The Knowledge Required to Interpret Noun Compounds

James Fan, Ken Barker, Bruce W. PorterIJCAI 2003

Introduction

Noun compounds, such as "concrete floor" or "animal virus," are ubiquitous in natural language but present a significant challenge for automated interpretation. The relationship between the constituent nouns is often implicit: a "concrete floor" is a floor made of concrete, whereas a "gymnasium floor" is a floor located in a gymnasium. To reason effectively, an AI must perform "loose-speak interpretation," translating these phrases into formal logic.

A central hurdle in this field is the Knowledge Acquisition Bottleneck. Traditional views suggest that an AI might require exhaustive, granular knowledge about every specific noun to understand a compound. However, this paper investigates a different hypothesis: that a robust high-level ontology—the "top levels" of a knowledge base—provides the essential logic required to achieve high accuracy, potentially bypassing the need for exhaustive detail.

The researchers treat the interpretation of noun compounds as a pathfinding problem within a Knowledge Base (KB) conceptual graph. The process involves three primary steps:

  1. Mapping: The nouns (e.g., "animal" and "virus") are mapped to specific nodes within the KB.
  2. Search: A breadth-first search (BFS) explores semantic relations—such as agent, object, or has-part—to identify a path connecting the nodes.
  3. Pathfinding: Unlike systems that assign a single category (e.g., "location"), this approach identifies a sequence of relations. For instance, "animal virus" might be resolved as: virus → agent of → invade → object is → cell part of → animal.

To determine which types of knowledge are most vital, the authors conducted ablation studies. They systematically removed levels of the KB hierarchy, starting from the most general concepts (e.g., Entity vs. Event) down to specific leaf nodes, measuring the resulting degradation in performance.

Experimental Setup

The study evaluated the system across three distinct domains to ensure generalizability:

  • Cell Biology: 224 compounds sourced from a college-level textbook (Alberts et al.).
  • Small Engine Repair: 294 compounds from a repair manual.
  • Computer Hardware: 224 compounds from a technical manual.

The researchers utilized the "Component Library" for the biology domain and an augmented version of WordNet for the others, focusing on the structural relationships between concepts.

Results and Key Insights

The experiments yielded several critical insights into the structure of semantic knowledge and its role in language processing:

The "Top-Heavy" Importance

The most vital knowledge resides at the highest levels of the ontology. The authors observed a "Level 1 Cliff": deleting the most general distinctions (such as the difference between an Entity and an Event) caused a catastrophic drop in both precision and recall. High-level distinctions act as necessary "guardrails" for the search algorithm; without them, the system cannot distinguish between physical objects and processes, leading to a flood of erroneous interpretations.

Axioms as Connective Tissue

Top-level levels contain universal axioms, such as "every Action involves an Object." These axioms serve as the essential "steps" in the search path. Their removal often results in the system failing to find any connection at all, significantly impacting recall. In contrast, ablating Level 2 created a significant gap where the system became much more likely to find no interpretation at all rather than an incorrect one.

Diminishing Returns on Detail

One of the most surprising findings was that granular, low-level knowledge about specific nouns is relatively less important. Once the high-level structural framework is established, adding specific details provides only marginal improvements in interpretation accuracy. With a full KB, the system achieved approximately 80% precision and recall across all domains, but performance remained remarkably stable even as lower leaf nodes were removed.

Conclusion and Future Work

This research suggests that the path toward solving the Knowledge Acquisition Bottleneck may lie in perfecting "upper ontologies" rather than attempting to hand-code every specific fact about the world. While the current study focused on noun pairs and required manual mapping of nouns to KB concepts, the results provide a strong foundation for future work in general text understanding. By focusing on the high-level logic that governs semantic relationships, AI systems can achieve robust interpretation with a surprisingly lean set of core principles.

Summary updated

BibTeX
@inproceedings{fan2003knowledge,
  author = {James Fan and Ken Barker and Bruce W. Porter},
  title = {{The Knowledge Required to Interpret Noun Compounds}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2003},
}
Figure