📄 An English Generator for a Case-Labelled Dependency Representation

Summary

Tait describes a program that turns Boguraev's case-labelled dependency structures, the meaning representation of a Cambridge analyser, back into English. The work focuses on how to say a given content, not on what to say. Generation runs in two passes. The first maps the semantic dependency structure onto a surface syntactic tree. The second flattens that tree into a string, and the paper only sketches this step.

The design keeps four knowledge sources apart:

• bare phrase and clause templates
• a lexicon with morphology
• production rules keyed to case labels, which fill the templates
• rules that turn trees into text

A single recursive control loop drives all of them. The knowledge is mostly declarative and avoids assignment. The system can list every syntactically legal variant of an utterance and can substitute synonyms. It usually produces a sentence in a fraction of a second, which the author contrasts with unification-based generators. It was used in document retrieval and in a database front end. The paper matters as an early example of a modular, fast, rule-based dependency-to-text generator.

Key points

• The input is Boguraev's (1979) representation, which uses "case" very broadly: labels such as det, nmod and grammatical number count as cases alongside prepositional attachment. The running example is "the high frequency oscillator".
• Each case-label production rule receives four inputs:
  • the dependent item
  • an environment passed downward, carrying things like tense or how many paraphrases to produce
  • a partially filled template
  • the dictionary entry of the dominant word
• The control loop works in steps. It classifies the structure as clausal, simple nominal or complex nominal, then splits it into a head and a case list. It looks up the head's entry, picks templates, applies the rules recursively, and finally inflects the head for agreement.
• The implementation has a lexicon of a few hundred words. Its grammar covers much of the English noun phrase and some simple sentence types, and empty slots in trees are marked NIL.
• The generator grew out of Meehan's Micro-Mumble and is contrasted with Goldman's BABEL, Appelt's TELEGRAM and Jacobs (1983). The author argues it could be ported to Schank's Conceptual Dependency by swapping the rules, the lexicon and the dismantling functions.

Interesting ideas

• I1. Templates fix only the shape of a phrase, for example how many slots a verb phrase may have (no tritransitive verbs), while separate rules impose category and word-order limits; this split keeps structural and ordering constraints independently editable.
• I2. One generation rule is attached to each relation label, so producing output means walking the dependents of a head and calling that label's rule, recursing into sub-structures.
• I3. A downward-flowing environment carries context such as tense into embedded clauses and also acts as a control knob, for instance on the number of paraphrases.
• I4. Rules can return every legal alternative, including synonym variants, yielding the full set of syntactically valid paraphrases of one meaning.
• I5. Rules written declaratively without assignment could in principle run in reverse as a parser, though the author leaves this untested.

For Mova

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

• I1 (back): Generation from frames could use small UD-shaped skeletons per clause type and let separate ordering rules decide linearisation, so each failure points at one rule.
• I2 (back): A text → graph → text compressor could register one realiser per UD relation or frame role, so every generated word traces back to the label that caused it.
• I3 (back): Mova could pass tense, register and a paraphrase budget top-down through generation instead of storing them in global state, keeping each decision explainable.
• I4 (learn): Enumerating all realisations of a gold graph would give paraphrase sets for testing whether parsers map variants to the same frames, a useful consistency check.
• I5 (tree): Sharing one declarative mapping between UD trees and frames for both reading and generation would let a round trip expose disagreements between the two directions.