Two Languages are Better than One (for Syntactic Parsing)

David Burkett, Dan KleinEMNLP 2008 · OpenReview

Abstract

We show that jointly parsing a bitext can substantially improve parse quality on both sides. In a maximum entropy bitext parsing model, we define a distribution over source trees, target trees, and node-to-node alignments between them. Features include monolingual parse scores and various measures of syntactic divergence. Using the translated portion of the Chinese treebank, our model is trained iteratively to maximize the marginal likelihood of training tree pairs, with alignments treated as latent variables. The resulting bitext parser outperforms state-of-the-art monolingual parser baselines by 2.5 F1 at predicting English side trees and 1.8 F1 at predicting Chinese side trees (the highest published numbers on these corpora). Moreover, these improved trees yield a 2.4 BLEU increase when used in a downstream MT evaluation.

Introduction

Syntactic parsing—the process of determining the grammatical structure of a sentence—is a cornerstone of natural language processing. However, monolingual parsers frequently struggle with ambiguity, out-of-domain text, and noise. This research explores whether the mutual constraints of two languages in a bitext (translated sentence pairs) can improve parsing for both, even when high-quality monolingual parsers are already available. The central insight is that while a sentence might be syntactically ambiguous in one language, its translation often clarifies the intended structure.

Methods: Joint Parsing via Reranking

The authors propose a log-linear reranking model that evaluates triples consisting of a source tree, a target tree, and a node-to-node alignment between them. Instead of building a joint parser from scratch, the system reranks the $k$-best lists from existing state-of-the-art monolingual parsers to find the pair of trees most consistent with each other.

Consistency is measured through several features:

  • Alignment Features: Derived from word-level alignments. For example, the "InsideBoth" feature counts how many word-alignment points fall within the spans of two aligned tree nodes. If two nodes match, their underlying words should also align.
  • Structural Features: These measure tree similarity, such as span length differences and child node label correspondences.
Alignment features, such as 'InsideBoth', quantify the consistency between source and target tree nodes based on word-level alignments.

Because gold-standard node alignments are rarely available, the model treats them as latent variables and uses an iterative training procedure to maximize the marginal likelihood of training tree pairs.

Experimental Results

The model was evaluated using the translated portion of the Chinese Treebank (CTB), which provides gold-standard parses for both languages.

Parsing Performance

The joint model significantly outperformed independent monolingual baselines across the board:

  • English Parsing: Improved by 2.5 $F_1$ points.
  • Chinese Parsing: Improved by 1.8 $F_1$ points, achieving the highest published numbers on this corpus at the time.
Bilingual constraints allow the model to correct errors (a) made by monolingual parsers, resulting in more accurate joint parses (b).

Downstream Impact on Machine Translation

Beyond parsing metrics, the improved trees were tested in a downstream machine translation (MT) system. The higher-quality syntactic structures led to a 2.4 BLEU point increase, demonstrating that better syntax directly enhances translation quality.

Conclusion

This work demonstrates that bilingual constraints are a powerful tool for resolving syntactic ambiguity. While the current reranking approach is limited by the initial $k$-best lists and carries higher computational costs (roughly 2-4 times slower than monolingual parsing), the substantial gains in both parsing and translation suggest that joint bilingual modeling is a promising direction for NLP. Future work may involve moving toward fully integrated joint decoders to bypass the limitations of reranking.

Summary updated

Abstract

We show that jointly parsing a bitext can substantially improve parse quality on both sides. In a maximum entropy bitext parsing model, we define a distribution over source trees, target trees, and node-to-node alignments between them. Features include monolingual parse scores and various measures of syntactic divergence. Using the translated portion of the Chinese treebank, our model is trained iteratively to maximize the marginal likelihood of training tree pairs, with alignments treated as latent variables. The resulting bitext parser outperforms state-of-the-art monolingual parser baselines by 2.5 F1 at predicting English side trees and 1.8 F1 at predicting Chinese side trees (the highest published numbers on these corpora). Moreover, these improved trees yield a 2.4 BLEU increase when used in a downstream MT evaluation.
BibTeX
@inproceedings{burkett2008two,
  author = {David Burkett and Dan Klein},
  title = {{Two Languages are Better than One (for Syntactic Parsing)}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2008},
  url = {https://openreview.net/forum?id=H1NZrfGO-S},
}
Figure