English Generation From Interlingua by Example-Based Method

What it is about

The paper describes an English generator that turns an interlingua (concepts, conceptual relations, and attributes such as tense and aspect) into a syntactic tree, using examples instead of rules. The examples are pairs of a small interlingua fragment and a small syntax fragment, cut from corpus sentences and dictionary entries. There are two example sets. The first is keyed by the central concept and chooses content words; it is searched by the similarity of concepts in the EDR concept hierarchy, counting shared ancestors up to three layers up. The second is keyed by attributes (tense, aspect, modality) and chooses inflection and function words, and is searched deterministically. The generator is built to test and correct the EDR dictionaries. Its main experiment is verb selection with a leave-one-out (jack-knife) test: remove one example, generate its verb from the rest, and compare. The rate at which the generated verb equals the original is 85% in Experiment 1, 13% in Experiment 2 and 16% in Experiment 3. The authors note that many non-matching verbs are still correct, so the true success rate is higher.

Ideas

• I1. world/src/knowledge.rs: a leave-one-out test for word choice. Remove one lesson sentence, generate its key word from the others, and compare. Report the rate per concept, and keep the rate per concept in a BTreeMap so that runs are comparable across releases.
• I2. Concept similarity from the shared ancestors in the school concept hierarchy, cut at a fixed depth (three layers, as the paper does), with the ratio of common to total ancestors as the score. Use it to rank candidate words for a concept. Keep ties broken by the word, in order.
• I3. Split the choice of word from the choice of form: a key-based table for the lexeme (get vs become), and an attribute-keyed table (tense, perfect, progressive) for the inflected form. Check the table with en/src/morph.rs inflection, and report the form accuracy separately from the lexeme accuracy.
• I4. Report the candidates that have no example and drop them from the list before scoring, as the paper does. Count how often this happens on the school lessons, since a candidate with no example can never win.

For Mova

Word choice for concepts belongs with the world: world/src/knowledge.rs and world/src/dictionary.rs (word senses). Inflection and tense forms belong in en/src/morph.rs, with the verb chapters 13. Verbs and 14. Irregular verbs. The verb-sense contrast (get, become, take) fits 19. The lexicon: vocabulary, multiword expressions and register. Measures: the leave-one-out rate per concept, run twice for identical output, and the lexeme and form accuracy reported apart.

Topics: Machine translation

Cite

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

@inproceedings{komatsu1994english,
  title = {{English Generation From Interlingua by Example-Based Method}},
  author = {Komatsu, Eiji and Cui, Jin and Yasuhara, Hiroshi},
  booktitle = {COLING 1994 Volume 1: The 15th International Conference on Computational Linguistics},
  year = {1994},
  url = {https://aclanthology.org/C94-1059/}
}

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. Cui, J., Komatsu, E. and Yasnhara, tl. (1993). A Calculation of Similarity between Words Using EDR Electronic Dictionary. Reprint of ll'SJ, Vol.93, No.1 (in Japanese) find
2. EDR (1993a). EDR Electronic Dictionary Specification Guide.TR.04 l. find
3. EDR (1993b). English Word Concept Dictionary. TR-026 find
4. Komatsu, E., Cni, J. and Yasuhara, II. (1993). A Mono-lingual Corpus-Based Machine Translation of the Interlingua Method. Fifth International Conference on Theoretical and Methodological Issues Nagao, M. (1984). A Framework of A Mechanical Translation between Japmlese and English by Analogy Principle. Artificial and Human Intelligence (A. Elithorn and R. Banerji, editors) Elsevier Science Publishers, B.V. find
5. Sadler, V. (1989). Working with Analogical Semantics, Disambiguation Tect, niques in DLT, Foris P u b l i c a t i o n s , Dordrecht Holland. find
6. Sato, S. (1991). E x a m p l e - B a s e d Translation Approach. Proc. (g'International Workshop on Fundamental Research for the Future Generation of Natural Language Processing, ATR Interpreting Telephony Research Laboratories, pp. 1-16. find
7. Sumita, g. and Iida, 11. (1992). Example-Based Transfer of Japanese Adnominal Particles into English. IEICE TRANS. INF. &SYST., VOL. E75-D, NO.4 find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/komatsu-1994-english-generation-interlingua-example-based/