📄 Book Reviews: English Word Grammar

Summary

This is Lynne Cahill's 1992 review of Richard Hudson's book English Word Grammar (Blackwell, 1990). The book presents Word Grammar (WG), formerly called Dependency Grammar and Daughter-Dependency Grammar, as a theory that avoids phrases and clauses. All syntax is stated as dependencies between words. WG now also claims to be a general theory of knowledge, linguistic and nonlinguistic alike. That knowledge sits in an inheritance hierarchy similar to a semantic net, with default inheritance.

Cahill credits Hudson for trying to bring WG closer to computational linguistics and AI. She finds the argument that language knowledge is continuous with world knowledge worth pursuing. Her criticisms are these:

• The English-like formalism is clumsy and hard to compute.
• The book covers only syntax in depth. Its morphology is limited to regular suffixation, and it gives no phonology.
• It offers no grammar statistics and only a small grammar listing.
• It brings in constituent structure ("word-string") just for coordination.

The review matters as an early assessment of dependency-based grammar from a computational point of view.

Key points

• The book has 445 pages. Seven chapters define the theory, and seven apply it to English, mostly syntax with a little morphology and semantics.
• In WG, black dogs eat cats gets three word-to-word dependencies: black–dogs, eat–dogs and eat–cats.
• Inheritance is blocked by explicit "NOT" propositions, unlike DATR's rule that more specific values override general ones. For example, go has the past form went, and a NOT statement blocks the regular stem + ed form. For dream, leaving out the NOT lets both dreamt and dreamed through.
• The appendices list only 14 verb entries and 22 noun entries. Partial implementations exist (Hudson 1989; Fraser 1987–1989), but the book does not discuss them.
• Coordination is the one place where WG uses a constituent-like unit, the "word-string".

Interesting ideas

• I1. Using explicit blocking statements instead of automatic override means a regular and an irregular form can coexist when the block is left out, which suits genuine variation like dreamt/dreamed.
• I2. Treating linguistic knowledge as a special case of general world knowledge means one inheritance network could hold word facts and world facts together.
• I3. The reviewer argues that a coordination mechanism should fit the main framework rather than be a special constituent device, a design caution relevant to any dependency system.
• I4. Cahill asks for figures, such as how many irregular entries need blocking versus how many need multiple forms, before accepting the cost of a design.
• I5. She warns that a formalism written in the same language it analyzes, such as using "it" as a variable, invites confusion.

For Mova

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

• I1 (tokens): the en lemmatizer/generator could keep irregular forms as overrides with an explicit "block regular" flag, so that variant pairs simply omit the flag and both forms stay valid.
• I2 (world): the compiled ontology and the lexicon could share one default-inheritance hierarchy, so verb-class and entity facts are inherited and explained in the same way.
• I3 (tree): handling conj/cc inside the ordinary UD dependency scheme, with only small adjustments, keeps parser and frame extraction uniform instead of adding a separate phrase layer for coordination.
• I4 (learn): when adopting a rule mechanism, count how many lexicon entries actually need it, as a cheap check of whether the mechanism earns its complexity.
• I5 (frames): keep frame and role notation clearly distinct from English surface words, so pronoun coreference in the text is never mixed up with references inside the formalism.