Two Theories for Computing the Logical Form of Mass Expressions

Francis Jeffry Pelletier, Lenhart K. SchubertACL 1984

TL;DR

This paper explores how to translate English sentences containing mass nouns (like "water" or "furniture") into logical representations that capture their meaning. The authors compare two distinct theories that successfully handle these tricky expressions, offering a more robust approach to natural language understanding than previous methods.

What Problem Does This Paper Address?

Understanding the meaning of sentences in natural language is a core goal in computational linguistics and artificial intelligence. A significant challenge arises from "mass expressions" – words that refer to substances or concepts that aren't easily countable, like "water," "furniture," or "information." Unlike count nouns ("chair," "book"), mass nouns don't typically take plural forms or determiners like "a" or "an" in their standard usage.

Traditional approaches to translating English into logical forms often struggle with this mass/count distinction. Key questions include:

  • Is the mass/count distinction a purely syntactic feature of words, or does it reflect their semantic meaning?
  • Should this distinction be applied to individual words in the lexicon, or determined by how they are used in a sentence?

The authors aim to present and compare two theories that can produce "plausible inferences and intuitively satisfying semantic representations" for sentences involving mass expressions, going beyond simpler, rule-based systems.

Methods

The paper focuses on two main theoretical frameworks for handling mass expressions:

  1. Syntactic Expression Approach (s-theory): This theory treats the mass/count distinction as a lexical feature. Nouns are classified as either mass or count in the lexicon. Special "lexical extension" rules are used to adjust the meaning of nouns based on their context. For example, a count noun like "man" might be extended to refer to the "stuff of which men are made" when used in a mass context. Syntactic rules use these features to guide the translation process.
  2. Semantic Occurrence Approach (p-theory): This theory argues that the mass/count distinction is not a syntactic feature but rather a description of the semantic representation. The core idea is that "reality" contains two types of entities: "ordinary objects" (like puddles or rings) and "kinds" (like wine or red wine). Mass predicates are true of both kinds and quantities of matter. The p-theory uses semantic operators (like $\lambda$-abstraction) to derive the appropriate meaning based on the context of occurrence.

Both theories rely on a formal semilattice structure to represent "kinds" of substances and their relationships. They also utilize meaning postulates to define the relationships between predicates and the entities they apply to. The paper presents a set of grammar rules (phrase structure rules paired with semantic translation rules) that are designed to produce "conventional" logical translations.

Datasets

Not discussed in the paper. The paper focuses on theoretical frameworks and linguistic examples rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • The mass/count distinction is complex: It's not simply a matter of whether a noun is pluralizable. The distinction impacts how we understand quantities, substances, and abstract types.
  • Two viable theoretical paths: The paper highlights that both a purely syntactic approach (s-theory) and a more semantically driven approach (p-theory) can successfully account for the nuances of mass expressions.
  • Ontological commitment: Both theories posit an ontology with "ordinary objects" and "kinds," which is crucial for representing the meaning of mass terms.
  • "Formal" semilattices: The concept of a semilattice of "kinds" allows for a systematic way to represent relationships between different types of substances (e.g., "wine" and "red wine").
  • Comprehensive denotation: The proposed theories can account for the validity of arguments involving mass terms that other theories struggle with, such as "Everything edible is food" and "Rice is food." This suggests a more powerful way of representing meaning.

Key Results

The paper demonstrates that both the syntactic expression approach (s-theory) and the semantic occurrence approach (p-theory) can:

  • Produce correct semantic representations for a wide range of sentences involving mass terms.
  • Generate plausible inferences, such as "Water is dripping from the faucet" entailing "sm water is dripping from the faucet."
  • Handle sentences like "Tap water is water" and "All wine is wine" as analytic (true by definition).
  • Make certain inferences valid (e.g., relating "Claret is a wine" and "wine is a liquid" to "claret is liquid") and others invalid.

The paper doesn't present quantitative results in terms of accuracy or performance metrics, as it's a theoretical comparison.

Limitations and Future Work

  • Theoretical comparison: The paper primarily focuses on comparing the theoretical underpinnings and linguistic coverage of the two theories, rather than providing empirical validation.
  • Meaning Postulates: The success of both theories relies on having appropriate "meaning postulates," which are not fully detailed for all cases.
  • Preference for one theory: While both theories are presented as successful, the paper concludes that the question of which theory is to be preferred is a topic for further research.
  • Future research: The authors suggest that future work should involve systematic attempts to account for large classes of mass expressions within formal syntactic-semantic-pragmatic frameworks to resolve remaining issues. They also note that extending inference mechanisms to allow for predicate modification and $\lambda$-abstraction is necessary to fully implement these theories.

Summary updated

BibTeX
@inproceedings{pelletier1984two,
  author = {Francis Jeffry Pelletier and Lenhart K. Schubert},
  title = {{Two Theories for Computing the Logical Form of Mass Expressions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure