Structural Relations - A Case Against Case

Ingeborg Steinacker, Harald TrostIJCAI 1983

Introduction

In Natural Language Understanding (NLU), "relational words" present a significant challenge for semantic parsing. Words such as the verb to have, to get, or possessive pronouns like my do not possess a fixed, inherent meaning. Instead, their semantic value is derived dynamically from the concepts they connect. Traditionally, systems attempted to resolve this ambiguity by maintaining exhaustive dictionaries of every possible sense. This paper argues that such an approach is fundamentally flawed and proposes a structural alternative: the Connectivity Condition.

The Challenge of Relational Words

Relational words are highly flexible, adapting to the context provided by the nouns they link. For instance, the phrase "my car" implies ownership, whereas "my sister" denotes a social relationship, and "my headache" refers to a physiological state.

Examples of the diverse semantic roles fulfilled by the relational verb "have".

As shown in the table above, the verb to have can represent ownership, social ties, sources of information, health states, or even serve as a paraphrase for an action. Attempting to hard-code these variations into a lexicon is inefficient; every time a new concept is added to the system, the dictionary entries for all relational words must be manually updated to account for new potential pairings.

The Connectivity Condition

To address this, the authors introduce a procedure based on the Connectivity Condition (CC) within Si-Nets (Structural Inheritance Networks). The core philosophy is to leverage the existing architecture of a knowledge base to infer relationships automatically.

The system distinguishes between the structure of the network and its content. When a parser encounters a relational word connecting two concepts, it bypasses the word list and examines the structural level of the Si-Net.

A basic structural representation of a role connecting two concepts in a semantic network.

The disambiguation process follows three primary paths:

  1. Original Sense: Used when a word has a primary, specific meaning (e.g., to give as a physical transfer).
  2. Event Paraphrasing: Used when the word facilitates an action description (e.g., "to give a kiss" as a synonym for "to kiss").
  3. Structural Connectivity: If the two concepts are already linked by a role in the network—for example, a PERSON filling a *PASSENGER role for an AUTOMOBILE—the system identifies that pre-existing role as the intended meaning.

Key Insights and Scaling

By treating structure as meaning, the system becomes inherently "future-proof." When new concepts are added to the semantic net, relational words automatically adapt to them based on the new structural links, requiring no additional programming.

Furthermore, the Connectivity Condition provides a robust framework for processing metaphors. In a sentence like "My car drinks gasoline," the system can identify a *FUEL role connecting AUTOMOBILE and GASOLINE, allowing it to resolve the metaphorical usage of "drinks" without needing a specific "fuel-consumption" sense for the verb.

Results and Limitations

The implementation of this approach in the V1E-LANG system (a German sentence parser) demonstrated a "drastic" reduction in lexical complexity. A single general rule for the Connectivity Condition successfully replaced dozens of hard-coded word senses across verbs, pronouns, and reference resolutions.

However, the authors note that structural connectivity is not a universal solution. Ambiguity persists in cases like "my present," which could refer to a gift being given or received. In such instances, the Connectivity Condition may identify multiple valid paths, necessitating higher-level situational context to achieve final disambiguation. Future research aims to integrate this structural method with broader context-aware processing.

Summary updated

BibTeX
@inproceedings{steinacker1983structural,
  author = {Ingeborg Steinacker and Harald Trost},
  title = {{Structural Relations - A Case Against Case}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure