a more SOV-like order.
These rules are defined for different parts of speech (verbs, adjectives, nouns, prepositions) and specify how their children nodes (arguments, modifiers) should be ordered relative to the head. For instance, a verb rule might dictate that the object should precede the verb in the reordered sentence. The rules use weights and order types (NORMAL/REVERSE) to manage complex reordering scenarios.
Crucially, this reordering is performed before the SMT system sees the data. This means the SMT system, whether phrase-based or hierarchical, operates on a "pre-reordered" input that is closer to the target language's word order. This simplifies the decoding process, as the need for extensive reordering during decoding is reduced. The authors use a deterministic inductive dependency parser for efficiency.
Datasets
The paper uses in-house parallel corpora for training SMT systems for English to Korean, Japanese, Hindi, Urdu, and Turkish. The size of these corpora varies significantly, from around 17 million words to over 350 million words. For evaluation, they collected approximately 10,000 English sentences, translated by humans into the target SOV languages, with each sentence having a single reference translation. This evaluation data is split into development (dev) and test sets for tuning and evaluation, and a blindtest set for final assessment.
Key Insights and Ideas
- Preprocessing Reordering as a Solution: The most significant insight is that reordering before SMT decoding can be more effective and efficient than integrating complex reordering models directly into the decoder.
- Dependency Parsing for Linguistic Knowledge: Leveraging dependency parse trees provides a structured way to inject linguistic knowledge about grammatical relationships, which is crucial for effective reordering.
- Simplifying the Decoder's Task: By transforming SVO to a more SOV-like structure, the SMT decoder faces a less challenging reordering problem, leading to improved performance.
- Flexibility of Precedence Rules: The proposed precedence reordering rules are designed to be flexible, handling various linguistic phenomena (subject, object, prepositional phrases, auxiliary verbs, etc.) with a unified rule format.
- Robustness to Parser Errors: While acknowledging parser errors as a potential limitation, the authors' approach is designed to be somewhat robust, and they suggest future work on handling parser uncertainty.
Key Results
The paper demonstrates significant improvements in BLEU scores across all five SOV languages (Korean, Japanese, Hindi, Urdu, Turkish) when using their dependency parser-based preprocessing reordering approach.
- Phrase-Based SMT: The proposed approach (PR) consistently outperforms the baseline system (BL) and a system with a maximum entropy lexicalized reordering model (-LR). For example, for Korean, the PR system achieved a blindtest BLEU score of 27.5, compared to 26.2 for the baseline. Similar gains were observed for other languages, with some showing statistically significant improvements (marked with ).
- Hierarchical Phrase-Based SMT: The approach also benefits hierarchical phrase-based systems, achieving statistically significant improvements for four out of the five languages when compared to a standard hierarchical system.
- Reduced Sensitivity to Reordering Distance: The reordered English sentences become less sensitive to the maximum allowed reordering distance in the decoder, meaning that even with a smaller, more efficient reordering distance limit, translation quality is maintained or improved. This is illustrated in Figure 5, showing consistent performance with the PR system across different reordering distances.
Limitations and Future Work
- Sensitivity to Parser Errors: The authors acknowledge that their approach is sensitive to errors made by the dependency parser. Some sentences suffer from parser errors, which can impact reordering quality.
- Future Work:
- Automatic Rule Extraction: While the current rules are manually created, the authors suggest that it's possible to learn these rules automatically from word-aligned corpora, similar to other approaches.
- Handling Parser Uncertainty: Investigating methods to handle parser errors more gracefully, perhaps by using word lattices instead of a single parse, is a promising direction.
- Training with Lattices: Exploring efficient procedures for training EM models and obtaining word alignments using word lattices on the source side of parallel data.
- Tree-to-String Models: Suggesting that tree-to-string models could be an interesting direction for future research.
- Extending to Other Languages: Extending the rules and approach to languages other than SOV languages.
Summary updated