📄 Syntax Augmented Machine Translation via Chart Parsing

Summary

The paper addresses how to bring target-side syntax into phrase-based statistical MT without giving up the coverage of ordinary phrase tables. The authors begin with a standard phrase table built from French–English Europarl data. They parse every English training sentence with Charniak's parser. Each phrase pair whose English side matches a constituent gets that constituent's label; all other pairs fall back to a generic "X". Following Chiang's Hiero approach, these labelled pairs are then turned into hierarchical rules by replacing sub-phrases with nonterminals, which yields a synchronous CFG. Decoding is done by chart parsing (a CKY+ variant over hypergraph nodes), with a log-linear model tuned by MERT. The system submitted to the workshop only used the n-gram language model to rescore a K-best list, and it did no better than the Pharaoh baseline. After a later change that brought the language model into K-best extraction, the syntactic systems clearly beat the baseline. The work shows that syntax-driven parsing decoders can scale to large data and that tight language-model integration is essential for them. The system (SAMT) is released under the GPL.

Key points

• Task and data: French-to-English shared task of the 2006 SMT workshop on Europarl. Weights were tuned by MERT on the 2005 development data. Evaluation used IBM BLEU on the '06 Development Test Set.
• Grammar size: Penn Treebank categories give about 75 nonterminals. The post-submission extended scheme gives about 4000, using labels such as NP+V, NP/NN and DT\NP for spans that do not match a single constituent.
• Systems compared:
  • Lex: phrase-decoder simulation.
  • XCat: single-X, Hiero-like.
  • Syn: treebank labels, up to 4 nonterminal sites per rule.
  • SynExt: extended labels, up to 2 sites per rule.
• Results with integrated LM and phrase length 7 (test BLEU): Baseline 30.61, Syn 31.31, SynExt 31.41. With phrases up to length 12, SynExt reaches 31.76, against 30.90 for the baseline. A difference of 0.78 BLEU corresponds to 95% significance.
• The submitted system (Syn, LM used only for rescoring) scored 30.58 on test, roughly equal to the baseline. CMU's phrase-based decoder scored 31.42 on test with the length-7 table.

Interesting ideas

• I1. Syntax as a label, not a filter. Phrase pairs that are not constituents are kept under a default label rather than thrown away. This adds linguistic information without losing coverage, a useful pattern whenever noisy annotation is layered onto existing resources.
• I2. Categorial-style labels for non-constituents. Spans that cover two adjacent constituents, or a constituent missing a neighbour on one side, get compound labels. This turns non-constituent spans into informative categories and could help any span-based model that has to deal with fragments.
• I3. Ablations that simulate rival systems inside one decoder. The authors reproduce a phrase-based decoder and Hiero within the same framework, so the gain from syntax is measured with the search procedure held fixed. This is a clean way to separate a modelling contribution from engineering differences.
• I4. Lazy, LM-aware K-best expansion. Only one back pointer is expanded at a time. The language-model cost is exact for words already produced and estimated from each unexpanded child's best hypothesis. The paper shows that this kind of integration, rather than rescoring afterwards, decides whether a richer model pays off.
• I5. Duplicate pruning in K-best lists. Partial hypotheses that produce the same target words from the same nodes are discarded, which increases list diversity. This is cheap and helps any rescoring or tuning step that relies on varied candidates.