📄 Book Reviews: Computations from the English

Summary

This is a short 1986 review by Alessandra Giorgi of Robert F. Simmons's book Computations from the English (Prentice-Hall, 1984). Giorgi sees the book as doing three things. It surveys AI work on language: parsing, question answering, questioning and summarizing texts, paraphrase, translation, and knowledge systems. It teaches procedural logic as a way to program these tasks. And, without saying so, it treats cognition as something to be modelled with algorithms. She values the book as a survey and as a programming introduction. Her main criticism is that its linguistics and cognitive psychology are thin and sometimes outdated. She doubts that serious systems can be built without that knowledge, naming translation, question answering and large vocabularies as examples. She expects that within a few years good products will need linguists and psychologists, and that AI should eventually yield findings about human cognition. Her verdict is that the book works as an introduction to procedural logic but offers little new to linguists or psychologists.

Key points

• Reviewed book: Robert F. Simmons, Computations from the English, Prentice-Hall, Englewood Cliffs, 1984, x + 308 pp., ISBN 0-13-164640-0. The review appeared in Computational Linguistics, Vol. 12, No. 1 (1986).
• Topics the book covers:
  • natural language parsing
  • question answering
  • questioning and summarizing texts
  • paraphrasing and translation
  • a survey of knowledge systems
• The book's method is procedural logic applied to natural language analysis and generation.
• The central criticism is the book's rough treatment of formal linguistics and cognitive psychology. Giorgi argues the book should at least point readers to further reading in those fields.
• The page also contains the end of a separate review by Helen Gigley of another book. That review stresses the neglected issue of process control and Marcus's case for a distinct syntactic component.

Interesting ideas

• I1. Giorgi argues that a question-answering system needs an explicit semantics of questions, not just matching against stored text, because what a question asks for depends on its form.
• I2. She argues that a large vocabulary is only manageable by modelling the morphological regularities of each language, not by listing words one by one.
• I3. She argues that translation needs a grammar theory that explains cross-language differences in a general way, rather than pair-specific hacks.
• I4. She treats cognitive tasks as explicit algorithms (the procedural view) and suggests such systems should also teach us something about human cognition, not only work.
• I5. She predicts that linguistic precision will soon become necessary for good products, even if it is optional for the moment.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (answer): Readers could classify each question by type (wh-target, yes/no, quantity) on the UD tree before searching the world of the text, so every answer can be traced to the asked-for slot.
• I2 (tokens): The en crate could favour rule-based inflection and derivation analysis with lexicon exceptions, which keeps lemmas explainable and covers unseen words.
• I3 (back): Translation through the meta-world could keep language-specific realization rules separate from a shared meaning graph, so adding a language means adding rules, not rebuilding the pipeline.
• I4 (learn): Evaluation could include probes that test whether Mova's step-by-step behaviour resembles known human patterns, alongside task accuracy, giving an extra check on whether the pipeline is principled.
• I5 (tree): Choices in UD dialect conversion and parser design could follow current linguistic analyses from the start, so later quality gains do not require redoing the treebanks.