Improving Statistical Machine Translation with a Multilingual Paraphrase Database

Ramtin Mehdizadeh Seraj, Maryam Siahbani, Anoop SarkarEMNLP 2015 · DOI

Introduction

Statistical Machine Translation (SMT) systems rely heavily on the breadth and quality of their training data. A persistent challenge in the field is the "Out-of-Vocabulary" (OOV) problem, where a system encounters words or phrases during inference that were absent from its training corpus. Traditional approaches often resort to copying the source word directly into the output or using basic transliteration, both of which degrade translation quality and disrupt syntactic reordering. This paper introduces a novel framework that leverages the Multilingual Paraphrase Database (PPDB) to address the OOV problem through graph-based label propagation.

Methodology

The proposed approach follows a three-stage pipeline to augment SMT phrase tables with translations for previously unknown terms.

Algorithm 1: The high-level workflow for integrating PPDB-based paraphrases into an SMT system.

1. Paraphrase Extraction

The foundation of the method is the PPDB, a massive repository of paraphrases generated via "bilingual pivoting." If two phrases in a source language (e.g., Arabic) are both translated to the same phrase in a target language (e.g., English) in a parallel corpus, they are identified as potential paraphrases. This method captures semantic equivalence more precisely than traditional monolingual context-based approaches.

2. Graph Construction and Propagation

The core of the system is a graph-based model where nodes represent source-language phrases.

  • Seed Nodes: Phrases with existing translations in the SMT phrase table.
  • Regular Nodes: OOV phrases requiring translation.
  • Edges: Weighted connections based on paraphrase likelihood scores from PPDB.

The researchers employ the Modified Adsorption (MAD) algorithm to propagate translation "labels" from seed nodes to OOV nodes. This allows an OOV term to "inherit" translations from its semantically related neighbors. To enhance connectivity, the authors utilize a tripartite graph structure, incorporating "bridge nodes" (unlabeled phrases) that link distant but related parts of the vocabulary.

The integrated framework showing the relationship between the phrase table, the paraphrase graph, and the resulting paraphrase pairs.

3. Phrase Table Integration

Once the propagation process is complete, the newly discovered translations for OOV phrases are integrated into the SMT system's phrase table. These new rules allow the decoder to treat previously unknown words as standard vocabulary during the translation process.

Experimental Setup and Results

The framework was evaluated across several challenging scenarios, including low-resource settings, domain adaptation, and morphologically complex languages.

  • Low-Resource Performance: In a French-English task using only 10,000 training sentences (Europarl v7), the system significantly outperformed existing graph-based baselines.
  • Domain Adaptation: When trained on parliamentary text but tested on medical (EMEA) or scientific (Science) corpora, the PPDB-augmented system effectively bridged the vocabulary gap, providing a substantial boost in translation accuracy.
  • Morphological Complexity: For Arabic-English translation, the method successfully identified translations for various morphological forms of the same root, resulting in a +1.53 BLEU improvement over the baseline on NIST OpenMT 2012 data.

Across all experiments, the PPDB-based method achieved BLEU score gains of up to 1.79 points. The results demonstrate that paraphrases derived from bilingual data are far more effective for SMT than those based solely on monolingual distributional profiles.

Discussion and Future Work

While the graph propagation method is highly effective, it is sensitive to "leakage"—the risk of propagating incorrect translations over too many graph hops. The authors mitigated this using early stopping and careful hyper-parameter tuning. Future research directions include extending the method to "infrequent" phrases (words seen too rarely for reliable translation) and integrating formal morphological analyzers to further refine the graph structure for complex languages.

Conclusion

By combining the semantic richness of PPDB with the reach of graph propagation, this work provides a robust solution to the OOV problem in SMT. The significant improvements in BLEU scores across diverse language pairs and domains underscore the potential of using large-scale paraphrase databases to enhance machine translation systems.

Summary updated

BibTeX
@inproceedings{seraj2015improving,
  author = {Ramtin Mehdizadeh Seraj and Maryam Siahbani and Anoop Sarkar},
  title = {{Improving Statistical Machine Translation with a Multilingual Paraphrase Database}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2015},
  doi = {10.18653/v1/d15-1163},
}
Figure