Abstract
Tree-to-string translation rules are widely used in linguistically syntax-based statistical machine translation systems. In this paper, we propose to use deep syntactic information for obtaining fine-grained translation rules. A head-driven phrase structure grammar (HPSG) parser is used to obtain the deep syntactic information, which includes a fine-grained description of the syntactic property and a semantic representation of a sentence. We extract fine-grained rules from aligned HPSG tree/forest-string pairs and use them in our tree-to-string and string-to-tree systems. Extensive experiments on large-scale bidirectional Japanese-English translations testified the effectiveness of our approach.
Introduction
Traditional tree-to-string translation systems often rely on coarse labels such as Part-of-Speech (POS) tags or simple phrasal categories. This coarseness can lead to significant translation errors because the model lacks the context to distinguish between different syntactic or semantic roles. For instance, the English word "killed" might be tagged simply as VBN (past participle). However, its Japanese translation varies significantly depending on whether it is used in an active sense (koroshita) or a passive sense (korosareta).
As shown in the table above, a standard system using only VBN must guess based on a blended probability distribution. By injecting "deep" syntactic information—such as verb voice and semantic arguments—directly into translation rules, we can resolve these ambiguities and improve translation accuracy.
Methodology
The core of our approach is the use of a Head-driven Phrase Structure Grammar (HPSG) parser to generate trees rich with feature structures (TFS). Unlike standard Context-Free Grammar trees, HPSG nodes include detailed attributes like tense, aspect, and voice.
Rule Extraction and PAS
We introduce several technical innovations to leverage this deep syntax:
- Rule Localization: We "flatten" complex HPSG feature structures into a format compatible with standard extraction algorithms while retaining critical features.
- Predicate-Argument Structure (PAS) Rules: We extract rules based on the relationship between a verb and its arguments.
- Minimum Covering Tree (MCT) Algorithm: We developed a linear-time algorithm to identify the smallest tree fragment containing a predicate and its associated arguments.
By using PAS rules, the model captures long-distance dependencies in a single rule, which is more effective than combining multiple independent POS-based rules. This "Minimum Covering Tree" approach allows the decoder to jump directly to relevant semantic structures.
Experimental Results
We evaluated the system on the JST Japanese-English paper abstract corpus, consisting of approximately 1 million parallel training sentences.
Translation Quality
The HPSG forest-based model significantly outperformed the baseline Joshua system. In English-to-Japanese (tree-to-string) tasks, it achieved a BLEU score of 24.75% compared to the baseline's 21.79%. Similar gains were observed in Japanese-to-English (string-to-tree) tasks, with our model reaching 22.67% BLEU against 19.73% for the baseline.
Efficiency
A notable finding was the efficiency of PAS-based rules (PRS). Despite the complexity of deep syntax, the PRS rules allowed the decoder to run 86 times faster than standard composed tree rules. The decoding time per sentence dropped to 0.031 seconds, demonstrating that fine-grained rules can actually simplify the search space for the decoder.
Conclusion
This research demonstrates that fine-grained syntactic and semantic information from HPSG parsers can substantially improve both the accuracy and speed of statistical machine translation. Future work will explore extending this approach to forest-to-tree models and applying deep syntax to both source and target languages simultaneously.
Summary updated