Multi-lingual Sentence Generation from the PIVOT interlingua

What it is about

The PIVOT interlingua is a directed network of conceptual primitives, with arcs for deep cases such as agent and object, and with node information, pragmatic information (topic, focus, position) and scope (negation, comparison, quantifiers). The authors build Spanish and French generators by reusing the English generation model, which has four steps: conceptual wording, syntactic selection (choosing a verb pattern by a plausibility score with priorities), word ordering, and morphology. Most of the work is in dictionary content, so the new languages need only their own tables; the authors compare subjects, cases, voice and morphology across the three languages to see what can be shared. The dictionaries reported are 70,000 English, 5,000 Spanish, 3,000 French and 110,000 Japanese words. The abstract also reports a cost reduction for a translation service that used the system, with no further detail.

Ideas

• I1. Make the verb-pattern choice a counted priority: score each candidate frame for a predicate by how many of its deep cases map onto grammatical cases, and whether the subject is new or old. Use it in the encoder for concepts with few entries. Measure the frame accuracy on dev, run twice.
• I2. Split the encoder's work in two, as the paper does: first choose the syntactic frame, then the order. Record frame errors and order errors separately, so that the residual shows which step fails.
• I3. Test reuse of the standard across a second target: count how many entries of the profile change when only the target tables change. The paper's claim is that most do not.
• I4. Carry topic and focus as a feature of meta tokens, as the paper does for its pragmatic layer (see the card for sgall-1987-machine-translation-linguistics-interlingua).

For Mova

The places are tools/translate/profile.rs (the encoder, as the verb-pattern choice) and en/src/lin.rs (the ordering step). I1 and I2 touch open problem 3 (the encoder should write trees): choosing a frame before the order is the tree step. The paper gives no accuracy figures, so each claim needs a measure on Mova's data.

Topics: Machine translation

Cite

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

@inproceedings{okumura1991multi,
  title = {{Multi-lingual Sentence Generation from the PIVOT interlingua}},
  author = {Okumura, Akitoshi and Muraki, Kazunori and Akamine, Susumu},
  booktitle = {Proceedings of Machine Translation Summit III: Papers},
  year = {1991},
  url = {https://aclanthology.org/1991.mtsummit-papers.11/}
}

References

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

1. K.Muraki “VENUS: Two-phase Machine Translation System,” Future Generations Computer Systems, 2, pp.117-119, 1986 find
2. K.Muraki “Conceptual Dependency Structure and English Sentence Generation,” Proc. WGNLC of the IEICE, NLC44-3, July 1984 (in Japanese). find
3. S.Ichiyama “Multi-lingual Machine Translation System,” Office Equipment and Products, 18-131, pp.46-48, August 1989 find
4. C.H.Kim, B.H.Choe, G.C.Kim, K.Choi, S.Ichiyama, “Generation of Korean from Conceptual Representation,” Proc. Information Processing Society of Japan , 2B-8, pp. 947-948, October 1986 find
5. H.Suematsu “A Simultaneous Processing Approach to the Selection of English Sentence Patterns,” Proc. WGNLC of the IECIE, NLC89-41, May 1990 find
6. H.Suematsu “PIVOT J-E: An Overview of English Generation,” Proc. IECIE, D-141, October 1988 (in Japanese). find
7. Y.Fukumochi “PIVOT: Subject and Predicate Selection in English Generation,” Proc. Information Processing Society of Japan, 5C-3, pp.1094-1095, October 1984 (in Japanese). find
8. C.Boite “Multilingual Machine Translation does not have to be saved by Interlingua,” MMT'90, Tokyo, November 1990 find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/okumura-1991-multi-lingual-sentence-generation-pivot/