Computation of Modifier Scope in NP by a Language-neutral Method

Computation of modifier scope in NP by a language-neutral method Richard CAMPBELL Microsoft Research 1 Microsoft Way Redmond, WA, USA 98052 richcamp@microsoft.com Abstract The relative logical scope of multiple modifiers within NP is often semantically significant. This paper proposes a structurally based method for computing the relative scope of such modifiers, based on their order, type, and syntactic complexity. The algorithm is language-neutral, in that it works with minimal errors for a wide range of languages without language-specific stipulations. Introduction Noun phrases quite commonly have multiple modifiers, including quantifiers, attributive adjective phrases, relative clauses, possessors, appositives and the like. As frequently noted in the literature (e.g., Shaw and Hatzivassiloglou, 1999), the linear order of modifiers can signify their logical scope (though other factors are involved, too), as in the English examples (1) and (2) (bracketing indicates logical scope): (1) my [favorite [new movie]] (2) my [new [favorite movie]] In (1) favorite modifies the phrase new movie; hence the NP refers to my favorite among the new movies (there may be an old movie I like better); in (2) new modifies favorite movie; hence the NP refers to my favorite movie, which has just become my favorite. The computation of the scope of modifiers is of inherent linguistic interest: it is necessary for determining the correct interpretation of NPs like (1) and (2). It follows that it is potentially useful in any application that may depend on such an interpretation. In addition, for multilingual applications such as transfer-based machine translation (MT) (as discussed for example by Richardson et al. (2001)) modifier scope may itself be used as an abstract, language-neutral representation of their surface configuration, including linear order. The generation component of the MT application could then make use of scope information, perhaps in addition to scope-independent ordering conventions (Malouf, 2000), to generate the modifiers in the correct order. The focus of the current paper is a method for computing the relative scope of modifiers based on structural information, which works independently of any particular language; that is, the same algorithm that computes the scope of the modifiers in the English NP (3) also correctly computes modifier scope in (4), its French translation, even though the two examples do not have exactly parallel surface structures: (3) the [twenty-ninth [American state]] (4) le [vingt-neuvième [État américain]] the twenty-ninth state American The proposed algorithm considers several structural factors in addition to linear order, including the type and internal structure of the modifiers themselves. The algorithm described here is currently implemented in the NLPWin system at Microsoft Research (Heidorn, 2000). The paper is organized as follows: Section 1 examines the various structural factors that determine modifier scope, and a preliminary algorithm for modifier scope assignment is proposed; in Section 2, we compare the predictions of the algorithm to a diverse set of examples from six languages, and propose a revised algorithm; Section 3 considers some related work; and Section 4 is a conclusion. 1 Modifier scope English examples like (1) and (2) seem to …

Content truncated to 500 words