Common Sense Retrieval

A. Julian CraddockAAAI 1992

Abstract

Traditional artificial intelligence systems in the early 1990s largely relied on rigid, declarative rule-based frameworks to simulate common-sense reasoning. This paper proposes a paradigm shift toward "Common Sense Retrieval," where reasoning is treated as a statistical inference problem over an Episodic Knowledge Base (EKB). By storing specific past experiences and applying mechanisms for generalization and associative chaining, the system can derive probabilistic conclusions about new situations. This approach addresses the brittleness of universal rules and provides a principled method for handling incomplete or atypical information.

Introduction

In the early 1990s, the dominant approach to common-sense reasoning involved constructing massive corpora of general declarative rules (e.g., "All birds fly"). However, Julian Craddock argues that such systems are inherently brittle and inefficient, particularly when encountering "new" or slightly atypical experiences that do not neatly fit into pre-defined categories.

The central challenge addressed in this work is how a system can utilize specific, partial descriptions of past events to reason about the unobserved properties of a current event. Rather than relying on a universal rule, Craddock suggests that common sense is better modeled as a retrieval process: looking back at a "memory" of similar instances to determine the most likely outcome based on historical frequency.

The Framework: EKB and Probabilistic Logic

The proposed architecture utilizes a first-order logic framework consisting of two distinct dialects:

  1. Language $L$: This is used to record specific experiences within the Episodic Knowledge Base (EKB). For example, an entry might state: "Observation 34: a red object that is a car."
  2. Language $QL$: An extension of $L$ used for querying the EKB. It introduces a probability term, $Prob(\alpha | \beta)_{ekb}$, which represents the conditional probability of property $\alpha$ being true given property $\beta$, derived directly from the data stored in the EKB.

The Retrieval Mechanism

The core reasoning process follows a four-step pipeline designed to move from a specific query to a statistical inference:

  • Input: The system receives a partial description of a new experience.
  • Select: It identifies a reference class, which is the subset of past experiences in the EKB relevant to the current situation.
  • Modify: If no direct matches exist in the EKB, the system employs two strategies: Generalization (relaxing constraints, such as looking for "any bird" instead of a specific "red bird") and Chaining (using associative links to connect related concepts).
  • Aggregate: Finally, the system uses statistical techniques to count occurrences within the reference class and outputs a conditional probability.

Craddock’s framework introduces several pivotal ideas for the field of automated reasoning:

The Reference Class Problem

The paper frames common-sense reasoning as a statistical problem. The primary difficulty in reasoning is often not the logic itself, but choosing the correct "reference class"—the most appropriate set of past examples—to serve as a basis for comparison.

To handle the modification step, the system organizes possible generalizations into a partial lattice. This allows the system to systematically "back off" from a specific, failed query to a more general one until it finds a sufficient sample size in the EKB to make a statistically significant inference.

Direct Inference

Unlike many machine learning models that pre-process data into abstract concepts during a training phase, this model performs "direct inference." It reasons using the raw, partial descriptions of experiences at the time of the query, allowing for greater flexibility as the EKB grows.

Formal Results and Qualitative Analysis

As a theoretical and architectural proposal, the paper focuses on formalizing the logic of retrieval. Key outcomes include:

  • Formalization of $QL$: The author successfully defines a logic where probabilities are grounded in the cardinality (count) of sets of experiences within the EKB.
  • Handling Incomplete Knowledge: Through Generalization and Chaining, the paper provides a principled way to provide answers even when the system has never encountered an identical situation before.
  • Computational Efficiency: Craddock argues that by reducing reasoning to pattern matching and counting, the system can potentially outperform complex deductive engines.

Limitations and Future Directions

While the framework is robust, it faces practical hurdles. The model is inherently "data hungry," requiring a large and representative set of experiences to function effectively—a significant challenge in 1992. Additionally, searching the generalization lattice and finding optimal associative chains can be computationally expensive. The author notes that while the retrieval itself is fast, the heuristics for modifying the reference class remain a primary area for ongoing research, a topic explored further in his subsequent doctoral work.

Summary updated

BibTeX
@inproceedings{craddock1992common,
  author = {A. Julian Craddock},
  title = {{Common Sense Retrieval}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure