Hierarchical Phrase-Based Translation with Jane 2

Matthias Huck, Jan-Thorsten Peter, Markus Freitag, Stephan Peitz, Hermann NeyPrague Bull. Math. Linguistics 2012 · OpenReview

Introduction

Jane 2 represents the second major iteration of the open-source statistical machine translation (SMT) toolkit developed by RWTH Aachen University. As a powerful framework for Hierarchical Phrase-Based Translation (HPBT), Jane 2 utilizes synchronous context-free grammars to bridge the gap between disparate languages. This update introduces a suite of advanced features—including insertion/deletion models, sophisticated lexical scoring, and "soft" syntactic constraints—that collectively yield a +2.6% BLEU improvement over standard hierarchical baselines.

Addressing the Limitations of HPBT

While HPBT is a robust framework, standard implementations frequently encounter two primary obstacles:

  1. Content Omissions: To maximize fluency or brevity, models may inadvertently omit critical content words during translation.
  2. Structural Rigidity: Standard hierarchical rules often lack the necessary flexibility to manage complex syntactic differences between distant language pairs, such as Chinese and English, or the nuance required to ensure output adheres to proper grammatical dependency structures.

Jane 2 addresses these issues through a modular C++ architecture that integrates several advanced SMT techniques into a single, efficient package.

Methodological Enhancements

Insertion and Deletion Models

To prevent the system from "hallucinating" or dropping content, Jane 2 implements insertion and deletion models that serve as a lexical sanity check. By utilizing lexical translation probabilities (derived from IBM Model 1 or relative frequencies), the system tracks words that appear without clear source or target counterparts. These counts are integrated as features to penalize hypotheses that fail to preserve the semantic content of the source text.

Reordering Extensions

Jane 2 expands the reordering capabilities of the decoder through two distinct methods:

  • Non-lexicalized Rules: The grammar is augmented with explicit "swap" and "jump" rules, providing the decoder with greater flexibility to move blocks of text.
  • Discriminative Lexicalized Reordering: A Maximum Entropy model predicts whether a neighboring block should be placed to the left or right based on the surrounding context.
Reordering models in Jane 2 allow for flexible block movement through swap and jump rules.

Soft String-to-Dependency Constraints

A significant innovation in Jane 2 is the transition from "hard" to "soft" syntactic constraints. Traditional models often fail if they cannot produce a perfect syntactic dependency tree. Jane 2, however, builds dependency trees on the fly and applies penalties when dependencies are malformed or contradictory. This allows the model to prioritize syntactic correctness without being paralyzed by noise in the training data.

The soft dependency model evaluates the validity of syntactic structures, penalizing malformed dependencies while maintaining robustness.

Lexical Scoring Variants

The toolkit explores four smoothing methods for phrase table probabilities, including the "Noisy-Or" model and standard Moses-style lexical weighting. These variants improve the system's ability to estimate the quality of rare or partially lexicalized phrases, leading to more reliable translations.

Experimental Evaluation

The researchers evaluated Jane 2 on the NIST Chinese→English translation task. The training set consisted of a parallel corpus of 3.0 million sentence pairs (~77.5M Chinese words and 81.0M English words). The NIST MT06 set was used for development, while the NIST MT08 set served as the unseen evaluation data. Target-side dependency structures for the soft syntax model were generated using the Stanford Dependency Parser.

Key Results

The cumulative impact of these enhancements is substantial. Starting from a standard hierarchical baseline, the addition of specific features resulted in incremental gains:

Model Configuration BLEU (%) TER (%)
Baseline 25.2 66.6
With Insertion/Deletion ~26.0 -
With Lexicalized Reordering 25.8 -
Full Combined Model 27.8 64.4

The full model achieved an absolute gain of +2.6% BLEU and a 2.2% reduction in TER (Translation Edit Rate), marking a significant advancement in translation quality.

Conclusion and Future Work

Jane 2 demonstrates that simple penalties and soft constraints can significantly enhance the robustness of hierarchical translation models. While the "soft" dependency model relies on merging heuristics that can be complex, it effectively balances syntactic accuracy with model flexibility. Future work may focus on refining these heuristics and reducing the computational overhead introduced by discriminative reordering and dependency language model scoring. Currently, the toolkit provides a highly modular and efficient environment for researchers to optimize high-dimensional feature weights using algorithms like MERT, MIRA, and Downhill Simplex.

Summary updated

BibTeX
@inproceedings{huck2012hierarchical,
  author = {Matthias Huck and Jan-Thorsten Peter and Markus Freitag and Stephan Peitz and Hermann Ney},
  title = {{Hierarchical Phrase-Based Translation with Jane 2}},
  booktitle = {Prague Bull. Math. Linguistics},
  year = {2012},
  url = {https://openreview.net/forum?id=4zvQHB4qngm},
}
Figure