📄 Processing English With a Generalized Phrase Structure Grammar

Summary

The paper reports interim progress from a project at Hewlett-Packard's Computer Research Center, running since summer 1981. The project asks whether Generalized Phrase Structure Grammar (GPSG), a strictly context-free formalism, can drive a practical English understanding system. The implemented pipeline has five parts:

• a GPSG grammar made of a lexicon, rules and metarules;
• a top-down parser with a grammar compiler;
• a semantics handler that turns parse trees into first-order logic;
• a disambiguator that uses sortal information to rewrite those formulas into queries;
• the HIRE relational database, hosted in SPHERE.

Following Montague's Rule-to-Rule Hypothesis, every syntactic rule carries its own semantic translation. Metarules derive related rules, such as passives from actives. Gap-carrying categories handle the long-distance dependencies in questions and relative clauses without leaving context-free power. The fragment covers many core English constructions. A worked example shows quantifier handling via Cooper storage.

The authors argue that syntax and compositional semantics are fully domain-independent. Porting to a new application would then mean only retuning the disambiguator and adding vocabulary. The paper gives no quantitative evaluation. Its contribution is an architecture that ties linguistic theory, model-theoretic semantics and database access together.

Key points

• Architecture: GPSG grammar → top-down parser and grammar compiler → logic transducer producing first-order logic → sortal disambiguator → query language of HIRE, a relational database inside the SPHERE system.
• Rule format: each rule is a triple of rule name, syntactic node-admissibility condition and semantic combination. An example is VP109, a verb taking two N!! objects, where ! marks X-bar levels.
• Coverage: the implemented subset handles prepositional phrases, noun-noun compounds, the auxiliary system, genitives, questions, relative clauses, passives and existential sentences.
• Lexicon: entries hold syntactic features, including rule-name features that encode strict subcategorization and irregular morphology. They also hold semantics: either a direct HIRE predicate or, for verbs and prepositions, case-role assignments.
• Theory and scope: the paper leans on Gazdar & Pullum (1982), who rebut claims that natural languages are not context-free. The current goal is limited to literal meaning of isolated sentences, with pragmatics and discourse left for later.

Interesting ideas

• I1. Pairing every syntactic rule with exactly one semantic composition makes every piece of meaning traceable to the rule that built it.
• I2. Metarules state a relation once, for example passive derived from active, and generate the derived rules automatically while preserving subcategorization.
• I3. Cooper storage keeps a quantified noun phrase in a side store until the clause is complete, then binds it with a placeholder variable, which postpones scope decisions.
• I4. Encoding in the lexicon which rules a word may appear in gives a cheap, explicit licensing check on structures.
• I5. Domain-independent grammar and logic, with all application-specific mapping confined to a thin sortal disambiguator, makes porting a matter of swapping one module.

For Mova

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

• I1 (frames): each UD relation pattern used by prag could be bound to one named frame-building rule, so every role in a frame cites the construction that produced it.
• I2 (frames): VerbNet role mappings could be written once for active clauses, with passive and other alternation variants (nsubj:pass, obl:agent) derived mechanically instead of hand-listed.
• I3 (world): phrases like "every applicant" could be stored at clause level and resolved into sets or universal statements when the world model is built, rather than becoming single entities too early.
• I4 (tree): a lexicon of verb argument frames could validate or rerank parser outputs, flagging trees where a verb appears in a construction it does not license.
• I5 (answer): keep frames and logical forms generic and put each test domain's (school tests, word problems, fables) mapping into a small swappable adapter.