Time to eat peaches: language-specific information in interlingua

What it is about

This short position paper answers Farwell's argument (the card for farwell-1997-representing-language-specific-information-interlingua) that the interlingua should carry only communicative intent. The author accepts that intent is what matters for translation, but argues that language-specific information still has to be in the interlingua, for two reasons. First, the form of a source text can be stored compactly in its own language, so re-deriving everything as intent may not be practical. Second, retranslation: when the source is already a translation from the target language, the system should see the original wording. The example is a Japanese story that quotes an English poem; a system that loses the form cannot give back the English lines. The paper has no experiments and no numbers. The text that was extracted seems to end early, so the argument is only partly available.

Ideas

• I1. Turn the retranslation argument into a measure with the aligned Bible pairs. Read a verse from one English translation into the meta-world, encode it into another translation, and count the tokens where the output returns the original wording. Those tokens say how much the source form is needed.
• I2. Make the reuse of the source form an author delta: when the source surface form is cheap and the author's overrides fix at least three more tokens on half of the contexts, the encoder may reuse it. This uses the existing delta rule, with its threshold, rather than a new rule.
• I3. Compare the cost of the two representations. Measure the bits needed to store the source form with the bits needed for the intent-level description, using the residual measure. Keep the cheaper one for each class of token.

For Mova

The retranslation test in I1 is the natural check for tools/translate: the Bible pairs are already verse-aligned, and the residual gives the bit cost. The author delta in I2 belongs to the existing delta code. The paper bears on open problem 2 (a lexicon merge should lower the bits) and on the question of how much of each author's wording to keep.

Topics: Machine translation

Cite

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

@inproceedings{helmreich1997time,
  title = {{Time to eat peaches: language-specific information in interlingua}},
  author = {Helmreich, Stephen},
  booktitle = {AMTA/SIG-IL First Workshop on Interlinguas},
  year = {1997},
  url = {https://aclanthology.org/1997.mtsummit-workshop.6/}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/helmreich-1997-time-eat-peaches-language-specific/