Word Alignment with Cohesion Constraint

Dekang Lin, Colin CherryNAACL 2003 · OpenReview

Abstract

We present a syntax-based constraint for word alignment, known as the cohesion constraint. It requires disjoint English phrases to be mapped to non-overlapping intervals in the French sentence. We evaluate the utility of this constraint in two different algorithms. The results show that it can provide a significant improvement in alignment quality.

Introduction

Early Statistical Machine Translation (SMT) models, such as the IBM models, relied primarily on statistical co-occurrences, often ignoring the underlying linguistic structure of sentences. While subsequent research attempted to incorporate syntactic knowledge, these models frequently assumed isomorphism—the idea that grammatical structures between source and target languages are nearly identical. However, structural divergences occur in approximately 31% of translated sentences due to non-literal translations. This paper introduces the cohesion constraint to leverage syntactic knowledge without requiring strict structural equivalence.

Methods

The cohesion constraint utilizes a dependency tree of the source (English) sentence to restrict word alignments. It mandates that if two phrases are disjoint in the source language, their translations must occupy non-overlapping intervals in the target (French) language.

Specifically, the constraint identifies two types of illegal overlaps:

  1. Head-modifier overlap: The translation of a head word falls within the interval of its modifier's translation.
  2. Modifier-modifier overlap: The translations of two modifiers sharing the same head overlap.
Example of word alignment between English and French sentences guided by dependency structure.

The authors implemented a fast checking algorithm to evaluate cohesion violations during the alignment process. This modular approach allows the constraint to be integrated into various alignment algorithms, such as the $\phi^2$ correlation method and a probabilistic $P(A|E, F)$ best-first search.

Results

The proposed constraint was evaluated on 50,000 English-French sentence pairs from the Hansard corpus, with a test set of 500 manually aligned sentences. The English sentences were parsed using Minipar to generate the required dependency trees.

Performance comparison of alignment methods with and without the cohesion constraint.

As shown in the results table, the cohesion constraint significantly improves alignment quality across different metrics:

  • $\phi^2$ Method: AER improved from 16.5% to 13.8%, with precision rising from 82.7% to 89.2%.
  • $P(A|E, F)$ Method: AER saw a 36% relative reduction, dropping from 13.6% to 8.7%. Precision reached 95.7%, while recall also saw a slight improvement (85.3% to 86.4%).

The cohesion-constrained probabilistic model outperformed the established IBM Model 4 baseline reported by Och and Ney (11.7% AER in their best refined configuration), demonstrating that the constraint effectively filters out statistically likely but syntactically invalid alignments.

Conclusion

The cohesion constraint provides a flexible, modular way to incorporate syntactic information into word alignment. By enforcing cohesion rather than strict isomorphism, the model accommodates natural translational divergences while significantly boosting precision. While the approach depends on high-quality source-side parsing, its success suggests that lightweight, syntax-based constraints can effectively prune search spaces and improve alignment accuracy in machine translation pipelines.

Summary updated

Abstract

We present a syntax-based constraint for word alignment, known as the cohesion constraint. It requires disjoint English phrases to be mapped to non-overlapping intervals in the French sentence. We evaluate the utility of this constraint in two different algorithms. The results show that it can provide a significant improvement in alignment quality.
BibTeX
@inproceedings{lin2003word,
  author = {Dekang Lin and Colin Cherry},
  title = {{Word Alignment with Cohesion Constraint}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2003},
  url = {https://openreview.net/forum?id=Sk4aAfW_-H},
}
Figure