Machine Translation Strategies: A Comparison of F-Structure Transfer and Semantically Based Interlingua

What it is about

The paper compares two development systems on the same task: English to Norwegian translation of technical text. The first, the Machine Translation Toolkit, is a transfer system whose grammars follow LFG; its transfer works on f-structures. The second, PONS, is declarative: a situation schema is a semantic description, neutral between analysis and generation, and two expressions are equivalent when they share one schema. PONS works in three modes. The first two translate directly or at a structural level, and the third is true interlingual translation, which runs only on the semantic representation. The author asks how far each strategy keeps translational equivalence and how easily the system can be extended. The text that could be read describes the modes and the feature structures, and does not give final scores; the conclusions on the comparison are hard to extract from the scan.

Ideas

• I1. Run the same test sentences through Mova's full path and a simpler baseline, and report which equivalences hold on each. The paper's design, one sentence set and two strategies, is a fair template. Count exact matches and divergence classes, run twice.
• I2. Test the declarative idea. In PONS one entry works in both directions, which is what the meta-world needs. Run the encoder in reverse on the reader's output and count round-trip failures by class.
• I3. Use the lowest mode that suffices. PONS falls back to the interlingual mode only when the simpler modes fail. Mova could copy the source for close authors and use the meta-world only when the residual is larger. Record how often each mode is used.
• I4. Note that PONS does not build parse trees at generation time. Check how often the encoder of Mova would need a tree, in light of open problem 3, before adding the tree step.

For Mova

The places are the translator's modes in tools/translate and the encoder's tree use in profile.rs. I2 and I3 test the design choices of the translator itself. The paper gives no quantities in the readable part, so the comparison has to be measured on Mova's data with the residual measure.

Topics: Machine translation

Cite

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

@inproceedings{thunes1994machine,
  title = {{Machine Translation Strategies: A Comparison of F-Structure Transfer and Semantically Based Interlingua}},
  author = {Thunes, Martha},
  booktitle = {Proceedings of the 9th Nordic Conference of Computational Linguistics (NODALIDA 1993)},
  year = {1994},
  url = {https://aclanthology.org/W93-0425/}
}

References

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

1. Banvise, J. and J. Perry. 1983. Situations and Attitudes. The MIT Press, Cambridge, Massachusetts. find
2. Bresnan, J. (ed.). 1982. The Mental Representation of Grammatical Relations. The MIT Press, Cambridge, Massachusetts. find
3. Dyvik, H. 1990. The PONS Project: Features of a Translation System. SKRIFTSERIE Nr. 39 S erie B, University of Bergen, Department of Linguistics and Phonetics. find
4. Executive Communication Systems, Inc. 1985. Machine Translation Toolkit. Utilities User’s Manual. Provo, Utah. find
5. Executive Communication Systems, Inc. 1985. Machine Translation Toolkit. LEGS User’s Manual. Provo, Utah. find
6. Fenstad, J.E., P.-K. Halvorsen, T. Langholm and J. van Benthem. 1987. Situations, Language and Logic. Reidel, Dordrecht. find
7. Karttunen, L. 1986. D-PATR - A Development Environment for Unification-Based Grammars. REPORT No. CSLI-86-61, Center for the Study of Language and Information, Stanford University. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/thunes-1994-machine-translation-strategies-comparison-f/