A Study of Translation Rule Classification for Syntax-based Statistical Machine Translation

Hongfei Jiang, Sheng Li, Muyun Yang, Tiejun ZhaoNAACL 2009

Abstract

In syntax-based Statistical Machine Translation (SMT), the extraction of translation rules often leads to a "rule explosion," creating a massive and unorganized set of syntactic and non-syntactic rules. This paper presents a theoretical framework for categorizing these rules, introducing a refined taxonomy that distinguishes between various levels of lexicalization and generalization. By providing formal definitions for Structure Reordering Rules (SRR) and Discontiguous Phrase Rules (DPR), the authors offer a standardized method for analyzing rule utility, which is essential for optimizing translation quality and computational efficiency.

Introduction

Before the dominance of deep learning in natural language processing, Statistical Machine Translation (SMT) evolved from simple phrase-based models to more sophisticated syntax-based models. While phrase-based models were effective for local translations, they struggled with long-distance structure reordering and discontiguous phrases—expressions where words act as a single unit but are separated by other text (e.g., "switch the light off").

Syntax-based models, which utilize grammar trees, were designed to address these limitations. However, they introduced a new challenge: "rule explosion." Systems became cluttered with a mixture of conventional syntactic rules, pure phrase rules, and "non-syntactic" rules. Without a standardized classification system, it was difficult for researchers to determine which rules were beneficial and which merely added noise. This paper, presented at the 2009 Workshop on Syntax and Structure in Statistical Translation, proposes a granular taxonomy to solve this problem.

The Rule Classification Framework

The authors build their taxonomy upon two primary dimensions: lexicalization level and generalization capability.

Lexicalization Level

Rules are classified based on whether their leaf nodes are actual words or placeholders:

  • FLex (Fully Lexicalized): All leaf nodes are specific words.
  • ULex (Unlexicalized): All leaf nodes are placeholders (non-terminals such as NP or VP).
  • PLex (Partially Lexicalized): A combination of words and placeholders.

Generalization Capability

Rules are further categorized into Initial Rules and Abstract Rules. Initial rules consist entirely of words and lack placeholders, whereas Abstract rules contain at least one placeholder. Abstract rules are critical because they provide the generalization power that allows a translation system to handle a wide variety of sentence structures.

Refined Taxonomy for Abstract Rules

The core contribution of the paper is a refined classification of Abstract Rules based on the patterns of their alignment links—the connections between source and target words.

The proposed taxonomy for translation rules, categorizing them by lexicalization and structural properties.

Structure Reordering Rules (SRR)

The authors distinguish between two types of reordering:

  1. SRR_NT²: Reordering that occurs between two placeholders.
  2. SRR_NT-T: Reordering between a placeholder and a specific content word.

The authors argue that reordering involving content words (nouns, verbs, etc.) is more semantically meaningful for translation than reordering involving function words (e.g., "the", "of").

Discontiguous Phrase Rules (DPR)

The paper provides a strict mathematical definition for Discontiguous Phrase Rules. Previous definitions were often too broad, capturing sequences that could be explained by simpler, smaller phrase rules. The new definition requires a specific "crossing" pattern of at least three alignment links to confirm that a rule is truly discontiguous.

Alignment patterns illustrating simple reordering (a) and a true discontiguous phrase (b).

Conclusion and Future Work

The primary result of this study is a formal taxonomy that allows researchers to perform "ablation studies"—systematically removing specific categories of rules to measure their impact on translation metrics like BLEU scores.

While the paper is theoretical and does not include empirical results on standard corpora, it establishes the necessary groundwork for future research. The authors intend to implement this classification in real-world synchronous grammar-based systems to evaluate the specific contributions of each rule category to final translation quality and to optimize decoding speed.

Summary updated

BibTeX
@inproceedings{jiang2009study,
  author = {Hongfei Jiang and Sheng Li and Muyun Yang and Tiejun Zhao},
  title = {{A Study of Translation Rule Classification for Syntax-based Statistical Machine Translation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
}
Figure