EL: A Formal, Yet Natural, Comprehensive Knowledge Representation

Chung Hee Hwang, Lenhart K. SchubertAAAI 1993

Abstract

Episodic Logic (EL) is a highly expressive knowledge representation framework designed to bridge the gap between the expressive richness of natural language and the formal rigor required for AI reasoning. By treating situations, events, and propositions as first-class entities within a "liberal ontology," EL allows AI systems to perform common-sense reasoning and process natural language without losing the nuance of the original text. This paper details the pillars of EL, its implementation in the EPILOG engine, and its application to complex narrative and technical domains.

Introduction

Historically, knowledge representation (KR) has faced a persistent trade-off: systems were either computationally efficient but limited in expressivity, or highly expressive but mathematically cumbersome. Natural language is replete with complex constructs such as beliefs ("John thinks..."), generic statements ("Birds fly"), and intricate temporal relationships ("While Mary was dropping the glass..."). Traditional first-order logic often struggles to represent these without becoming incredibly convoluted. Episodic Logic (EL) addresses this by proposing that an AI's internal "language of thought" should be as expressive as natural language itself.

The Episodic Logic Framework

EL is built upon several novel pillars designed to handle the complexity of the real world and the language used to describe it.

Liberal Ontology

Instead of simplifying the world to fit a restricted logic, EL expands the logic to fit the world. It adopts a "liberal ontology" that includes "possible individuals," which can be real, imaginary, or even non-existent (such as "tomorrow's cancelled lecture"). This ontology encompasses various types of entities, including kinds, propositions, and regions.

The liberal ontology of Episodic Logic, illustrating the inclusion of situations, propositions, kinds, and other first-class entities.

Situations and Episodes

EL treats situations (episodes, events, states of affairs) as actual objects in the logic. A central feature is the ** operator, which connects a formula to the episode it describes. For example, the expression [Mary drop Glass1] ** e1 signifies that the formula "Mary dropped the glass" characterizes the specific episode e1.

Deindexing and Inference

The framework includes a process called deindexing, which transforms raw English-like "Logical Forms" into fully specified representations. This involves automatically filling in contextual blanks, such as determining event timing relative to speech or resolving anaphoric references (e.g., identifying which "it" refers to which event).

To perform reasoning, the authors developed the EPILOG Inference Engine. It utilizes two primary mechanisms: Rule Instantiation (RI) for forward-chaining (deriving new facts from known ones) and Goal Chaining (GC) for backward-chaining (answering questions).

Evaluation and Applications

As a foundational work from 1993, EL was evaluated on specific complex reasoning tasks rather than modern large-scale datasets:

  • Narrative Inference: The "Little Red Riding Hood" story was used to test complex causal inferences within a narrative structure.
  • Planning and Dialogue: The TRAINS domain, a classic AI planning environment, tested EL's ability to handle logistics-oriented dialogue.
  • Technical Analysis: Boeing Aircraft Reports were processed to assist in reliability and maintainability analysis, demonstrating EL's utility in parsing technical telex reports.

Key Insights

  • Ontological Promiscuity: Borrowing a term from Jerry Hobbs, EL embraces the inclusion of many entity types (kinds, propositions, regions) to avoid the pitfalls of overly restricted logics.
  • Intentionality: EL makes a sharp distinction between actions and events. An event is an occurrence, while an action is a pair consisting of an agent and an event. This allows the system to reason about intentionality—one can intend an action, but not an event where they are not the agent.
  • Probabilistic Generics: EL uses probabilistic conditionals to represent "common sense" rules that are generally but not always true, such as the probability that a dropped container will lose its contents.

Conclusion and Future Work

EL demonstrates that a logic can remain close to the surface form of English while providing the formal power needed for causal reasoning and opaque contexts. Despite its success, several challenges remain. The authors note the lack of a general theory for combining probabilistic evidence and the difficulty of preventing "inference explosion" in such an expressive system. Furthermore, while EL provides a strong mapping for syntax and logic, achieving full grammar coverage for the breadth of human language remains a significant ongoing undertaking.

Summary updated

BibTeX
@inproceedings{hwang1993el,
  author = {Chung Hee Hwang and Lenhart K. Schubert},
  title = {{EL: A Formal, Yet Natural, Comprehensive Knowledge Representation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}

Loading related content...

Figure