Introduction
In the field of computational linguistics, the challenge of resolving ambiguity is central to developing machines that can truly understand human language. A significant contribution to this effort was made by Mary P. Harper in 1990, who proposed a novel method for representing singular definite noun phrases—such as "the teacher" or "his mother"—within a mathematical framework. This approach allows a system to maintain a "wait and see" posture, representing potential meanings compactly and refining them only as contextual evidence becomes available.
The Challenge of Ambiguity in Definites
Natural language is inherently ambiguous. Consider the sentence: "Every boy loves his mother." For a computer to process this, it must determine the referent of the phrase "his mother." Is it a distributive reading, where each boy loves his own mother? Or is it a referential reading, where there is one specific mother previously mentioned in the discourse?
Traditional computational systems often attempted to resolve these ambiguities immediately upon encountering the phrase. However, such systems are prone to error; if a system commits to the wrong interpretation too early, it often lacks a mechanism to "backtrack" without re-processing the entire input. Harper identifies the need for an intermediate Logical Form (LF) that satisfies three criteria:
- Compactness: It should represent multiple possible meanings without enumerating every permutation.
- Modularity: It must be constructible based solely on the syntax of the current sentence.
- Consistency: It should allow for incremental refinement as more information arrives, ensuring that new data specifies meaning rather than contradicting the original structure.
The Functional Representation Approach
Harper’s core innovation lies in treating definite noun phrases not as fixed constants, but as named functions (e.g., $def_1, def_2$).
Functional Notation and Argument Selection
Instead of representing "the dog" as a static object, Harper represents it as a function of the variables surrounding it. The arguments of these functions are determined by the sentence's syntax—specifically, variables associated with noun phrases that "c-command" (syntactically dominate) the definite phrase. For instance, if a sentence contains a subject $x$ and an object $y$, a definite phrase might be initially represented as $def_n(x, y)$.
The "Upper Bound" and Incremental Updating
This initial Logical Form represents the "upper bound" of the phrase's potential meaning. It encompasses every variable that could theoretically influence the referent. As the system processes subsequent sentences or integrates general world knowledge, it performs argument reduction. This process simplifies the function or equates it to a specific value, narrowing the scope of the "upper bound" until the ambiguity is resolved. This mirrors human cognition, where we often maintain a "fuzzy" understanding that sharpens as a conversation progresses.
Linguistic Testing and Results
To validate this model, Harper utilized a suite of 14 complex linguistic examples designed to test edge cases in natural language processing:
- Verb Phrase (VP) Ellipsis: In sentences like "Fred loves his wife; George does too," the system must account for both "strict" readings (George loves Fred’s wife) and "sloppy" readings (George loves his own wife).
- Bach-Peters Sentences: These involve complex cross-references, such as "The pilot who shot at it hit the MiG that chased him."
- Relative Clause Islands: Constraints where certain words cannot refer to entities outside their immediate sub-clause.
Harper demonstrated that the functional approach successfully handles these phenomena where previous models—including those based on Bertrand Russell’s theory of descriptions—often failed. Crucially, the approach was fully implemented, proving that the functional representation of definites is computationally viable and not merely a theoretical construct.
Key Insights
One of the most significant insights of the paper is the failure of constants. Harper points out that treating definites as fixed individuals fails in the presence of universal quantifiers (e.g., "Every boy..."). By using a functional approach, the definite can change its value depending on the variable it is bound to. Furthermore, the model ensures that the internal structure of a phrase (e.g., the modifiers in "the old lady's diary") remains part of the function's "restriction," keeping the meaning linguistically grounded.
Limitations and Future Directions
While the functional LF is robust, Harper acknowledges certain limitations:
- Structural Ambiguity: The system currently cannot provide a single LF for sentences with multiple syntactic interpretations (e.g., "I saw the man with the telescope"). Harper suggests that future iterations might utilize "parse forests" to store multiple partial LFs until the syntax is clarified.
- Pragmatics: The LF provides the framework for resolution, but the actual "common sense" reasoning required to select the correct antecedent remains the domain of a separate pragmatic module.
By providing a consistent, modular, and compact way to represent definite noun phrases, Harper’s work remains a foundational step in the development of sophisticated natural language understanding systems.
Summary updated