📄 A Comparison of Pivot Methods for Phrase-Based Statistical Machine Translation

Summary

The paper tackles translation between language pairs that lack a direct parallel corpus, apart from a few hundred sentences used for tuning, by routing through English as a pivot. Using the Pharaoh phrase-based SMT baseline (GIZA++, SRILM, minimum error rate training on BLEU), the authors compare two pivoting schemes. PhraseTrans builds a source–target phrase-table by combining source–English and English–target tables: it multiplies phrase and lexical probabilities and sums them over the shared English phrases. SntTrans passes each sentence through two independent systems. It produces n English hypotheses, translates each one into the target language, and reranks the combined outputs with jointly tuned weights. Controlled experiments use the multi-way aligned Europarl corpus with Spanish, French and German, so a directly trained Direct system can serve as an upper reference. PhraseTrans was significantly better than both SntTrans variants and reached 0.92–0.97 of Direct performance. This shows that a pivot-induced phrase-table is a practical replacement when no direct bitext exists.

Key points

• Four systems were compared for each language pair: Direct, PhraseTrans, SntTrans15 (n = 15) and SntTrans1 (n = 1, which is plain sequential pivoting). Experiments covered Spanish, French and German, with English as the pivot.
• All systems used the 2006 NAACL/HLT shared-task baseline: the Pharaoh decoder, eight log-linear features and MERT with BLEU as the objective.
• SntTrans15 is roughly n times slower than its component systems. Translating 3064 test sentences took more than two days on a 3.06GHz Linux machine.
• Pivot-built phrase-tables held 168.2–190.8M phrases, compared with up to 18.2M in Direct tables. Measured against Direct, recall was about 34.7–44.1% and precision only about 1.8–3.4%, yet translation quality barely suffered.
• As a step toward realistic conditions, the training data was split into two disjoint halves of 292,915 sentences each. One half trained the source–English systems and the other trained the English–target systems.

Interesting ideas

• I1. Triangulating probability tables through a shared middle element. The source–target scores are obtained by marginalizing over common English phrases, assuming conditional independence given the pivot. The same trick could induce dictionaries or alignment tables for language pairs such as Ukrainian–X where only bitexts with English exist.
• I2. Fanout analysis to predict table blow-up. The authors estimate that the size of the combined table grows with the mean squared fanout rather than the mean fanout. For German–English, an average fanout of 1.5 implies a table roughly 29× larger. This gives a quick sanity check before composing any two mappings.
• I3. Recall matters more than precision for phrase-tables. Even very noisy induced entries did little harm because the decoder and language model filter them out. This suggests that generous candidate generation followed by downstream scoring is a reasonable design.
• I4. Controlled evaluation on multi-way aligned data. Europarl is aligned across all languages, so a direct system serves as an upper bound and pivot quality can be reported as a relative score. This design is reusable whenever a cleanly attributable comparison is needed.
• I5. Joint reranking across a pipeline. SntTrans tunes the feature weights of both stages together over n-best lists rather than chaining 1-best outputs. This is a general approach for any cascaded system, but its cost grows linearly with n.