Generation-Heavy Hybrid Machine Translation

What it is about

The paper proposes a way to translate between structurally divergent languages when only the target language is rich in resources. The source side needs only a dependency parser and a simple one-to-many word lexicon. There are no transfer rules and no interlingua. All the work of resolving divergences is moved into generation, using target resources only. The source dependency tree, glossed with target words, goes through five steps. Thematic linking turns syntactic relations into thematic roles, using English verb classes, a preposition-to-role map and a max-flow linking network. Structural expansion overgenerates conflations (give stabs → stab) and head swaps, using a categorial-variation database of 28,305 clusters over 46,037 words. Syntactic assignment applies subcategorization frames, and linearization builds a word lattice. Finally, a unigram and bigram language model extracts one sentence. The paper uses Dorr's five divergence types (categorial, conflational, structural, head swapping, thematic). On the first 48 verb divergences of a Spanish–English sample, 39 (81%) could be generated. Seven (14.5%) need conceptual knowledge, such as dar muerte → kill, and two (4%) are fixed idioms.

Ideas

• I1. Move divergence handling to the target side. Mova's meta-world reader could stay shallow, with the target encoder (profile.rs) generating the variants. Candidates are the conflation of light-verb + noun (make a covenant / covenant), head swaps and categorial variants. Each variant is scored by the target author's counts. That is the generation-heavy split in a count-based form.
• I2. A categorial-variation table for English, built from en/src/morph.rs derivations and the lexicon (hunger/hungry, death/die). It lets two meta concepts that differ only in category (his death / he died) count as one concept with a category feature. This is a candidate for the residual-driven lexicon (open problem 2). Keep a merge only if reader and encoder bits go down.
• I3. A preposition-to-role map, as in the paper's 341 links. For each English preposition, the set of frame roles it can mark, counted from prag frames on the Bibles. The reader then gives a PP a role and not just an attachment, which bears on attachment noise (open problem 1).
• I4. Report the translator's failures with the paper's three buckets: generable from lexical knowledge, needing conceptual knowledge, and fixed idioms. For each Hebrew→English verse with a large lexical residual, label which bucket it falls in. A sample of 50 verses would show how much the meta-world could gain from a categorial table alone.

For Mova

The paper is the closest published design to "standard encoder plus target-side knowledge" in tools/translate. Its overgenerate-then-rank step is what profile.rs would become if the encoder proposed several chunks per concept. The language-model ranker becomes the most-specific-context count. The divergence buckets of I4 give a measure for hebrew.rs.

Topics: Machine translation

Cite

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

@inproceedings{habash2002generation,
  title = {{Generation-Heavy Hybrid Machine Translation}},
  author = {Habash, Nizar},
  booktitle = {Proceedings of the International Natural Language Generation Conference},
  year = {2002},
  url = {https://aclanthology.org/W02-2125/}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/habash-2002-generation-heavy-hybrid-machine-translation/