Abstract Standard word alignment models often fail to identify rare or idiomatic translations due to insufficient statistical evidence. This paper introduces a two-stage "transpotting" method that enhances the signal of infrequent translations by constructing artificial bitexts from non-parallel sentences. Evaluation on the Canadian Hansards corpus demonstrates that this approach significantly improves alignment accuracy for rare translations with minimal computational overhead, requiring as few as ten non-parallel sentence pairs to achieve substantial gains.
Introduction
In the field of Computer-Assisted Translation (CAT), professional translators rely on bilingual concordancers to search massive parallel corpora (bitexts) for historical translation examples. The task of identifying a specific translation of a query within a sentence pair is known as "transpotting." While common translations are easily identified, rare or idiomatic ones—often appearing only once or twice in millions of sentences—are notoriously difficult for statistical models to align correctly.
As shown in the distribution of translation candidates for the query wait and see, many infrequent candidates are either incorrect or only partially accurate. This data sparsity leads to a significant recall problem, frustrating professional users who depend on these tools for specialized terminology.
Methodology
The authors propose a two-stage approach to improve the alignment of these "rare birds" of translation:
- Stage 1 (Standard Transpotting): A baseline model (e.g., HMM or IBM Model 2) generates an initial distribution of translation candidates.
- Stage 2 (The Booster): The system focuses on the "lower tail" of the distribution—candidates with a frequency of one. To verify these candidates, the authors create an artificial bitext.
The "seed" sentence pair containing the rare translation is paired with other target sentences from the training data that also contain that specific translation. Although these new pairs are non-parallel (they do not share the same meaning), they both contain the query and the suspected translation. Running an alignment model on this artificial bitext reinforces the co-occurrence signal. The resulting "local" lexical distribution is then blended with the "global" distribution to re-align the original seed sentence.
Experimental Setup
The researchers utilized the Canadian Hansards, a bilingual corpus of parliamentary proceedings containing over 8.3 million sentence pairs. To evaluate the method, they developed two specific benchmarks:
- TESTFREQ: 1,516 sentence pairs containing frequent translations.
- TESTRARE: 591 sentence pairs containing rare translations (occurring only once in the corpus).
- DEVRARE: A development set of 115 pairs used for tuning the interpolation parameter.
Results and Discussion
The two-stage method yielded significant improvements across all tested alignment models when dealing with rare translations. For the bidirectional HMM model (C-HMM-bi), the percentage of correctly identified reference translations increased from 65.6% to 69.7%. When measuring "region spotting" (identifying at least one correct word in the translation phrase), the accuracy reached 86.5%.
A critical insight from the study is that accuracy gains plateau quickly. Most of the performance benefit is achieved using as few as 10 non-parallel sentence pairs. This efficiency makes the method highly tractable for real-world deployment in CAT environments, as it does not require massive amounts of additional data or compute time.
Conclusion
This paper demonstrates that "noisy" non-parallel data can be leveraged to overcome the data sparsity issues inherent in aligning rare translation events. By artificially reinforcing the signal of infrequent terms, the proposed two-stage method significantly improves recall for professional translators. Future work will investigate more intelligent sampling strategies for artificial bitext construction and the potential for applying these online adaptation techniques to standard Machine Translation decoding.
Julien Bourdaillet, Philippe Langlais. Proceedings of the 10th Conference of the Association for Machine Translation in the Americas: Research Papers. 2012.
Summary updated