# Tech Report CS-89-46

## The Representation of Noun Phrases In Logical Form

### Abstract:

Several researchers in artificial intelligence have recognized the usefulness of a two-stage model of sentence comprehension for building a computer model of language. In the first stage, an intermediate level of representation called logical form is derived. During the second stage, logical form is updated with additional information (e.g., quantifier scoping). We introduce three constraints we consider necessary to make this model of language computationally feasible: \begin{enumerate} \item Logical form should compactly represent ambiguity. \item Logical form should be initially computable from syntax and local (sentence-level) semantics. In particular, logical form should not be dependent on pragmatics, which requires inference and hence internal representation. \item Further processing of logical form should only disambiguate or further specify logical form. Logical form has a meaning. Any further processing must respect that meaning. \end{enumerate}

Within this framework, we have devised logical-form representations for pronouns, singular definite noun phrases, and singular indefinite noun phrases. For example, we represent a pronoun as a function of all of the variables corresponding to operators that can bind the pronoun. This representation allows us to indicate a meaning for the pronoun without deciding on the antecedent for the pronoun. Later, when we can determine the antecedent for the pronoun, we replace the pronoun function with the variable or function used to represent its antecedent. Like pronouns, definites are represented as functions. However, indefinites cannot initially be represented as a function in logical form. Initially, we represent an indefinite as an existentially quantified variable. Later, when more information is available about the meaning of a noun phrase, the initial representation is limited to indicate the intended meaning of that noun phrase.

We demonstrate that these representations both model the appropriate linguistic behavior and satisfy our computational constraints. This work has been implemented and tested on a wide variety of examples.

(complete text in pdf)