TL;DR
This paper tackles the challenge of building Statistical Machine Translation (SMT) systems for low-resource language pairs, like Japanese to Hindi, where large parallel corpora are scarce. They demonstrate that by using multiple "pivot" languages (intermediate languages with existing parallel corpora), they can significantly improve translation quality, even when the direct parallel corpus is small. Their key innovation is the effective use of the Moses decoder's Multiple Decoding Paths (MDP) feature to combine translations from multiple pivot languages simultaneously.
What Problem Does This Paper Address?
Building high-quality machine translation systems typically relies on massive parallel corpora (millions of sentence pairs). However, many language pairs, especially those outside major language families, lack such resources. This "low-resource" problem severely limits translation quality. While using a single "pivot" language (e.g., translating Japanese to English, then English to Hindi) is a known strategy, this paper addresses the limitations of this approach when the pivot and target languages are from different domains or when the direct corpus is very small. They propose a more robust multilingual approach by leveraging multiple pivot languages to overcome the scarcity of direct parallel data.
Methods
The core of their approach is pivot-based SMT, specifically focusing on how to best combine information from multiple pivot languages.
- Phrase Table Triangulation: They use a standard method to build source-pivot and pivot-target phrase tables. These tables are then combined to create a source-target phrase table.
- Phrase Table Combination Techniques:
- Linear Interpolation: This method combines phrase tables by taking a weighted average of their probabilities. They explore different weighting schemes, including fixed ratios (9:1) and ratios based on individual pivot table performance (BLEU scores).
- Fillup Interpolation: This technique adds phrase pairs from secondary tables only if they are not present in the primary table.
- Multiple Decoding Paths (MDP): This is their key method. Instead of merging phrase tables into one, MDP allows the Moses decoder to explore translation options from multiple phrase tables (direct and various pivot-based) simultaneously during decoding. This preserves the distinct contributions of each table.
- Multilingual Setting: They leverage a small multilingual corpus containing translations of the same sentences across many languages, allowing them to extract multiple source-pivot and pivot-target phrase tables from a single dataset.
Datasets
The primary dataset used is a Bible domain multilingual parallel corpus.
- It contains sentence alignments across 9 languages.
- After alignment, they obtained 29,780 sentence tuples.
- This was split into 29,000 training tuples, 280 tuning tuples, and 500 testing tuples.
- The focus language pair is Japanese to Hindi.
- They considered 7 pivot languages: Chinese, Korean, Esperanto, Paite, Marathi, Kannada, and Telugu.
Key Insights and Ideas
- More Pivots are Better: Using multiple pivot languages significantly improves translation quality compared to using a single pivot, especially when the direct corpus is small.
- MDP is Superior for Multiple Pivots: The Multiple Decoding Paths (MDP) feature in Moses is more effective than linear interpolation for combining multiple pivot phrase tables. It allows the decoder to leverage the strengths of each pivot without diluting their individual contributions.
- Pivoting Induces New Phrase Pairs: Pivoting helps discover phrase pairs that are not present in the direct source-target corpus, effectively augmenting the translation knowledge base. This is crucial for low-resource scenarios.
- Morphological Similarity Matters, But Isn't Everything: While languages closer in morphology (like Korean to Japanese) can be good pivots, distant languages (like Esperanto) can also yield significant improvements, suggesting that other factors are at play.
- Bad Pivots Can Be Handled: While some pivots are clearly less effective (e.g., Telugu for Hindi-Japanese), MDP can effectively prune bad translation options, mitigating their negative impact.
Key Results
- Baseline: The direct Japanese-Hindi system achieved a BLEU score of 33.86.
- Single Pivot Improvement: Using the best single pivot (Esperanto for Jap-Hin, Korean for Hin-Jap) improved BLEU to 35.74 and 39.49 respectively.
- Multiple Pivots with MDP: Combining all 7 pivot tables with the direct table using MDP achieved a BLEU of 38.42 for Japanese-Hindi (a 4.56 point, or 13% relative, improvement over the direct baseline). For Hindi-Japanese, it reached 40.19.
- Significant Improvements: Many of the improvements achieved using multiple pivots with MDP were statistically significant (p<0.05) over the direct baseline.
- Unique Phrase Pairs: Each pivot language contributes a substantial number of unique phrase pairs, highlighting their role in expanding the translation vocabulary.
Limitations and Future Work
- Oracle Scenario: The paper notes that using BLEU scores from the tuning set for weighting in linear interpolation would be more appropriate than using test set scores (as done in some experiments for comparison).
- Selection of Pivots: While they found multiple pivots beneficial, they acknowledge that choosing the best set of pivots remains an open question. They suggest further research into effective pivot selection strategies.
- Large Corpora: The paper focuses on small corpora. It's unclear how much benefit pivoting would provide when already large parallel corpora are available.
- Morphological/Syntactic Treatment: The impact of deeper linguistic processing of pivot languages on translation quality is an area for future exploration.
- Reordering Information: Investigating whether pivoting can help extract better reordering information is suggested.
- Multi-Source/Multi-Pivot Synergy: Exploring the combined benefits of multi-source and multi-pivot settings is a future direction.
- Noise Control: Developing methods to control noise induced by pivoting, beyond simple probability cutoffs, is also proposed.
Summary updated