Computer Generation of Multiparagraph English Text

What it is about

The paper argues that multi-sentence text should not be produced by cutting a knowledge base into sentence-sized pieces (the "Partitioning" paradigm), because real knowledge bases have no boundary marks and no adjacency for that. The KDS system instead uses Fragment-and-Compose. A fragmenter breaks the knowledge of a fire-alarm scenario into small propositional fragments. A problem solver orders them and factors them into paragraphs, using style rules. A knowledge filter marks fragments the reader already knows as "do not express", using a simple reader model (a set of known propositions). A hill climber then joins fragments with aggregation rules (common cause, conjoin, delete mid-state, if-then-else), scoring each candidate by a preference rule set: a start score of -1000 per fragment, -10 per embedded fragment, +100 for advised terms and time-sequential links, +200 for some constructions, and a large penalty for applying the same rule twice. The paper has no quantitative evaluation. It shows one scenario, with the input, the output of each module and the final text. The main reported finding is that brevity dominates: avoiding saying things matters as much as the choice of words.

Ideas

• I1. world/src/narr.rs: treat each fact of a lesson as a fragment and join fragments with a fixed table of aggregation rules (common cause, conjoin, delete mid-state). Keep the rules in a BTreeMap and apply them in a fixed order, so the output is deterministic.
• I2. world/src/knowledge.rs: a knowledge filter with a reader model made of the facts the student has already read in the lesson. Mark a fact "do not express" when it is known, and measure the repeated-content rate of the generated text with and without the filter. The filter off is the negative control.
• I3. Integer preference scores as an explicit, explainable table (start value, penalty per embedded fact, bonus per advised term, bonus per time link, penalty for a repeated rule). Each output then carries the list of contributions that explain its rank, which fits the transparency goal.
• I4. Brevity as a measured quantity: the number of sentences, words and repeated subjects per answer on the school lessons. Report it before and after each change of the rules, as the paper's "avoid saying things" finding suggests.

For Mova

The fragment and aggregation step belongs with the world: world/src/narr.rs and world/src/story.rs. The reader model and the filter fit world/src/knowledge.rs. Sentence joining relates to 36. Discourse relations and coherence; the referring phrases (first mention, reduced description, pronoun) relate to 35. Reference and coreference and the rules in coref/src/sieve.rs. Measures: sentence count and repeated-content rate on the school lessons, with a shuffled-fact negative control, run twice for determinism.

Cite

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

@article{mann1981computer,
  title = {{Computer Generation of Multiparagraph English Text}},
  author = {Mann, William C. and Moore, James A.},
  journal = {Computational Linguistics},
  year = {1981},
  url = {https://aclanthology.org/J81-1002/}
}

References

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

1. Badler, N.I., "The Conceptual Description of Physical Activities," In Proceedings of the 13th Annual Meeting of the Association for Computational Linguistics, AJCL Microfiche, 35, 1975. find
2. Carbonell J.R., and A.M. Collins, "Natural Semantics in Artificial Intelligence," In Proceedings of the Third International Joint Conference on Artificial Intelligence, 1973, 344-351. find
3. Davey, Anthony, Discourse Production, Edinburgh University Press, Edinburgh, 1979. find
4. Swartout, William R., "Producing Explanations and Justifications of Expert Consulting Programs," Technical Report TR-251, MIT Laboratory for Computer Science, January 1981. find
5. Heidorn, George E., "Natural Language Inputs to a Simulation Programming System," Technical Report NPS-55HD72101A, Naval Postgraduate School, 1972. find
6. Levin, J.A., and N.M. Goldman, "Process Models of Reference in Context," Research Report 78-72, USC/Information Sciences Institute, 1978. find
7. McDonald, D.D., "Natural Language Production as a Process of Decision-Making Under Constraints," PhD Thesis, MIT, Dept. of Electrical Engineering and Computer Science, 1980. find
8. Meehan, James R., "TALE-SPIN, An Interactive Program that Writes Stories." In Proceedings of the Fifth International Joint Conference on Artificial Intelligence, 1977. find
9. Nida, Eugene, Toward a Science of Translating, E.J. Brill, Leiden, 1964. find
10. Schank, Roger C., and the Yale A.I. Project, "SAM - - A Story Understander," Research Report 43, Yale University, Dept. of Computer Science, 1975. find
11. Simmons, R., and J. Slocum, "Generating English Discourse from Semantic Networks," Comm. A C M 15, 10 (October 1972), 891-905. William C. Mann is a member of the research staff of Information Sciences Institute at the University of Southern California. He received the Ph.D. degree in computer science from Carnegie-Mellon University in 1973. James A. Moore is a member of the research staff of Information Sciences Institute at the University of Southern California. He received the Ph.D. degree in computer science from Carnegie-Mellon University in 1974. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/mann-1981-computer-generation-multiparagraph-english-text/