Abstract
Translating from English to morphologically rich languages like Japanese presents significant challenges due to the lack of explicit case markers in the source language. This paper introduces a syntax-aware statistical model designed to predict these markers—the grammatical "glue" (e.g., ga, wo, ni) that defines noun roles. By integrating this model into a translation pipeline via a novel "expanded $n$-best re-ranking" technique, the authors demonstrate significant improvements in translation fluency and accuracy.
Introduction
In Statistical Machine Translation (SMT), grammatical elements such as case markers are often treated as standard content words. However, case markers are unique; they frequently lack direct counterparts in English, and their correct usage depends on long-range syntactic dependencies, such as the relationship between a distant subject and a passive verb.
The authors identified that in baseline English-to-Japanese systems, approximately 16% of translation errors stem solely from incorrect case markers. For instance, a system might correctly translate the lexical items for "patch" and "replace" but erroneously use an accusative marker (wo) instead of a nominative one (ga) in a passive construction, rendering the sentence ungrammatical. Standard SMT systems struggle with these markers because their context is typically limited to local $n$-grams (3-4 words).
Methodology
The researchers proposed a two-stage approach to address this bottleneck:
1. The Case Prediction Model
A multi-class log-linear (Maximum Entropy) classifier was developed to predict one of 18 Japanese case markers (or "NONE") for every noun phrase (bunsetsu). Unlike standard MT decoders, this model utilizes a rich feature set, including source-side English syntax, target-side Japanese dependency trees, and part-of-speech (POS) tags. This allows the model to consider a broader context, including parent-child relationships within the syntactic tree.
2. Expanded $n$-best Re-ranking
A key architectural contribution of this work is the "expanded $n$-best re-ranking" method. Traditional re-ranking evaluates the top $N$ outputs of an MT system. However, the authors found that the correct case marker variation is often absent from even the top 1000 hypotheses. To overcome this, they take the top $n$ translations and programmatically generate the top $k$ case marker variations for each using their prediction model. This expanded pool of candidates is then re-scored using a combination of original MT features and new features from the case model.
Experimental Setup
The study utilized a technical/computer domain dataset:
- Training: 500,000 English-Japanese sentence pairs.
- Evaluation: Disjoint sets including
lambda-1K(tuning),dev-1K(selection), andtest-2K(final evaluation). - Reference Testing: A
5K-testset evaluated the case marker model in isolation on gold-standard human translations.
Results and Discussion
The log-linear model achieved 94.9% accuracy in predicting case markers on reference text, significantly outperforming a standard trigram language model (87.2%).
Translation Performance
The expanded re-ranking method yielded the best results. The system improved the BLEU score from a baseline of 35.53 to 36.29 on the test set. While a gain of ~0.75 might appear modest, it represents a statistically significant improvement in SMT. Human evaluations further supported these findings; in head-to-head comparisons where translations differed only in case markers, judges preferred the proposed system's output in 42% of cases, with zero instances of degradation compared to the baseline.
Key Insights
- Search Space Bottleneck: Standard re-ranking fails because decoders often do not generate the correct grammatical variations in the initial $n$-best list.
- Syntax as a Bridge: Using English dependency structures to predict Japanese markers captures deep grammatical intent that surface patterns miss.
- Fluency vs. Adequacy: Improvements were primarily seen in fluency (naturalness) rather than adequacy (meaning preservation), suggesting case markers function as "grammatical oil."
Conclusion
This research demonstrates that syntax-aware modeling and expanded re-ranking can effectively address the complexities of case marker generation in English-to-Japanese translation. Future work may extend this "predict-and-expand" framework to other grammatical phenomena, such as article generation or complex verb inflections in other morphologically rich languages.
Summary based on "Generating Case Markers in Machine Translation" by Kristina Toutanova and Hisami Suzuki (2007).
Summary updated