Generalizing Hierarchical Phrase-based Translation using Rules with Adjacent Nonterminals

Hendra Setiawan, Philip ResnikNAACL 2010

TL;DR

This paper tackles a limitation in hierarchical phrase-based translation (Hiero) by allowing "adjacent nonterminal" rules, which were previously forbidden due to computational complexity and spurious ambiguity. They introduce methods to manage these rules, leading to improved generalization and significant gains in Chinese-English translation.

What Problem Does This Paper Address?

Hierarchical phrase-based translation (Hiero) is a powerful framework for machine translation that balances syntactic structure with data-driven approaches. It learns translation rules from parallel text, enabling it to handle both short- and long-distance reorderings effectively. However, a common constraint in Hiero implementations is to disallow rules where nonterminal symbols appear next to each other on the source side. This "non-adjacent nonterminals constraint" is primarily to avoid "spurious ambiguity," where multiple derivations lead to the same translation, thus complicating the decoding process and potentially reducing efficiency. The authors argue that this constraint severely limits Hiero's ability to generalize and capture important reordering phenomena, particularly in languages like Chinese where word order can be quite flexible.

Methods

The core contribution of this paper is the relaxation of the non-adjacent nonterminals constraint. To address the challenges this introduces, they propose two main strategies:

  • Computational Management: Instead of allowing all adjacent nonterminal rules, they focus on a subset called "XX-nonmono rules." These rules specifically model non-monotone reordering (where the order of phrases changes between source and target languages). This targeted approach helps manage grammar size and avoid spurious ambiguity by ensuring that non-monotone reorderings are always grounded in lexical evidence. Monotone reorderings are still handled by existing "glue rules."
  • Modeling Enhancement: They introduce a new model called the "target-side function words orientation-based model" (Porit). This model addresses the competition between the new XX rules and the existing glue rules. Porit leverages the "function words reordering hypothesis," suggesting that function words (like prepositions) encode crucial information about the reordering of neighboring phrases. Unlike previous work that focused on source-side function words, Porit analyzes the influence of function words on their target-side neighbors. This allows for a more direct evaluation of how XX rules and glue rules handle phrases adjacent to function words, leading to a more informed decoding process.

Datasets

  • Training Data: NIST MT06 Eval corpus (excluding UN data, approx. 900K sentence pairs) for Chinese-English translation.
  • Language Model Training: English side of the training data plus portions of the Gigaword v2 English corpus.
  • Development Set: NIST MT03 test set for feature weight optimization.
  • Evaluation Sets: NIST 2006 evaluation test (MT06) and NIST 2008 evaluation test (MT08).

Key Insights and Ideas

  • The "non-adjacent nonterminals constraint" is a significant bottleneck for Hiero's generalization: By disallowing these rules, Hiero misses opportunities to learn important reordering patterns.
  • Targeted relaxation is key: Not all adjacent nonterminal rules are beneficial. Focusing on rules that explicitly model non-monotone reordering (XX-nonmono) is crucial for managing complexity and avoiding spurious ambiguity.
  • Function words are powerful indicators of reordering: The Porit model demonstrates that analyzing the relationship between function words and their neighboring phrases on the target side can significantly improve translation quality, especially when combined with more flexible rule sets.
  • Combining rule relaxation with better modeling yields synergistic gains: The best results are achieved when both the XX-nonmono rules and the Porit model are employed, suggesting that these approaches complement each other effectively.

Key Results

The paper presents experimental results in an incremental fashion:

  • Baseline Hiero: Achieved BLEU scores of 30.58 (MT06) and 23.59 (MT08).
  • Adding ITG rules (+itg): Hurt performance consistently.
  • Adding general XX rules (+XX): Led to a grammar size increase of 25-26% but resulted in a performance drop, attributed to spurious ambiguity.
  • Adding XX-nonmono rules (+XX-nonmono): Increased grammar size by only 5% and yielded a marginal but consistent gain (30.96 MT06, 24.07 MT08).
  • Adding the Porit model (+orit): Showed mixed results, worse for MT06 but better (with significance) for MT08 (30.19 MT06, 23.69 MT08).
  • Combining XX-nonmono rules and Porit model (+XX-nonmono+orit): Produced a significant and consistent gain across both test sets (31.49 MT06, 24.73 MT08). This combination is highlighted as the best performing system.

Limitations and Future Work

  • Limitations:
    • The authors note that the Porit model's potential might not have been fully realized because Hiero, in its standard configuration, only considers monotone reordering for unseen cases.
  • While XX-nonmono rules performed better than general XX rules, the authors believe general XX rules might still be useful. Manual inspection of system output revealed that translations were sometimes shorter than references, with missing words potentially corresponding to function words that could be modeled by these more general rules.
  • Future Work:
    • Investigating the use of general XX rules to model legitimate word insertions.
  • Further exploring how to leverage XX rules to capture more complex reordering phenomena.

Summary updated

BibTeX
@inproceedings{setiawan2010generalizing,
  author = {Hendra Setiawan and Philip Resnik},
  title = {{Generalizing Hierarchical Phrase-based Translation using Rules with Adjacent Nonterminals}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2010},
}
Figure