📄 A Rapidly Extensible Language System (REL English)

Summary

The paper presents REL English, one of the languages in the REL (Rapidly Extensible Language) system. REL is a conversational information system that lets people query and extend their own data bases without learning a programming language. The authors treat English as a family of idiolects. Each user gets a personal "version" that is tied to the structure of the data and can grow through definitions built from existing concepts or through new rules.

The central claim is that English behaves like a formal language when its subject matter is organized into a limited set of precisely structured categories. The limiting factor is how the data is structured, not the topic itself.

The grammar has two parts. A syntactic component combines context-free rewrite rules with transformational rules that produce deep structure as kernel sentences. A semantic component applies transformation rules directly to memory structures stored in ring form. The interpretation of a sentence follows its parse step by step and can reject phrases that have no meaning.

Ambiguity is resolved against the data where possible and kept when the data itself is ambiguous. Sample dialogues show income averages, time-bounded answers, an ambiguous result, and a user-defined phrase.

Key points

• REL is designed for large or small data bases accessed through highly individualized languages ("versions"). Users extend their version by defining new concepts and structures or by adding rules.
• The paper distinguishes two kinds of ambiguity:
  • Structural ambiguity, as in "Boston ships", is resolved by specifying the relation.
  • Semantic ambiguity, as in "King" referring to either a captain or a destroyer, is resolved by wider context.
  • The examples are linked to Chomsky's "Flying planes" sentence and to Katz and Fodor's "bachelor".
• The paper separates two kinds of restriction on language use:
  • Structural restrictions come from how the data is organized.
  • Discourse restrictions come from the topic.
  • The authors argue that only structural ones make semantic processing feasible. For example, "collections of boys" and "boys' collections" are deliberately treated as synonymous because the data cannot tell them apart.
• Function words (all, of, what) carry no memory structure; referent words map to data structures.
• Quantifiers and conjunctions are handled by "generators" that check elements one at a time:
  • "and" yields an "all" generator.
  • "or" yields a "some" generator.
  • Generators can be nested, though the authors defer full treatment of nesting.
• In the demo, "When was the number of Smiths equal to 5?" returns two labelled alternative intervals. The user defines "the 1940s" inline and then uses it in a follow-up query.

Interesting ideas

• I1. Meaning is checked against the actual structure of stored knowledge: each semantic rule either builds an interpretation or reports that the phrase is meaningless, which prunes parses.
• I2. Ambiguity is resolved where the data allows it, but genuine ambiguity in memory is preserved and shown to the user as numbered alternatives rather than silently picking one.
• I3. Quantifiers are treated as generators that iterate over a set and test a clause per element. Universal readings need every element to pass, existential ones need at least one, and wh-questions collect the elements that pass.
• I4. The language is extended by user definitions phrased in the same language, such as defining a decade as a ten-year span, which immediately become usable concepts.
• I5. Coordination is mapped directly to quantificational behaviour: "and" works like a universal and "or" like an existential over the conjoined items.

For Mova

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

• I1 (world): Candidate frames could be validated against the world-of-text entities and holdings, and readings that fail to attach to anything known could be discarded with a logged reason.
• I2 (answer): Readers could return several labelled answers when the world graph really supports more than one referent or interval, keeping the explanation honest.
• I3 (answer): This gives a transparent, step-by-step evaluation of "all/some/which" questions over world entities, with each element's verdict traceable.
• I4 (world): Compiled knowledge could accept definitional sentences from text or teachers and turn them into new ontology entries or measure spans without retraining.
• I5 (frames): When building frames from coordinated arguments, the conjunction type could be recorded as a distributive all/some flag so later reasoning knows how to evaluate them.