Introduction
When humans encounter novel compound phrases like "mountain chair" or "robot dog," they instinctively perform "concept combination." These interpretations generally fall into two categories: relational (e.g., a chair located on a mountain) or property-based (e.g., a dog made of metal). Cognitive science suggests that "diagnosticity"—the degree to which a feature uniquely identifies a concept—is the primary driver of these interpretations. While identifying diagnostic properties (like "yellow" for "banana") is intuitive, mathematically formalizing diagnosticity for abstract relations has remained a challenge. This research presents a formal framework that quantifies diagnosticity using Information Retrieval (IR) techniques and implements it within a Constraint Logic Programming (CLP) system.
Quantifying Diagnosticity via IR
The core innovation of this work is the adaptation of the $tf-idf$ (term frequency-inverse document frequency) weighting schema to cognitive features. By treating a concept as a "document" and its associated properties or relations as "terms," the authors provide a uniform measurement for both physical traits and abstract connections.
The diagnosticity $D$ of a feature $F$ for a concept $C$ is calculated as:
$$D_{C,F} = ff_{C,F} \times \log_2(N_{CS}/n_F)$$
where $ff$ is the feature frequency, $N_{CS}$ is the total number of concepts, and $n_F$ is the number of concepts possessing that feature. This formula effectively penalizes common, non-distinctive features (e.g., "has a weight") while highlighting those that are statistically rare and thus highly diagnostic.
System Architecture: The CLP Engine
The reasoning engine, implemented in Prolog, utilizes two primary algorithms to generate interpretations for noun-noun combinations:
- Relational Algorithm: Identifies highly diagnostic relations that link the Modifier Noun (MN) and the Head Noun (HN).
- Property Algorithm: Focuses on the transfer of highly diagnostic properties from the MN to the HN.
The system employs a "Highly Diagnostic" (HD) threshold, flagging features that score at least one standard deviation above the mean. This constraint mimics human cognition by prioritizing the most defining characteristics during the interpretation process.
Evaluation and Results
The system was evaluated using a knowledge base derived from descriptions provided by 30 human participants. It was then tested on 30 novel noun-noun combinations (15 property-based, 15 relational).
The results demonstrate significant alignment with human reasoning:
- Sensibility: 81% of the 62 generated interpretations were rated as "sensible" by human judges.
- Human Alignment: 50% of the system-generated meanings were identical to those produced by human subjects in a separate empirical study.
- Property Effectiveness: The system was particularly adept at identifying the specific property-based interpretations most frequently favored by humans.
Conclusion and Future Directions
This work demonstrates that cognitive importance can be modeled as statistical rarity. By providing a uniform mathematical treatment for properties and relations, the system bridges a long-standing gap in concept combination research.
However, the current system is limited by its reliance on a manually curated, "self-contained" knowledge base. Future research will focus on scaling this $tf-idf$ approach to massive, standardized ontologies like WordNet and refining the CLP constraints to better filter out interpretations that are plausible but lack "world knowledge" or utility.
Summary updated