Syntactic Models for Structural Word Insertion and Deletion during Translation

What it is about

The treelet translation system uses unlexicalised order templates for reordering, separate from lexical treelets. The authors let the templates also insert and delete unaligned words, such as English-to-Japanese case markers or Spanish de in nombre de archivo, without needing a content-word anchor. Unaligned nodes attach to their highest aligned neighbour when the target tree is projected, and some templates keep the inserted word as a lexical item. Inserted words are counted as template features. The gain was 1.1 BLEU over the treelet baseline and 3.6 over phrasal for English-Japanese (general-domain data), and 0.5 to 1.1 on Europarl English-Spanish, with about 0.7 on broad web text. Insertions were common (4,728 on the web test) and deletions rare (53), and deletion helped English-Japanese but not English-Spanish.

Ideas

• I1. Attach unaligned target words high. When the meta-world projects the target tree, attach unaligned tokens to the highest aligned neighbour, not the lowest. Count the unaligned tokens and the function-word residual on dev, run twice.
• I2. Structural insertion templates by POS. Learn from verse pairs rules such as NOUN NOUN to NOUN of NOUN, keyed by enum tags, and use them in the encoder when the concept is unseen. This gives the encoder a fallback below the concept-level contexts. Measure the function-word accuracy on unseen concepts.
• I3. Deletion only where it helps. Keep structural deletion off as a standard rule and allow it through the author DELTA, which the paper's direction finding suggests. Compare the standard with and without deletion per author pair, using the rule that a delta must fix at least three more tokens.
• I4. Tabulate the most common inserted and deleted words per target, as the paper does for Japanese, as a measure of which function-word classes the encoder gets wrong.

For Mova

The paper gives the template-level counterpart of open problem 4: function words inserted or deleted by structure, not by lexical context. The deletion result is a warning for the author delta: deletions may help one pair and hurt another, so they need to be per author.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{menezes2008syntactic,
  title = {{Syntactic Models for Structural Word Insertion and Deletion during Translation}},
  author = {Menezes, Arul and Quirk, Chris},
  booktitle = {Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing},
  year = {2008},
  url = {https://aclanthology.org/D08-1077/}
}

References

12 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Chiang, David. A hierarchical phrase-based model for statistical machine translation. ACL 2005. find our card
2. Crego, Josep, José Mariño and Adrià de Gispert. Reordered search and tuple unfolding for Ngram-based SMT. MT Summit 2005. find
3. He, Xiaodong. Using Word Dependent Transition Models in HMM based Word Alignment for Statistical Machine Translation. Workshop on Statistical Machine Translation, 2007 find
4. Heidorn, George. “Intelligent writing assistance”. In Dale et al. Handbook of Natural Language Processing, Marcel Dekker. 2000 find
5. Koehn, Philipp, Franz Josef Och, and Daniel Marcu. Statistical phrase based translation. NAACL 2003. find
6. Chi-Ho Li, Dongdong Zhang, Mu Li, Ming Zhou, Hailei Zhang. An Empirical Study in SourceWord Deletion for Phrase-based Statistical Machine Translation. Workshop on Statistical Machine Translation, 2008 find
7. Marcu, Daniel, Wei Wang, Abdessamad Echihabi, and Kevin Knight. SPMT: Statistical Machine Translation with Syntactified Target Language Phrases. EMNLP-2006. find
8. Menezes, Arul, and Chris Quirk. Microsoft Research Treelet translation system: IWSLT evaluation. International Workshop on Spoken Language Translation, 2005 find
9. Menezes, Arul, and Chris Quirk. Using Dependency Order Templates to Improve Generality in Translation. Workshop on Statistical Machine Translation, 2007 find
10. Och, Franz Josef. Minimum error rate training in statistical machine translation. ACL 2003. find
11. Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. ACL 2002. find our card
12. Wang, Wei, Kevin Knight and Daniel Marcu. Binarizing Syntax Trees to Improve Syntax-Based Machine Translation Accuracy. EMNLP-CoNLL, 2007 find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/menezes-2008-syntactic-models-structural-word-insertion/