A Common-Sense Conceptual Categorization System Integrating Heterogeneous Proxytypes and the Dual Process of Reasoning

Antonio Lieto, Daniele Paolo Radicioni, Valentina RhoIJCAI 2015

Introduction

Traditional artificial intelligence systems often struggle with "common-sense" categorization. While classical logic defines concepts through necessary and sufficient conditions (e.g., a "bird" must have feathers and lay eggs), human cognition relies on more flexible structures. Psychological research suggests that humans utilize prototypes (typical instances like a robin for "bird") and exemplars (specific instances like a pet parrot) to navigate the world.

Presented at IJCAI 2015, DUAL-PECCS is a hybrid knowledge representation system designed to bridge this gap. By integrating geometrical representations with formal ontologies, the system mimics the human ability to categorize objects using a combination of fast, similarity-based intuition and slow, logical verification.

The DUAL-PECCS Architecture

The core of DUAL-PECCS lies in its ability to handle heterogeneous "proxytypes"—temporary mental constructs that represent concepts in different ways depending on the context. The system is built upon the Dual Process Theory of mind, which posits two distinct reasoning systems:

  1. System 1 (S1) - Conceptual Spaces: This layer uses a geometrical framework where concepts are regions in a multi-dimensional space. Similarity is determined by the distance between points. Within this space, prototypes are represented as centroids of regions, while exemplars are specific points.
  2. System 2 (S2) - Ontologies: This formal symbolic layer (utilizing OpenCyc) performs deductive reasoning. It ensures that any categorization proposed by S1 is logically consistent with established taxonomic rules.
Figure 1: The DUAL-PECCS hybrid knowledge base integrates typicality-based knowledge (prototypes and exemplars) with classical ontological information.

The interaction between these systems follows a Dual-Process Loop: S1 first generates a fast, similarity-based hypothesis. S2 then acts as a filter, rejecting proposals that violate "hard" logical constraints (e.g., rejecting "fish" for a swimming creature if it possesses mammalian traits). This architecture was integrated into the ACT-R cognitive architecture, extending its declarative memory capabilities.

Experimental Evaluation

The system was evaluated using a dataset of 90 linguistic descriptions, structured as simple riddles. These were divided into 45 prototype-targeted descriptions (e.g., "The big carnivore with yellow fur and black stripes") and 45 exemplar-targeted descriptions (e.g., "The big carnivore with white fur and black stripes"). The "gold standard" for these targets was validated against human subject data from prior psychological experiments.

Testing was conducted in two modes:

  • With no IE (Information Extraction): Using manually encoded, "clean" input.
  • With IE: Using an automated natural language processing pipeline to parse the descriptions.

Results and Analysis

The system demonstrated robust performance across both input types. The primary metrics were Concept Categorization Accuracy (CC-ACC) and Proxyfication Accuracy (P-ACC), the latter measuring the system's ability to correctly distinguish between a general prototype and a specific exemplar.

Figure 2: Categorization and proxyfication accuracy results for DUAL-PECCS with and without automated Information Extraction (IE).

As shown in the results, DUAL-PECCS achieved a high CC-ACC of 95.6% with clean input, which remained strong at 85.6% even when subjected to the noise of automated NLP. Proxyfication accuracy followed a similar trend. Error analysis revealed a specific cognitive bias: "Ex-Proto" errors occurred when specific features of an exemplar "outshone" the general features of a prototype during similarity calculations, leading the system to return a specific instance when a general category was expected.

Conclusion and Future Directions

DUAL-PECCS successfully demonstrates that integrating multiple representational layers—prototypes, exemplars, and ontologies—significantly enhances an AI's ability to handle common-sense reasoning. By using System 1 as a fast filter for the more computationally expensive System 2, the architecture achieves both efficiency and logical rigor.

Future work aims to address current limitations, such as the relatively shallow language processing pipeline and the limited scope of features. The authors plan to incorporate more complex language models and expand the representational scope to include affordances (functional uses of objects) and sensory-motor data, further closing the gap between artificial and human-like categorization.

Summary updated

BibTeX
@inproceedings{lieto2015common,
  author = {Antonio Lieto and Daniele Paolo Radicioni and Valentina Rho},
  title = {{A Common-Sense Conceptual Categorization System Integrating Heterogeneous Proxytypes and the Dual Process of Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}

Loading related content...

Figure