📄 Machine Translation by Triangulation: Making Effective Use of Multi-Parallel Corpora

Summary

Phrase-based SMT degrades sharply when bilingual training data is scarce: phrase translation probabilities become unreliable and many longer phrases are never seen. Cohn and Lapata tackle this with multi-parallel corpora (here Europarl, covering many languages aligned at the sentence level). Their approach, triangulation, builds a source–target phrase-table by pivoting through a third language. The probabilities of the two bilingual hops are multiplied, and the result is summed over all pivot phrases. These triangulated tables can be built from several pivot languages and then linearly interpolated with the conventional direct phrase-table. The merged table goes straight into the decoder, so no separate system outputs need to be combined. Experiments on small Europarl samples across several language pairs show BLEU gains in every tested setting, with the largest gains for hard pairs involving German. Interpolation gives roughly the benefit of doubling the training data. Even without any direct bitext, the triangulated tables still produce usable translations. This matters for low-resource language pairs that lack direct parallel data but share translations with other languages.

Key points

• Model: p(s|t) ≈ Σᵢ p(s|i)·p(i|t). It assumes the pivot phrase carries all the information needed, so the target becomes redundant given the pivot. Pivot phrases missing from the other bitext are simply dropped.

• Merging uses linear interpolation of joint distributions with non-negative weights that sum to one. Both p(s|t) and p(t|s) from the merged table serve as decoder features.

• For fr→en with a 10K-sentence sample, Romance pivots (es, it, pt) work best, followed by Germanic ones. Greek and Finnish help least; Finnish alone scores −14.26 BLEU relative to the baseline, yet it still helps once interpolated. For en→de, Germanic pivots come out on top.

• Adding more pivot languages raises BLEU at every step, and the purely triangulated scores converge toward the interpolated ones. With disjoint halves of Europarl, triangulation via Spanish still works: 28.37 BLEU from the same half versus 28.13 from different halves, against a one-half baseline of 28.85.

• Phrase-table statistics compare the standard and triangulated tables:

  Table	Source phrases	Target phrases	Events	Overlapping mass
  Standard	8M	7M	12M	0.646
  Triangulated	2.5M	2.5M	70M	0.750

  The triangulated table covers fewer distinct phrases but contains far more phrase pairs.

Interesting ideas

• I1. Pivoting inside the model, not after decoding. Pivot information is folded into the phrase-table itself rather than by selecting among outputs of several systems. This keeps a single decoder and avoids system-combination overhead, an idea that carries over to any table-based transfer task, such as lexicon induction for a low-resource pair through a related language.
• I2. Interpolate instead of replace. The triangulated table is noisier because alignment errors compound across two hops, so it complements the direct table rather than replacing it. Mixing a precise but sparse resource with a broad but noisy one is a widely reusable pattern.
• I3. Choose pivots by relatedness. Pivots closely related to the source or target perform best, likely because word alignment is easier between such languages. For Ukrainian work, this suggests testing closely related Slavic languages as pivots first.
• I4. Isolate the factor under study. Minimum error-rate training was switched off when comparing pivot languages, so differences reflect the translation estimates rather than tuning. This is a useful habit for any ablation where an optimizer could mask the effect being measured.
• I5. Inspect resources directly, not only end-task scores. Besides BLEU, the authors compare phrase-tables via coverage counts, overlapping probability mass and Jensen–Shannon divergence. This intrinsic evaluation shows the pivot table adds new information rather than merely approximating the direct one.