📄 Interactive Semantic Analysis of English Paragraphs

Summary

Wilks tackles word-sense disambiguation at the paragraph level, assuming that a single sentence often gives too little context to choose a word's meaning. His LISP 1.5 system, run on the Q32, does not use a conventional syntactic parse. It cuts each paragraph into clause-sized fragments and encodes every word sense as a bracketed semantic formula built from 53 primitive elements. It then looks for templates, which are coded patterns of an elementary message such as "a kind of man is in some state". Chaining templates that fit together semantically produces a nested representation of the paragraph, and that representation keeps exactly one sense per word. The test material was ten paragraphs: five Times editorials and five passages from philosophers. When no coherent reading can be found, the program can enlarge its dictionary interactively by proposing new senses for an operator to accept. The output is printed in plain-English sense descriptions, so anyone can judge it without learning the code. Wilks also suggests that this kind of resolvability could serve as a graded test of whether a text is meaningful.

Key points

• The data was ten paragraphs. Five came from randomly picked Times editorials, and five from Descartes, Leibniz, Spinoza, Hume and Wittgenstein. A concordance of the texts was used to write the semantic dictionary.
• Each dictionary entry is a list of sense-pairs. Each pair joins a formula over 53 primitives (such as THING, STUFF, KIND, CAUSE, CHANGE) with an English gloss that is shown to the operator but not used in computation.
• Formulae are defined recursively, and their head is the last atom. A formula is well-formed only if its head belongs to a fixed subset of the primitives. Six bracket-group types guide how the dictionary maker builds formulae.
• Before template matching, fragments are re-ordered so that qualifiers precede the word they modify. For example, "a book of rules" becomes "a of rules fo book".
• When every reading path is blocked, a constructor mode suggests a new sense (for instance, attribute as nature) and asks the operator to confirm it.

Interesting ideas

• I1. Text is first cut into units sized between a word and a sentence, and complete clauses are lifted out of their host sentence. This yields cleaner semantic chunks even when the chunks look ungrammatical, such as a singular subject left with a plural verb.
• I2. Senses are chosen by coherence across fragments. Two fragments that both say roughly the same thing, such as "a structure is of some sort", decide the reading, which is how "habits" comes out as behaviour patterns rather than clothing.
• I3. Every internal code carries a human-readable gloss that only appears in the output. People can then judge the results without knowing the formalism.
• I4. Each primitive has a negation class of partly synonymous but mutually exclusive elements; for example, a thing cannot be both STUFF and THING. This gives a principled way to detect clashes.
• I5. An argument counts as "valid and useful" only when it is formally valid and also gets a complete and consistent sense assignment, meaning one sense per key term everywhere it occurs. The crown/coins syllogism fails this test.

For Mova

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

• I1 (frames): clause-sized sub-units cut from the UD tree could be the input for building frames, so that a single frame is never forced to straddle a coordinated subject.
• I2 (world): when a lemma has several WordNet senses, prefer the sense that makes the paragraph's entity and state assertions agree with each other, not the sense that best fits the local sentence.
• I3 (learn): attaching plain-English glosses to frame and sense decisions would let reviewers, or LLM teachers, check silver data without having to read the internal notation.
• I4 (world): contrary sets over ontology types could flag entities assigned incompatible types and trigger re-resolution.
• I5 (answer): before accepting a chain of reasoning on a test or fable question, check that each key term keeps one sense throughout, to catch equivocation.