Random generation of English sentences

What it is about

The paper writes English sentences at random from a phrase-structure grammar, so that people can judge the grammar. Rules are applied left to right, with a short last-in-first-out store for the parts still to be produced. Yngve hypothesises that English needs only about seven items in that store. The test grammar covers the first ten sentences of a children's book with 77 rules (24 of the form A = B + C, 26 of the form A = B, and choice rules). If each recursion were limited to one repetition, the paper counts over 10^20 sentences from a 38-word vocabulary. A run of 100 random sentences is mostly grammatical but nonsensical; the problems found are in coordination and in the animacy of adjectives such as "oiled".

Ideas

• I1. Test the seven-item store on Mova's own trees: measure the maximum number of pending constituents (or the maximum stack depth of a left-to-right traversal) of the UD trees of the EWT dev set, and report the share of trees above 7. Run twice, and check the figure against a negative control with reversed word order.
• I2. Random sampling as a grammar test, in en/src/lin.rs: generate strings from a small rule set, then let en parse-eval accept or reject them. Report the acceptance rate of generated strings and of corrupted strings. A good grammar accepts the first and rejects the second.
• I3. Coordination restrictions as a rule table in 24. Coordination and subordination: the same item may not occur twice in one series, and only items of one class may be coordinated. Count how many of the coordinated structures in the school lessons violate the table.
• I4. An animacy feature on adjectives and nouns, as the paper suggests, keyed in en/src/dict.rs: "oiled" with an animate subject is a different sense from "oiled" with an inanimate subject. Test on the school senses in world/src/dictionary.rs, with the feature as a deterministic check and not as a learned score.

For Mova

The rule-based grammar and the depth bound are for en/src/parse.rs and 20. Syntax: the Universal Dependencies frame. Coordination belongs in 24. Coordination and subordination. The animacy feature belongs in 15. Adjectives and adverbs and 19. The lexicon: vocabulary, multiword expressions and register, with code in en/src/dict.rs. Measures: the stack-depth distribution on UD EWT dev, the acceptance rate of random strings under en parse-eval, and the count of coordination violations on the school lessons.

Cite

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

@inproceedings{yngve1961random,
  title = {{Random generation of English sentences}},
  author = {Yngve, Victor H.},
  booktitle = {Proceedings of the International Conference on Machine Translation and Applied Language Analysis},
  year = {1961},
  url = {https://aclanthology.org/1961.earlymt-1.4/}
}

References

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

1. YNGVE, V. H. A Framework for Syntactic Translation, Mechanical Translation, 1959, 4, No. 3 59. find
2. CHOMSKY, N. Syntactic Structures, Mouton and Company, The Hague, 1957 MATTHEWS, G. H., Denumerability of the Sentences of Natural Languages American Mathematical Society, New York, April 1960. YNGVE, V. H., "A Model and an Hypothesis for Language Structure," Proc. Amer. Phil. find
3. LENSKI, L. The Little Train, Oxford University Press, New York, 1940 YNGVE, V. H. "A Programming Language for Mechanical Translation," Mechanical Translation, 1958, 5, No. 1 p. 25. (98026) find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/yngve-1961-random-generation-english-sentences/