Representing Pronouns in Logical Form: Computational Constraints and Linguistic Evidence

Mary P. HarperAAAI 1988

Introduction

In computational linguistics, the transition from a sentence's surface syntax to its deep semantic meaning is mediated by a "Logical Form" (LF). One of the most persistent challenges in constructing a robust LF is the treatment of pronouns, particularly in the context of Verb Phrase (VP) ellipsis. A classic example is the sequence: "Fred loves his wife. George does too." This second sentence is inherently ambiguous, allowing for either a strict identity interpretation (George loves Fred’s wife) or a sloppy identity interpretation (George loves his own wife).

Historically, models of pronoun resolution struggled to balance mathematical soundness with computational efficiency. This paper proposes a novel framework that represents pronouns as functions rather than static variables or strings, satisfying key computational constraints while expanding linguistic coverage.

Computational Constraints for Logical Form

The paper argues that any viable Logical Form must adhere to three fundamental principles to be effective in a computational system:

  1. Compactness: The LF should represent ambiguity within a single structure. Generating $2^n$ distinct versions of a sentence to account for every possible interpretation is computationally prohibitive.
  2. Local Computability: The LF must be derivable from syntax and local semantics. It should not require immediate access to high-level pragmatics or world knowledge to establish its initial structure.
  3. Monotonicity: The processing of a sentence should be an additive process. As more information becomes available (e.g., through pronoun resolution), the system should add information or disambiguate existing structures rather than deleting or "destructively" modifying the initial LF.

The Functional Approach to Pronouns

To satisfy these constraints, the author introduces a functional representation for pronouns. Instead of treating a pronoun as a fixed label, it is treated as a function whose arguments are the variables currently in scope.

For example, in the sentence "Fred loves himself," the pronoun himself is represented as a function $himself_1(x)$, where $x$ represents the subject. The resolution of the pronoun does not involve replacing the function; instead, the system adds an equality statement: $(himself_1(x) = x)$.

This approach allows the system to maintain a stable underlying structure. In the case of VP ellipsis, the ambiguity between strict and sloppy identity is managed by a disjunction of equalities. This allows the system to defer the final choice between interpretations until more contextual information is available, without needing to reconstruct the entire logical representation.

Linguistic Evidence and Evaluation

The model was evaluated against a suite of 23 linguistic "challenge cases." These cases involve complex structures that typically break simpler models of pronoun resolution, such as nested possessives (e.g., "Every boy showed his mother her clock").

In such cases, the pronoun "her" refers to "his mother," but "his mother" is not a fixed entity—it is a dependent entity that changes based on which "boy" is being discussed. The functional approach handles this naturally by nesting functions: the representation of "her" becomes a function of the representation of "his mother," which in turn is a function of the subject variable. This recursive capability allows the model to handle deep dependencies that previous string-replacement models (such as those by Sag or Webber) could not resolve efficiently.

Comparison with Prior Models

The proposed model offers significant advantages over the then-standard approaches of the late 1980s:

  • Efficiency: By avoiding the generation of multiple separate logical representations for ambiguous sentences, the functional model reduces the search space for the computer.
  • Stability: Unlike models that rely on "destructive" updates—where parts of the logic are swapped out as context is added—this model follows the principle of monotonicity, making it easier for a system to backtrack or refine its understanding without losing the original structural data.
  • Coverage: The model successfully addresses Example 16 and similar complex cases where pronouns refer to possessive noun phrases that are themselves dependent on other variables.

Conclusion

By reimagining pronouns as "chameleons" that adapt their behavior through functional arguments, this work provides a bridge between rigid logical structures and the fluid nature of natural language. While the paper focuses on the theoretical framework and the representation of pronouns, it lays the groundwork for more sophisticated compositional rules that can automatically transform raw syntax into computationally efficient and linguistically accurate Logical Forms.

Summary updated

BibTeX
@inproceedings{harper1988representing,
  author = {Mary P. Harper},
  title = {{Representing Pronouns in Logical Form: Computational Constraints and Linguistic Evidence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure