The Semantic Interpretation of Nominal Compounds

Timothy W. FininAAAI 1980

Introduction

In English, we frequently combine nouns to create new meanings, such as "aircraft flight arrival" or "engine repairs." For humans, these nominal compounds are intuitive; however, for computational systems, they present a significant challenge. The core difficulty lies in the fact that the semantic relationship between the nouns is implicit. Without prepositions or verbs to guide interpretation, a system must rely on a structured world model to understand the roles different objects play. This 1980 paper proposes a frame-based system that utilizes semantic rules and a "role-fitting" algorithm to automatically discover these hidden relationships.

Methodology

The system, implemented as part of the JETS natural language program, relies on three primary components to decode noun-noun sequences.

1. Frame-Based Representation

Concepts are organized in a hierarchy using a system derived from Frame Representation Language (FRL). Each "frame" contains slots, or roles, that can be filled by other concepts. This structure allows the system to define a "pilot" not just as a person, but as the agent of a flying event.

2. Three Tiers of Rules

The interpretation process is governed by three levels of semantic rules:

  • Idiomatic Rules: These handle "canned" phrases with fixed, non-compositional meanings (e.g., "hanger queen").
  • Productive Rules: These capture common semantic patterns. For instance, a rule might specify that a "ChemicalCompound + Liquid" usually implies the compound is "dissolved in" the liquid.
  • Structural Rules: These are the most versatile, examining the slots available in a concept's frame. A RoleValue + Concept rule attempts to see if the first noun can fill a functional slot in the second. For example, in "washing machine," the machine fills the instrument slot of the washing event.
Examples of nominal compounds mapped to their underlying semantic frames.

3. The Role-Fitting Algorithm

This "matching engine" scores how well a modifier fits into the roles of a head noun. The algorithm evaluates various "facets" of a role to determine compatibility.

The facets used to evaluate and score the fit between a noun and a semantic role.

The system uses a weighted scoring function rather than a binary match. If a modifier satisfies a "Preference," it receives a positive score; if it violates a "Requirement," the interpretation is discarded (negative infinity). This allows the system to rank multiple possible meanings and resolve ambiguity.

Key Insights: Role Nominals and Directionality

A significant contribution of the paper is the treatment of role nominals—nouns that inherently refer to a role in another event. By defining a "pilot" through the "to-fly" concept, the system can interpret "pilot school" by linking the underlying actions of teaching and flying.

Additionally, the paper identifies a fundamental directionality in interpretation:

  • Modifier fills a role in the head: e.g., "engine repair" (engine is the object of repair).
  • Head fills a role in the modifier: e.g., "service man" (man is the agent of service).

Results and Evaluation

The system was evaluated using examples from the JETS/PLANES query system, a natural language interface for Navy aircraft maintenance records. Qualitative results demonstrated the system's ability to handle complex, nested compounds:

  • "F4 pilot school": Correctly identified as an organization teaching people to fly F4 aircraft.
  • "Aluminum water pump": Distinguished "aluminum" as the material and "water" as the substance being moved.
  • Ambiguity Resolution: For "Chicago flights," the system could identify and rank interpretations (to, from, or stopping in Chicago) based on the database context.

Limitations

While pioneering, the system faced several hurdles. The author noted that "correct" interpretations often depend on pragmatic context beyond the immediate sentence. Furthermore, the boundary between idiomatic and structural rules is often blurry, requiring manual tuning. Finally, as compounds grow in length, the search space for possible structures increases exponentially, necessitating efficient pruning strategies.

Summary updated

BibTeX
@inproceedings{finin1980semantic,
  author = {Timothy W. Finin},
  title = {{The Semantic Interpretation of Nominal Compounds}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure