Bilingual Markov Reordering Labels for Hierarchical SMT

Gideon Maillette de Buy Wenniger, Khalil Sima'anSSST @ EMNLP 2014 · DOI

TL;DR

This paper introduces a novel method for improving hierarchical Statistical Machine Translation (SMT) by automatically inducing "bilingual Markov reordering labels" directly from existing Hiero grammars and word alignments. These labels capture phrase reordering patterns, and when used with a soft constraint decoding approach, significantly boost translation quality, especially for Chinese-English.

What Problem Does This Paper Address?

Hierarchical SMT models like Hiero are powerful but can struggle with complex phrase reordering. Existing methods to improve reordering often rely on external linguistic resources (like parsers or taggers), which can be costly and may not perfectly align with the data. This paper aims to address this by developing a way to automatically learn effective reordering signals from the data itself, without needing extra linguistic tools, and integrating them into the Hiero framework.

Methods

The core innovation lies in how they derive and use "bilingual Markov reordering labels."

  • Label Derivation: They start with word alignments and represent them as Normalized Decomposition Trees (NDTs). Each node in these trees is augmented with an "operator" that describes how source phrases are reordered to form target phrases. These operators are then clustered into categories based on their complexity.
  • Phrase-Centric (0th-order) Labels: These labels look inside a phrase pair to describe how its sub-phrases are reordered. They categorize these reorderings into types like Monotonic, Inverted, Permutation, and Complex.
  • Parent-Relative (1st-order) Labels: These labels capture how a phrase pair is reordered relative to its parent phrase in the derivation tree. This adds context about the broader reordering landscape.
  • Soft Constraint Decoding: Instead of strictly enforcing these labels (which can be brittle), they employ a "soft constraint" decoding approach. This allows the decoder to consider the labels as preferences rather than hard rules, making the system more robust to imperfect labels. They use "label substitution features" to integrate these preferences into the decoding process.

Datasets

  • German-English: Europarl corpus (parliament proceedings) for training, WMT-07 development and test data.
  • Chinese-English: MultiUn (United Nations documents) and Hong Kong Parallel Text (Hansards, Laws, News) for training. Multiple-Translation Chinese datasets from LDC for development and testing.
  • Language models were trained on approximately 1M (German-English) and 5.4M (Chinese-English) sentences.

Key Insights and Ideas

  • Bilingual Labels are Powerful: Automatically learned labels that capture bilingual reordering patterns (i.e., how source and target phrases align and reorder) are more effective than purely monolingual syntactic labels.
  • Context Matters (1st-order): Looking beyond just the internal reordering of a phrase pair (0th-order) to its reordering relative to its parent (1st-order) provides significant additional benefit.
  • Soft Constraints are Crucial: Strict enforcement of reordering labels can hurt performance. Soft constraints, allowing the decoder to learn how to best utilize these labels, are essential for achieving consistent improvements.
  • No External Resources Needed: The method successfully extracts useful reordering signals without requiring external tools like parsers or taggers, making it more accessible and potentially more adaptable.

Key Results

  • Chinese-English: The best system (Hiero-1st-Sft, using fine-grained 1st-order labels with soft matching) achieved significant improvements: +0.98 BLEU, +0.42 METEOR, and +1.81 KRS over the baseline Hiero. TER worsened slightly.
  • German-English: While improvements were more modest, the soft-matched systems showed positive gains, particularly in METEOR and TER.
  • The fine-grained labels (0th and 1st order) generally outperformed the coarser ITG+ labels.
  • Soft matching consistently outperformed strict matching across the board.

Limitations and Future Work

  • TER Worsening: For Chinese-English, the best system saw a slight worsening in TER (translation edit rate), indicating a potential trade-off between reordering quality and other translation aspects.
  • Coarse vs. Fine Labels: While fine-grained labels generally performed better, the paper notes that there wasn't always a clear winner between coarse and fine variants, suggesting further tuning might be beneficial.
  • Future Work: The authors suggest exploring the complementarity of their bilingual reordering labels with target-side syntax, potentially leading to even better fluency and translation quality. They also hint at combining their approach with other reordering techniques.

Summary updated

BibTeX
@inproceedings{wenniger2014bilingual,
  author = {Gideon Maillette de Buy Wenniger and Khalil Sima'an},
  title = {{Bilingual Markov Reordering Labels for Hierarchical SMT}},
  booktitle = {SSST @ EMNLP},
  year = {2014},
  doi = {10.3115/v1/w14-4002},
}
Figure