EBMT by Tree-Phrasing: a Pilot Study

Philippe Langlais, Fabrizio Gotti, Didier Bourigault, Claude CoulombeMTSummit (Workshop) 2005 · OpenReview

Abstract: Philippe Langlais, Fabrizio Gotti, Didier Bourigault, Claude Coulombe. Workshop on example-based machine translation, 2005.

TL;DR

This pilot study investigates "Tree-Phrases" (TPs)—a novel translation unit that pairs source-language dependency treelets with "elastic" target phrases. By allowing for non-contiguous words and flexible gaps, this approach aims to overcome the rigid sequence limitations of standard phrase-based machine translation, potentially offering better generalization for languages with divergent word orders.

Introduction: Beyond Contiguous Phrases

Standard Phrase-Based Machine Translation (PBMT) relies on contiguous sequences of words. While effective for local reordering, PBMT struggles with "gaps"—discontinuous phrases like the French negation ne... pas—and fails to generalize well when word orders differ significantly between languages.

The authors address the need for a translation unit that is syntactically informed but flexible. They propose moving away from flat word sequences toward dependency-based "treelets" on the source side and "elastic phrases" (EPs) on the target side—sets of words that can be spread across a sentence with arbitrary gaps.

Methodology

The researchers developed a pipeline to create and test a "Tree-Phrase" memory using the French-English Canadian Hansards corpus (1.7 million training pairs).

  1. Parsing: The SYNTEX dependency parser analyzed the source (French) text to identify head-dependent relationships.
  2. Alignment: Word-level alignment was performed using a recursive splitting method and IBM Model 2.
  3. Extraction: The system extracted "Tree-Phrases," consisting of a depth-one source treelet (a governor and its dependents) and its corresponding target words (the elastic phrase).
Examples of depth-one dependency treelets extracted from the French sentence 'on a demandé des crédits fédéraux'.
  1. Matching Policies: To evaluate these units, the authors defined several criteria:
  2. Exact (E): Words must appear with the exact same relative offsets as in the training data.
  3. Relaxed (R): Words must appear in the same order, but gap sizes can vary.
  4. Relaxed with Stretch (R+S): Gaps can grow up to three times their original size.

Key Insights: The Prevalence of Non-Contiguity

A core finding of the study is that standard PBMT models ignore a massive amount of structural data by forcing units to be adjacent strings. As shown in the statistical analysis of the extracted units, less than half of the treelets and only about 40% of the target phrases were contiguous.

Statistical breakdown of treelet and elastic phrase (EP) counts and their respective contiguity percentages (%-c) across different sizes.

Beyond contiguity, the TP approach excels at handling "unknowns." A treelet can match a structure even if it contains an unseen word (like a specific date or name), provided the surrounding syntactic "skeleton" is recognized. Furthermore, the study found that "Oracle" translations (the best possible translations within a candidate list) consistently showed higher Tree-Phrase coverage than standard decoder outputs, suggesting that TP coverage is a strong indicator of translation quality.

Results and Performance

  • Coverage: Using relaxed matching, the Tree-Phrase memory covered roughly 80% of tokens in the test bitext. In a more realistic, non-overlapping decomposition, source coverage remained at ~60-65%, while target coverage was ~56-57%.
  • Validation: In experiments using the Pharaoh decoder, "Oracle" translations achieved a target coverage of 75.4%, compared to 71.6% for the top-ranked output. This 3.9% absolute difference supports the idea that TPs capture essential aspects of "good" translation that standard models might miss.

Limitations and Future Work

As a pilot study, the authors did not build a full translation engine from scratch, instead using TPs to analyze existing system outputs. The computational complexity of finding optimal decompositions remains a challenge, requiring heuristic searches like beam search. Future work aims to explore "treelet-to-treelet" mappings to incorporate internal dependency structures on the target side, moving beyond the current "bag-of-words" elastic phrase model.

Summary updated

BibTeX
@inproceedings{langlais2005ebmt,
  author = {Philippe Langlais and Fabrizio Gotti and Didier Bourigault and Claude Coulombe},
  title = {{EBMT by Tree-Phrasing: a Pilot Study}},
  booktitle = {MTSummit (Workshop)},
  year = {2005},
  url = {https://openreview.net/forum?id=LAGHMbonTYW},
}
Figure