📄 Object-oriented semantics of English in natural language understanding system

Summary

Ostapov describes LEIBNIZ, an experimental English understanding system built in Visual Prolog at the Institute of Cybernetics in Kyiv. Its domain is human social behaviour. The system parses an English sentence with a grammar written in Backus-Naur style and turns the result into records of a semantic database. These records cover people, organizations, projects, actions, events, messages, places, times, machines, things, animals, relations, thoughts, causes and abstract objects. Questions are then answered with Prolog's inference engine.

A separate knowledge base explains word meaning in frame-like articles, which are themselves entered as plain English sentences. They describe what an object does or consists of, the motives behind an action, and multi-step operations. The author contrasts this with the Conceptual Dependency and MOP work at Yale and Stanford, where word knowledge was buried in program code. He argues that keeping linguistic data apart from code lets the method scale and move to other European languages.

The paper reports no quantitative evaluation. Its contribution is the design of the representation.

Key points

• The dictionary holds about 2,400 words. The system ran on Windows with Visual Prolog, and grammar structures are encoded as Prolog compound objects (functors).
• The semantic database distinguishes 17 kinds of predicate. Listed from last to first, these are number, abstr, cause, intelligence, relation, animal, thing, machine, tim, place, message, event, action, project, organization, job and person.
• The main records have many fields:
  • person has 21 fields, from appearance to temperament.
  • action has 18 fields: a semantic type in the style of Schank's primitives (PROPEL, MOVE, GO, TRANSFER…), real or supposed status, negation, tense and aspect, agent, patient, source, recipient, locations, instrument and purpose.
  • event has 12 fields describing a change of state.
• State changes are measured on scales. Health runs from −100 (death) to +100, hunger has a similar range, and physical scales such as Celsius are also allowed.
• Semantic analysis of a simple sentence follows a fixed order: fronted adverbial, then subject, then objects, then adverbials, and finally the predicate. The predicate type of a noun phrase comes from the semantic code of its head noun.

Interesting ideas

• I1. A state change is stored as an event with a named scale plus a start value and an end value, so "he fell ill" becomes a measurable shift rather than a label.
• I2. Each action record carries a flag saying whether it really happened or is only supposed, kept apart from negation and tense.
• I3. The knowledge base is written in restricted English (for example, an entry for doctor followed by "doctor examines a person") and parsed by the same analyzer that reads stories. A small set of built-in verbs grounds all other definitions.
• I4. Verb articles list typical motives together with the conditions under which each applies, and goal-directed "operations" are stored as staged scripts with alternative routes.
• I5. The author insists on keeping word knowledge out of program code and in queryable data, arguing that hand-coded semantics was what stopped earlier systems from scaling.

For Mova

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

• I1 (world): Mova's world states could carry an explicit scale and before/after values, which makes questions like "did things get better or worse?" answerable and traceable.
• I2 (frames): prag frames could keep a factuality status separate from polarity, so hypothetical, modal and negated actions never enter the world as facts.
• I3 (world): Compiled knowledge such as verb classes and ontology glosses could be authored as controlled English run through Mova's own pipeline, which keeps the knowledge explainable and self-consistent.
• I4 (answer): Readers answering "why did X do Y?" about fables could match an action against stored motive/condition pairs and script stages, and cite the matched stage as the explanation.
• I5 (learn): Self-induced repair rules and lexical knowledge should be stored as inspectable data that can be audited and tested with negative controls, rather than as Rust special cases.