Abstract: Yang Feng, Yang Liu, Haitao Mi, Qun Liu, Yajuan Lü. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing.
2009.
This paper introduces a novel lattice-based approach to combining outputs from multiple Machine Translation (MT) systems. Traditional methods rely on confusion networks, which are constrained by 1-to-1 word mappings. By utilizing word lattices, the proposed framework supports many-to-many phrase alignments, effectively preserving the integrity of multi-word expressions. Experimental results on Chinese-to-English translation tasks demonstrate significant improvements, with gains of up to 1.23 BLEU points over state-of-the-art confusion network baselines.
Introduction
In the field of Statistical Machine Translation (SMT), "system combination" has emerged as a powerful technique to enhance translation quality by merging the outputs of several different MT engines. The goal is to produce a "consensus" translation that is superior to any individual input. Historically, the standard data structure for this task has been the confusion network.
The Limitations of Confusion Networks
Confusion networks operate by finding a 1-to-1 mapping between the words of different candidate translations. However, natural language often requires $n$-to-$n$ mappings. For example, one system might translate a concept as "is fond of" (three words) while another uses "prefers" (one word).
To accommodate these differences, confusion networks must insert "null" ($\epsilon$) words. This process often results in "degenerate" or "Franken-phrase" translations—nonsensical hybrids like "is fond to" or "prefers of"—because the network lacks the structural flexibility to treat multi-word phrases as single, atomic units.
The Lattice-Based Framework
To overcome these limitations, the authors propose a framework based on word lattices. Unlike confusion networks, lattices allow edges to represent entire phrases and can span multiple nodes, naturally supporting many-to-many alignments.
The proposed system follows a five-step pipeline:
- Backbone Selection: A primary translation is chosen as the "backbone" to determine the global word order. This is achieved via Minimum Bayes Risk (MBR) decoding, selecting the hypothesis with the lowest average edit distance from all others.
- Hypothesis Alignment: The system employs an Indirect-HMM (IHMM) alignment method. By aligning candidates to the backbone in both directions and applying a "grow-diag-final" algorithm, the model extracts robust many-to-many phrase mappings.
- Alignment Normalization: Candidate hypotheses are reordered to match the backbone's flow, ensuring that aligned phrases remain contiguous and coherent.
- Lattice Construction: A directed acyclic graph is built where edges can represent multi-word expressions. This preserves the semantic integrity of phrases like "is fond of" as a single unit.
- Decoding: The optimal consensus translation is found by searching the lattice using a log-linear model. This model incorporates features such as language model scores, arc posterior probabilities, and word counts.
Experimental Evaluation
The researchers evaluated the lattice-based approach on NIST Chinese-to-English translation tasks (MT02, MT05, MT06, and MT08). The system was trained using the FBIS corpus and the XinHua portion of the Gigaword corpus.
Key Results
The lattice-based method consistently outperformed both individual MT systems and traditional confusion network combinations:
- BLEU Improvements: On the NIST MT05 set, the lattice system achieved a score of 35.80 BLEU, representing a 1.23-point gain over the IHMM confusion network baseline and a 3.73-point improvement over the best single system.
- Robustness: Similar trends were observed across the MT06 and MT08 sets, with the lattice system maintaining a lead of approximately 3.0 BLEU points over the best individual engine.
- Dictionary Sensitivity: The study found that semantic alignment benefits from a medium-sized bilingual dictionary. While a 289K-pair dictionary yielded the best results, an excessively large dictionary (2.5M pairs) introduced noise that slightly degraded performance.
Conclusion
By transitioning from word-level confusion networks to phrase-level word lattices, this research provides a more flexible and linguistically sound method for MT system combination. The ability to handle $n$-to-$n$ mappings prevents the creation of disjointed translations and significantly boosts overall fluency and accuracy. Future work may focus on reducing the dependency on the initial backbone selection and exploring more advanced alignment techniques to further expand the lattice's expressive power.
Summary updated