Introduction
In the localization industry, professional translators often rely on Translation Memories (TMs) for their high fidelity and human-vetted quality. However, TMs are inherently rigid; they provide little utility unless a new sentence is an exact or near-exact match to an existing entry. Conversely, Statistical Machine Translation (SMT) offers broad flexibility but often lacks the domain-specific precision and stylistic consistency found in TMs. This paper presents a hybrid framework designed to "seed" SMT engines with high-quality fragments extracted from fuzzy TM matches, leveraging tree-based structural alignment to bridge the gap between these two technologies.
Methodology: A Structural Hybrid Approach
The proposed system integrates a TM backend, a Sub-Tree Alignment (STA) engine, and a Moses-based SMT backend. The workflow is designed to identify and preserve correct fragments from a fuzzy match while allowing the SMT engine to fill in the remaining gaps.
The process begins with Fuzzy Matching, where the system retrieves the best match from the TM for a given input sentence and calculates a Fuzzy Match Score (FMS). The core innovation lies in the Tree-Based Structural Alignment. Unlike traditional word-to-word alignment, this approach uses a phrase-based sub-tree aligner to map the input sentence to the TM source sentence, and subsequently to its target translation. This structural awareness allows the system to identify precisely which chunks of the TM translation are valid for the current input.
Once these chunks are identified, the system performs XML Seeding. The input sentence is marked up with XML tags that instruct the SMT engine to use specific TM-derived translations for certain segments. Finally, the SMT engine decodes the sentence, translating the "gaps" while adhering to the constraints provided by the TM seeds. To improve integration, a Reordering Tool pre-processes training data to align the source language's word order with that of the target language.
Experimental Setup and Datasets
The researchers evaluated the system using industrial data from Symantec Ireland, consisting of approximately 109,000 translation units of English-to-French customer support documentation. This dataset presented significant challenges, including "messy" content such as XML tags, file paths, and URLs, which required specialized tokenization and meta-tag handling.
Results and Discussion
The performance of the hybrid system (referred to as "xml") was compared against pure SMT ("direct") and raw TM output ("tm") using metrics such as BLEU, METEOR, and TER.
The evaluation revealed a distinct "sweet spot" for the hybrid approach. As shown in the performance analysis, the hybrid system significantly outperformed pure SMT when the FMS was between 80% and 100%. In the 70% to 80% range, the hybrid system achieved parity with pure SMT. However, performance dropped below that of pure SMT when the FMS fell below 70%. This suggests that at lower match scores, the noise introduced by incorrect TM fragments can confuse the SMT engine, highlighting the importance of the precision-oriented nature of the sub-tree aligner.
Conclusion
This research demonstrates that seeding SMT with structurally aligned TM fragments can significantly enhance translation quality, particularly for high-quality fuzzy matches. By prioritizing precision over recall in the alignment phase, the system ensures that only the most reliable fragments are passed to the SMT engine. Future work will focus on improving reordering capabilities for lower-quality matches and conducting human evaluation studies to measure the impact on post-editing efficiency in professional settings.
Summary updated