TL;DR
This paper tackles the problem of "noisy" translation models in phrase-based Statistical Machine Translation (SMT) that lead to slower and less accurate translations. They propose a novel "triangulated filtering" approach using a third, intermediate language to identify and remove redundant or incorrect translation rules, significantly reducing model size and improving translation quality.
What Problem Does This Paper Address?
Statistical Machine Translation (SMT) systems learn to translate by extracting translation rules (phrases) from large parallel corpora. While more data generally leads to better systems, it also results in larger, more complex translation models. These large models increase computation time and resource demands during translation. More critically, the automatic extraction process, especially from word alignments, introduces "noise" – incorrect or less probable translation rules. This noise can lead to suboptimal translation choices and degrade the overall quality. Existing filtering methods have only offered marginal improvements. This paper aims to address both the efficiency and quality issues by reducing the size of these noisy translation models without sacrificing translation accuracy.
Methods
The core contribution is the triangulated filtering approach. Instead of directly filtering the source-target translation model, the authors introduce a third, "bridge" language. The key idea is to leverage the intersection of information from two "bridge models": one mapping the source language to the bridge language, and another mapping the target language to the bridge language.
Here's how it works:
- Construct Bridge Models: Two translation models are built: Source-to-Bridge and Target-to-Bridge.
- Filter Source-Target Model: For each phrase pair in the original Source-Target model, the system checks if both the source phrase and the target phrase can be mapped to common phrases in the bridge language via the respective bridge models.
- Intersection: If a common link exists through the bridge language, the phrase pair is considered more reliable and is kept. If no such common link is found, the phrase pair is removed from the original model.
- Re-optimization: After filtering, the feature weights for the reduced model are re-optimized using Minimum Error Rate Training (MERT) to adapt to the new, smaller model.
This method contrasts with traditional "union" approaches in multilingual SMT, aiming for a more precise and compact model by focusing on shared information.
Datasets
The experiments were conducted using the Europarl corpus (Release v3), specifically focusing on Spanish-English translations.
- Parallel Corpora: The Europarl corpus was used to build both the main translation models and the bridge models.
- Subsets: The authors created subsets of the Europarl corpus based on maximal sentence length ("EP-20", "EP-40", "EP-50") to build translation models of varying sizes.
- Bridge Languages: For Spanish-English, the bridge languages used were French (fr), Portuguese (pt), Danish (da), German (de), and Finnish (fi).
- Test Data: A shared task test set from the third Workshop on Statistical Machine Translation (2008) was used for evaluation. Development data was excluded from training.
- German-English Experiments: Additional experiments were conducted on German-English translation using Dutch (nl), Spanish (es), and Danish (da) as bridge languages.
Key Insights and Ideas
- Noise is a Major Bottleneck: The paper strongly emphasizes that noise in phrase-table entries is a significant issue, not just for efficiency but also for translation quality. Incorrectly aligned phrases can lead to translations with opposite meanings.
- Intersection Over Union: The core insight is that using the intersection of information through a third language is more effective for filtering than simply combining information from multiple languages (union). This intersection acts as a strong signal of reliability.
- Triangulation as a Filter: The bridge language acts as an intermediary, allowing the system to "triangulate" the validity of a source-target phrase pair. If two different paths (source-bridge and target-bridge) converge on a common concept in the bridge language, the original pair is likely correct.
- Smaller Models Can Be Better: The results demonstrate that significantly reduced models can achieve higher or comparable translation quality, challenging the "bigger is always better" assumption for SMT models.
- Bridge Language Matters: The choice of bridge language has a noticeable impact on filtering effectiveness and translation quality, suggesting that language similarity and typological characteristics play a role.
Key Results
- Significant Model Reduction: The triangulated filtering approach achieved substantial reductions in model size, with reductions of up to 94% in phrase-table and reordering table size.
- Quality Improvement: Despite model reduction, translation quality improved, with gains of up to 2.3 BLEU points.
- Efficiency Gains: Reduced model sizes led to faster translation times. For example, a reduction of 939 seconds (almost 16 minutes) was observed for translating 2,000 sentences.
- Robustness: The approach showed robustness across different model sizes and bridge languages, with quality gains converging for larger models.
- German-English: Similar positive results were observed for German-English translation, with size reductions of 40-85% and translation speed increases up to 17%.
- Compactness: Filtered models were more compact and contained more useful information per entry than naive pruned ones. In one extreme case, a system achieved reasonable quality using only 1.4% of the original phrase table, reducing disk space from 12.5 GB to under 100 MB.
Limitations and Future Work
- Fixed Filtering Decision: The current filtering decision for a phrase pair is binary (keep or discard). The authors note that this is fixed and doesn't consider probability distributions directly.
- Further Filtering: They propose a simple extension to further filter tables based on probability distributions of phrases, which could lead to even greater reductions.
- Language and Corpus Diversity: The authors suggest exploring their approach with more diverse languages (e.g., Chinese, Arabic) and different types of corpora to verify the conclusions.
- Language Clustering: The paper hints at language similarity/dissimilarity influencing results, suggesting further investigation into language clustering and its impact on triangulation.
- Re-optimization Necessity: The need to re-optimize feature weights after filtering is acknowledged, as weights optimized for a larger model may not be optimal for a reduced one.
Summary updated