On representing language-specific information in interlingua

What it is about

This is a position paper arguing that an interlingua should not carry language-specific facts about the source text. What it should carry is the communicative act: who said what to whom, and why, and how. The author goes through four cases that seem to need language-specific material. In the first, the source talks about a word, so the form of the word matters. In the second, the source relies on sounds, as in a tongue twister. In the third, the source is a pun. In the fourth, the source is a joke built on a rhyme. In each case the author argues that the translator needs the speaker's goal and the effect, not the surface form, and that a target text can reach the same goal with its own means. The paper has no experiments and no numbers; the claim rests on examples. A reply by Helmreich in the same workshop argues that some language-specific information must stay.

Ideas

• I1. Test the claim as a rule for the meta-world: a source form (a quotation, a word as a mention, a rhyme) is kept on the meta token only when keeping it changes the target output. Measure the residual with and without the rule on dev, run twice.
• I2. Add a small closed set of intended effects to meta tokens (emphasis, quotation, joke, archaic register, mention). The encoder reads them as a context key. Counts can come from labelled lines of the author data.
• I3. Keep the source words on each meta token, as now, and use them only as a fallback when the intent of a token is unknown. The pun and the rhyme cases show where this fallback is needed.
• I4. Measure how many verses in the Bible pairs carry a mention, a quotation or a rhyme. Most probably carry none, so the claim matters more for school and story text. Say that honestly in the measure.

For Mova

The work touches the meta-world reader (tools/translate/meta.rs) and the encoder (profile.rs). The intent set in I2 is the same idea as open problem 8, style and register deltas per class, and open problem 4 for function-word features. The paper offers no measure, so the rule in I1 has to be proven on dev before it is kept.

Topics: Machine translation

Cite

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

@inproceedings{farwell1997representing,
  title = {{On representing language-specific information in interlingua}},
  author = {Farwell, David},
  booktitle = {AMTA/SIG-IL First Workshop on Interlinguas},
  year = {1997},
  url = {https://aclanthology.org/1997.mtsummit-workshop.5/}
}

References

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

1. Shakespeare, W. 1600. Hamlet. Electronic version at: The Complete Works of William Shakespeare, http://the-tech.mit.edu/Shakespeare/works.html. ↗
2. Bravo-Villasante, C. 1976. Una, Dola, Tela, Católa: el Libro del Folklore Infantil. Valladolid: Editorial Miñon. find
3. Helmreich, S., 1997. Workshop on Interlinguas at MT-Summit ’97. Electronic communication, 1 Jul 1997 Electronic version at: http://crl.nmsu.edu/Events/FWOI/call.html ↗
4. Helmreich, S. 1996. What Is an Interlingua and What Information should it Contain? Paper presented at the AMTA SIG-IL Pre-workshop on Interlinguas and Interlingual Approaches to Machine Translation, Montreal, Canada, October 1, 1996. Electronic version at: http://crl.nmsu.edu/Events/FWOI/PreWorkshop/prewkshop.papers.html. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/farwell-1997-representing-language-specific-information-interlingua/