A joint inference of deep case analysis and zero subject generation for Japanese-to-English statistical machine translation

Taku Kudo, Hiroshi Ichikawa, Hideto KazawaACL 2014 · DOI

Abstract

Translating Japanese to English presents significant challenges due to structural differences, specifically the "pro-drop" nature of Japanese and the ambiguity of surface case markers. This paper proposes a joint inference framework that simultaneously performs deep case analysis and zero subject generation. By utilizing Integer Linear Programming (ILP) to enforce linguistic constraints, the system identifies hidden subjects and determines semantic roles more accurately. Experimental results on a large-scale Japanese-to-English translation task demonstrate that this joint approach significantly improves translation quality, outperforming baseline and independent models in both BLEU and RIBES metrics.

Introduction

Japanese-to-English statistical machine translation (SMT) is hindered by fundamental linguistic gaps. English follows a strict Subject-Verb-Object (SVO) order and requires explicit subjects. In contrast, Japanese is a Subject-Object-Verb (SOV) language where subjects are frequently omitted (zero pronouns) when they can be inferred from context. Furthermore, Japanese case markers like ga, wo, and wa are often ambiguous, mapping to different semantic roles depending on the verb's voice.

Example of Japanese sentence structure highlighting particle ambiguity and omitted subjects.

As shown in Figure 1, the particle ga can mark an object in potential verb forms, and the topic marker wa can obscure the underlying subject or object. Previous attempts to address these issues often treated subject generation and reordering as independent tasks. However, these problems are inherently coupled; determining if a subject is missing requires knowing the semantic roles of the existing words.

Methodology

The proposed framework introduces a Japanese deep syntactic parser that operates on dependency trees. It employs a joint probability model consisting of three components:

  1. Zero Subject Model ($P_z$): A Maximum Entropy classifier predicting the presence and type of omitted subjects (e.g., "I", "you", "it").
  2. Voice Type Model ($P_v$): A classifier identifying the verb's voice (active, passive, or potential), which dictates the interpretation of case markers.
  3. Deep Case Model ($P_d$): An estimation of the likelihood that a word serves as a subject or object based on its surface particle and verb frequency data.

Global Inference with ILP

To ensure linguistic consistency, the authors use Integer Linear Programming (ILP) to enforce global constraints during inference. Key constraints include the Mandatory Subject rule (every English sentence must have a subject, either from the text or generated) and the Uniqueness rule (a predicate cannot have multiple subjects or objects). This prevents the model from making linguistically impossible choices, such as generating a zero subject when a subject is already present in the surface text.

The output of this joint analysis is used for pre-ordering, where identified subjects are moved to the front of the sentence to align with English SVO structure before translation.

Evaluation

The system was trained on a massive parallel corpus of 300 million source words and evaluated on a test set of 10,000 web-sampled sentences. The zero subject and voice type classifiers were trained on manually annotated datasets.

Comparison of translation performance across different reordering and subject generation strategies.

As detailed in Figure 2, the joint approach ("joint deep reordering + zero subj.") achieved the highest scores, reaching 19.90 BLEU and 61.89 RIBES. This represents a significant improvement over the baseline and standard surface reordering techniques. The results confirm that addressing deep case analysis and zero subject generation together provides a more robust foundation for Japanese-to-English translation.

Conclusion

This research highlights the interdependence of deep case analysis and zero subject generation in Japanese-to-English translation. By framing these as a joint inference problem constrained by ILP, the authors successfully bridged structural gaps between the two languages. Future work may involve integrating causative voices and handling more complex coordination structures, as well as exploring full joint inference that includes the initial dependency parsing stage.

Summary updated

BibTeX
@inproceedings{kudo2014joint,
  author = {Taku Kudo and Hiroshi Ichikawa and Hideto Kazawa},
  title = {{A joint inference of deep case analysis and zero subject generation for Japanese-to-English statistical machine translation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2014},
  doi = {10.3115/v1/p14-2091},
}
Figure