Merged bilingual trees based on Universal Dependencies in Machine Translation

What it is about

The paper merges two parallel sentences into one dependency tree. Each node holds a pair of words, one from each language, with their two part-of-speech tags, and a word with no counterpart becomes a leaf marked as empty. The merge uses the two directional word alignments from GIZA++ on CzEng 1.0: a source word joins a target word only when the alignments agree, and the rest stay alone. Because the function words of Universal Dependencies are leaves, the two monolingual trees that come out of a merged tree have the same shape, apart from their leaves. The merged trees are parsed with an unsupervised parser, and a simple translation system is trained on them, without a language model or a reordering model. On WMT 2016 the system reaches BLEU 9.5 for English-to-Czech and 15.6 for Czech-to-English. The Czech-to-English result is close to another tree-based system, according to the abstract.

Ideas

I1. A merged tree is the meta-world in another form: the paper's pair node (source word, target word, both tags) is a count-based version of the meta token. Build the Bible meta-world as merged trees for each verse pair, and measure how many verses keep the same tree shape after the merge, as the paper's isomorphism claim suggests.
I2. Empty leaves for function words: the paper marks words with no counterpart as leaves with an empty partner, so the function words stay in the tree without a head decision. Use this for the function words of open problem 4 (determiners, "that", quotes): keep them as leaves of the meta token, and count their share per author.
I3. Alignment agreement as the merge rule: a link is used only when two directional alignments agree. Use the same rule for the lexicon's concept merges (open problem 2) and count how many merges the rule removes, compared with the current count thresholds.
I4. Unsupervised parsing of the merged sentence: the paper parses the merged sentence without gold trees. Test the same on the Hebrew-English verse pairs, and compare the parsed merged tree with the MACULA tree, to see how far the reader's attachments can be trusted (open problem 1).

For Mova

The code is tools/translate/src/meta.rs (the merge and the concepts) and tools/translate/src/hebrew.rs (the alignment). The merge is a count rule over two alignments, which fits Mova. The paper's BLEU figures are low and it uses no language model, so its numbers are a floor for the translator, not a target.

Topics: Universal Dependencies · Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w16-2318

@inproceedings{marecek2016merged,
  title = {{Merged bilingual trees based on Universal Dependencies in Machine Translation}},
  author = {Mareček, David},
  booktitle = {Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers},
  year = {2016},
  doi = {10.18653/v1/w16-2318},
  url = {https://doi.org/10.18653/v1/w16-2318}
}

References

2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In Proceedings of ICLR. find
Ondřej Bojar, Zdeněk Žabokrtský, Ondřej Dušek, Petra Galuščáková, Martin Majliš, David Mareček, Jiřı́ Maršı́k, Michal Novák, Martin Popel, and Aleš Tamchyna. 2012. The Joy of Parallelism with CzEng 1.0. In Proceedings of LREC2012, Istanbul, Turkey, May. ELRA, European Language Resources Association. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/marecek-2016-merged-bilingual-trees-based-universal/