File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/92/j92-4002_metho.xml
Size: 98,849 bytes
Last Modified: 2025-10-06 14:13:11
<?xml version="1.0" standalone="yes"?> <Paper uid="J92-4002"> <Title>Ambiguous Noun Phrases in Logical Form</Title> <Section position="5" start_page="422" end_page="425" type="metho"> <SectionTitle> 2 The only exception are pronouns like those in paycheck sentences (first noticed by Kartunnen \[1969\]). </SectionTitle> <Paragraph position="0"> Consider the sentence Fred gave his paycheck to his wife. George gave it to his mistress. The pronoun it is not referential. For that matter, it is not bound. The pronoun seems to take his paycheck as its antecedent where the pronoun is instantiated to a different individual than in the original sentence. 3 This is related to the Complex NP Constraint introduced by Ross (1967), which prevents wh-movement out of a relative clause attached to an NP.</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 18, Number 4 Donkey sentences provide evidence that all of the following cannot be simultaneously true (adapted from Heim \[1982, p. 102\]).</Paragraph> <Paragraph position="2"> 1. Indefinites should be represented using existential quantifiers.</Paragraph> <Paragraph position="3"> 2. Indefinites obey the same scope-island restriction as universals.</Paragraph> <Paragraph position="4"> 3. Pronouns are either bound variables or referential.</Paragraph> <Paragraph position="5"> Many researchers have attacked one or more of these assumptions, but we prefer to modify the third by adding an additional type of pronoun: pronouns that adopt the functional behavior of their antecedents. The first two assumptions, together with our modification of the third, allows us to build a simpler LF for a sentence and also handle donkey sentences (as described in Section 5.2).</Paragraph> <Paragraph position="6"> Pronouns are a source of ambiguity in verb phrase ellipsis (VPE). To signal a VPE, a full verb phrase (VP) is replaced with an auxiliary, as in the second sentence of Example 3. A sentence with VPE is called an elided sentence. The index on Fred and his indicates that they are co-referential.</Paragraph> <Paragraph position="7"> The elided sentence has little meaning independent of the first sentence, called a trigger sentence. Hence, before determining the meaning of the elided sentence, the meaning of the trigger sentence must be completely determined. Even though the antecedent of the pronoun in the trigger sentence is Fred, the meaning of that pronoun is still ambiguous because pronouns in a trigger VP can refer to a subject NP either directly or indirectly. This example indicates that care is needed to design a pronoun representation capable of handling VPE since the representation must be compatible with the two behaviors of a pronoun whose antecedent is the syntactic subject of a trigger sentence. It also demonstrates how the meaning assigned to the trigger VP limits the possible meanings of the elided sentence. Though the meaning of the elided sentence is ambiguous, it cannot mean George loves some other person's wife (other than Fred's or George's).</Paragraph> <Section position="1" start_page="423" end_page="425" type="sub_section"> <SectionTitle> 3.2 Pronouns: The Initial Representation </SectionTitle> <Paragraph position="0"> Before introducing our LF for pronouns, we briefly describe the LF for the rest of a sentence. A sentence is represented as a predicate-argument structure, with subjects lambda abstracted to handle VPE (following Sag \[1976\]; Williams \[1977\]; Webber \[1978\]; and Partee and Bach \[1981\]). By lambda abstracting syntactic subjects in LF, a pronoun whose antecedent is a syntactic subject can refer to that subject in two different ways, either directly by using a value depending on the type of the subject NP or indirectly by using the subject's lambda variable. The logical roles of all NPs in a sentence are indicated by position in LF (logical subject first, logical object second, logical indirect object third, etc.). Following Webber \[1978\], we represent universal NPs as universally quantified and restricted variables (as in 4a) and existentially quantified NPs as existentially quantified and restricted variables (as in 4b). The colon between the quantifier and its restriction expands differently depending on the type of the quantifier.</Paragraph> <Paragraph position="1"> Mary P. Harper Ambiguous Noun Phrases in Logical Form Example 4 a. Sentence: Every man is happy.</Paragraph> <Paragraph position="2"> Representation: Vx: (man x) (happy x) Meaning: Vx (if (man x) (happy x)) b. Sentence: A man is happy.</Paragraph> <Paragraph position="3"> Representation: 3x: (man x) (happy x) Meaning: 3x (and (man x) (happy x)) Quantifier scoping is handled in the same way as in Section 2. Initially, quantifiers are placed in the predicate-argument structure for the sentence, except for subjects, which are necessarily abstracted. Abstraction of a quantified subject does not imply that it must have scope over quantifiers placed in the lambda function corresponding to the VP. Later, when information becomes available for making scoping decisions, quantifier scoping is indicated using a method similar to Allen's (1987) (described in Section 5.2). Possessive NPs are represented as functions of the possessive nouns (following Webber \[1978\]) and proper nouns as skolem constants (i.e., skolem functions without arguments). These representations will be replaced with a general representation for definite NPs in Section 4.</Paragraph> <Paragraph position="4"> The LF representation for a pronoun must be compatible with our computational constraints. To be consistent with the modularity constraint, a pronoun's representation must be generated without utilizing the contextual information needed to select its antecedent. To obey the compactness and formal consistency constraints, a pronoun must be represented using a single representation that is consistent with the ways the pronoun can act given its position in a sentence. To conform with these constraints, we represent a pronoun as a pronoun function in LE This representation of pronouns is similar in spirit to the representation of pronouns as unique skolem constants in Charniak and McDermott (1985). Their representation allows the construction of basic logical structure of the sentence to precede pronoun resolution, a division consistent with our compactness and modularity constraints. However, because a constant is incompatible with a variable, their pronoun representation is incompatible with a bound variable meaning.</Paragraph> <Paragraph position="5"> A pronoun function is a composite representation reminiscent of a skolem function. Its role is to limit the range of possible antecedents for the pronoun without committing to one in particular. Each pronoun function is assigned a unique name (supplied by adding a unique number to the pronoun to distinguish it from other pronouns), and its argument list is specified using only syntax and sentence-level semantics to avoid violating the modularity constraint.</Paragraph> <Paragraph position="6"> A pronoun should be represented as a function of all the variables corresponding to quantified NPs that can affect its meaning (because they are representations for possible antecedents or can affect the meanings of other nonquantified antecedents such as definite NPs). By concentrating on variables of possible antecedents, we automatically include those variables that affect potential nonquantified antecedents. The argument list should also contain the variables of lambda operators that have scope over the position that the pronoun function fills in LF (in order to allow sloppy readings of elided sentences).</Paragraph> <Paragraph position="7"> To provide an algorithm for automatically generating the LF for a pronoun, we must develop a mechanism for specifying its argument list. This mechanism should not automatically assign all of the variables associated with quantified NPs in the sentence to the argument list of a pronoun function because, in English, some quantified NPs are syntactically incompatible antecedents for the pronoun. Consider, for example, He loves every man, in which the antecedent for he cannot be every man. To determine Parse tree for Every man who saw every boy kicked his dog.</Paragraph> <Paragraph position="8"> which NPs can bind a pronoun in a sentence, we adapt Reinhart's (1983) c-command (or constituent-command), which is a relation on nodes in the sentence's parse tree. Node A c(constituent)-commands node B iff the branching node ~1 most immediately dominating A either dominates B or is immediately dominated by a node a2 which dominates B, and a2 is of the same category type as al. (Reinhart 1983, p. 23) To illustrate the concept of c-command, consider the parse tree for the sentence in Example 5, shown in Figure 2.</Paragraph> <Paragraph position="9"> Example 5 Every man who saw every boy kicked his dog.</Paragraph> <Paragraph position="10"> In Figure 2, NP1 c-commands his, but NP2 and NP3 do not. Reinhart claims that a pronoun can be bound by an NP if and only if the NP c-commands the pronoun.</Paragraph> <Paragraph position="11"> Hence, every man who saw every boy can bind his, but every boy cannot.</Paragraph> <Paragraph position="12"> C-command is very useful for ruling out incorrect antecedents for pronouns, as illustrated by Example 5. Unfortunately, it also makes some incorrect predictions. For example, it does not allow every man's to bind him in Every man's mother loves him or each candidate to bind him in A friend of each candidate supported him, even though the universal variables provide reasonable meanings for the pronouns. The difference between Example 5 and the previous two sentences is that in 5 the quantified NP in question is contained in a relative clause attached to an NP, whereas in the other two sentences they are not. A quantified NP is prevented from binding any pronoun outside of the clause even though the NP containing it c-commands the pronoun. Hence, we adapt Reinhart's binding rule 4 to allow quantified NPs to bind a pronoun if they c-command the pronoun or are embedded in another c-commanding NP but not contained in a relative clause.</Paragraph> </Section> </Section> <Section position="6" start_page="425" end_page="430" type="metho"> <SectionTitle> 4 We could have also adapted the binding constraints found in Bach and Partee (1980). </SectionTitle> <Paragraph position="0"> Mary P. Harper Ambiguous Noun Phrases in Logical Form Armed with this syntactic rule for determining which quantified NPs can have an impact on the meaning of a pronoun, we can now specify a pronoun's representation. A pronoun is represented as a uniquely named function of all the variables corresponding to operators that can have scope over it. These operators include: * any lambda operator that has scope over the pronoun function in LF, * any nonsubject quantified variable corresponding to an NP that can bind the pronoun given our modification of Reinhart's binding rule.</Paragraph> <Paragraph position="1"> Because a quantified subject's variable is abstracted by the sentence's lambda operator, the lambda variable subsumes the quantified variable as the only subject-related variable required in the argument list. The lambda variable makes available both the direct quantified variable meaning as well as the indirect lambda variable meaning since we can choose to apply the quantified variable or not, depending on our need.</Paragraph> <Paragraph position="2"> Now that the initial representation for pronouns has been specified, consider a series of examples, beginning with Example 6: Example 6 Fred loves himself.</Paragraph> <Paragraph position="3"> Fred22, A(x)(love x (himselfl x)) The subject Fred is represented as the skolem constant Fred22 and the VP as the lambda function, &(x)(love x (himself1 x)). To create a representation for the entire sentence, we apply the subject to the lambda function; this is indicated by placing the subject to the left of the function. Though it is more traditional to apply the term to the right of the function, we chose this method to make it easier to see the correspondence between the sentence and its representation (following Sag \[1976\]). The logical subject fills the first slot after the predicate in the VP and the logical object fills the second. Since the sentence in Example 6 contains no universal or indefinite NPs, himself is represented simply as a pronoun function of the lambda variable x. And as we mentioned previously, the name of this function is created by concatenating a unique integer onto the end of the pronoun string (i.e., himselfl).</Paragraph> <Paragraph position="4"> In the next example, the representation of the pronoun is affected by universal and indefinite NPs in the sentence.</Paragraph> <Section position="1" start_page="426" end_page="427" type="sub_section"> <SectionTitle> Example 7 </SectionTitle> <Paragraph position="0"> Every man showed a boy his picture.</Paragraph> <Paragraph position="1"> Vx: (man x) x, My)(show y (picture-of (his2 y z)) \[3z: (boy z) z\]) The syntactic subject of the sentence is universally quantified, the indirect object is existentially quantified, and the logical direct object his picture is represented as a function of the pronoun. As discussed above, the subject's lambda variable subsumes the universal variable; hence his is represented as a function of the subject's lambda variable plus the existential variable for a boy. The reader should note that there is no order implied by quantifiers in our LF (following Allen \[1987\]) and so the quantifiers can be ordered in two different ways to provide two possible final meanings for the sentence.</Paragraph> <Paragraph position="2"> Next, consider Example 8.</Paragraph> <Paragraph position="3"> Although this sentence contains a universal NP, he is represented as a pronoun function of the lambda variable x alone. The pronoun function's argument list does not include the variable y because every woman neither c-commands nor is embedded in an NP that c-commands the pronoun.</Paragraph> <Paragraph position="4"> Finally, consider a sentence containing only pronouns and definite NPs.</Paragraph> <Paragraph position="5"> Example 9 Fred showed his mother her picture.</Paragraph> <Paragraph position="6"> Fred22, Mx)(show x (picture-of (her1 x)) (mother-of (his2 x))) Both of the pronouns are represented as functions of the lambda variable x.</Paragraph> </Section> <Section position="2" start_page="427" end_page="429" type="sub_section"> <SectionTitle> 3.3 Pronouns: Updating Logical Form </SectionTitle> <Paragraph position="0"> When a pronoun's antecedent is known, 5 the LF containing it must be updated in a way compatible with its initial representation (to conform with the formal consistency constraint). To augment LF with antecedent information without creating an ill-formed LF (i.e., a LF with unbound variables), we assert equality statements in the lambda environment containing the pronoun function and limit the types of updates allowed.</Paragraph> <Paragraph position="1"> A pronoun function constrains its possible antecedents, and depending on the type and location of its antecedent, the pronoun function is equated with four different values.</Paragraph> <Paragraph position="2"> 1. If a universal or indefinite NP is the intrasentential antecedent for the pronoun, and its variable is an argument of the pronoun function (or is abstracted by a lambda operator whose variable is an argument of the function), then equate the pronoun function with that quantified variable.</Paragraph> <Paragraph position="3"> 2. If an NP represented as a function (i.e., a pronoun or definite) is the intrasentential antecedent for a pronoun, and its argument list is compatible with that of the pronoun function (either immediately or after more is known about its meaning), then equate the pronoun function with that function.</Paragraph> <Paragraph position="4"> 3. If a syntactic subject of a sentence is the intrasentential antecedent for the pronoun, and the lambda variable is an argument of the pronoun function, then equate the pronoun function with either the subject's lambda variable or some other value depending on the subject's type.</Paragraph> <Paragraph position="5"> 4. If an NP in a different sentence or some nonlinguistic entity is the antecedent for the pronoun (and it is compatible in number and gender with the pronoun), then equate the pronoun function with the discourse entity created for the antecedent.</Paragraph> <Paragraph position="6"> 5 In this paper, we are not concerned with how the correct antecedent for a pronoun is determined. However, our LF provides useful hooks for an implementation to request information of either a user or a pronoun resolution module in order to resolve ambiguity.</Paragraph> <Paragraph position="7"> Mary P. Harper Ambiguous Noun Phrases in Logical Form Some updates to logical form must be postponed until more is known about the meaning of a possible antecedent (as we will show in Section 4.3.2).</Paragraph> <Paragraph position="8"> Consider some examples of how LF is refined following pronoun resolution. Suppose, for example, that we decide that the antecedent for his, in Example 7, is a boy, then the LF for the sentence is modified as follows: Example 10 Every mani showed a boyj hisj picture.</Paragraph> <Paragraph position="9"> Vx: (man x) x, A(y)(and (show y (picture-of (his2 y z)) \[3z: (boy z) z\]) (= (his2 y z) z)) Simplification: Vx: (man x) x, A(y)(show y (picture-of z) \[3z: (boy z) z\]) To indicate that the antecedent for his is a boy, the pronoun function (his2 yz) is equated with the existentially quantified variable z (by pronoun update case 1). Notice that the equality statement is placed in the environment of the A(y) operator. If we had placed it outside of this environment, the variable y would have been unbound. The variable z is compatible with the pronoun's initial representation because we are limiting the function of y and z to be the identity function on z. Once the equality statement is asserted, we can simplify the LF as shown above. Notice that the meaning of the sentence is still underspecified since quantifier scoping has not been determined. Next, consider how the representation in Example 6 is augmented, given that the antecedent for himself is the subject Fred.</Paragraph> <Paragraph position="11"> The pronoun can refer to the subject either directly or indirectly, so its function is equated with Fred22 (by pronoun update case 2) or x (by case 3), respectively. Notice that we use Ora as a meta-or to annotate our logical form with multiple possible meanings for the pronoun (outside of the logical system for mapping to truth value, see Appendix A). No simplification of the LF is possible until one of the alternatives is chosen, but by using this device we can at least compactly represent the ambiguous ways that pronouns refer to syntactic subjects. In fact, if there are n pronouns whose antecedents are subjects, we can specify this ambiguity with O(n) updates, compared with providing 2 n different representations for the sentence. This allows us to more easily put off decisions about the pronoun's intended meaning until we process the required information, in contrast to an approach that uses an exponential amount of memory.</Paragraph> <Paragraph position="12"> Finally, consider how the LF in Example 9 is updated given that Fred is the antecedent of his and his mother is the antecedent for her.</Paragraph> <Paragraph position="14"> with those values. Also (her1 x) is equated with (mother-of (his2 x)) (by case 2) since the functions are compatible.</Paragraph> </Section> <Section position="3" start_page="429" end_page="430" type="sub_section"> <SectionTitle> 3.4 Pronouns and Verb Phrase Ellipsis </SectionTitle> <Paragraph position="0"> In this section, we discuss how to handle VPE, using Example 3 to illustrate our approach.</Paragraph> <Paragraph position="1"> Because the meaning of an elided VP is constrained by the meaning of the trigger VP, the trigger's meaning must be fixed before we provide the meaning of the elided VP. In particular, we must locate the antecedent for his; otherwise, the final meaning of the pronoun function in the trigger cannot limit the meaning of the pronoun in the elided sentence. Given that the antecedent for his is Fred, we augment the LF as shown in 14. To determine the intended meaning of the elided sentence, we must locate the trigger sentence, select a single meaning for that sentence, and replace the dummy lambda function with the lambda function representing the trigger VP. The trigger sentence's VP, shown in Example 14, contains a meta-or of equality statements signalling an ambiguity in the pronoun's meaning; hence, before providing the meaning of the elided VP, we must select a single meaning for the pronoun.</Paragraph> <Paragraph position="2"> The procedure for choosing between the two meanings is beyond the scope of this work; consequently, we will demonstrate that for each choice, we are able to provide a reasonable meaning for the elided sentence. If the pronoun his refers indirectly to the Mary P. Harper Ambiguous Noun Phrases in Logical Form subject of the trigger sentence, the first disjunct in 14 is selected. This choice provides the sloppy reading of the elided sentence, as shown below.</Paragraph> <Paragraph position="3"> There is no way to determine whether a particular sentence will be a trigger sentence until an elided sentence is processed, so all sentences are treated as potential trigger sentences. Later, when an elided sentence is detected, its trigger sentence must be located and disambiguated. Contextual information is needed to select trigger sentences, to choose antecedents for pronouns, and to select a single meaning for a pronoun function whose antecedent is a syntactic subject.</Paragraph> <Paragraph position="4"> We have described our LF representation of pronouns, which allows us to model pronoun behaviors both in normal use and in VPE and is compatible with our computational constraints. Next, we discuss the LF for singular definite NPs, which is slightly more complex than the logical form for pronouns.</Paragraph> </Section> </Section> <Section position="7" start_page="430" end_page="439" type="metho"> <SectionTitle> 4. Singular Definite NPs in LF </SectionTitle> <Paragraph position="0"> In this section, we develop an LF representation for singular definite NPs. As with pronouns, we wish to obey our computational constraints while providing a model of definite behavior. First, we discuss the behaviors we wish to model.</Paragraph> <Section position="1" start_page="430" end_page="432" type="sub_section"> <SectionTitle> 4.1 Definites: Linguistic Behavior </SectionTitle> <Paragraph position="0"> Like pronouns, definite NPs can be anaphoric. Anaphoric definites either depend on linguistic antecedents or denote salient individuals in the environment of the speaker/hearer. An anaphoric definite's antecedent can be found in previous sentences as in Fredi saw (hisi cat)j. The catj was chasing a mouse, or within the same sentence as in Every boyi who loves (hisi cat)j takes care of the animalj. In the first example, the antecedent for the cat is his cat, found in the previous sentence, hence, the cat adopts the discourse entity assigned to his cat. In the second, the animal cannot take a discourse entity as its meaning because its antecedent is his cat, which does not denote a particular cat.</Paragraph> <Paragraph position="1"> Definites, unlike pronouns, can have a complex syntactic structure. A definite NP's meaning can be affected by embedded NPs. While simple nonanaphoric definites (i.e., they contain no embedded NPs) act like constants when included in sentences with universal NPs, as in Every boy loves the woman; definite NPs containing pronouns often cannot be described as constants, as in Every boy loves his mother. The meaning of his mother depends on how the pronoun is resolved. If the antecedent for his is found in another sentence, then his mother could be represented as a constant, but if every Computational Linguistics Volume 18, Number 4 boy is the antecedent for his, then the universal quantifier corresponding to every boy distributes over his mother. When a quantifier distributes over a definite, the definite denotes different entities depending on the values assigned to the quantified variable.</Paragraph> <Paragraph position="2"> Any pronoun embedded in a definite NP can affect the definite's meaning: a possessive pronoun, one contained in a prepositional phrase (PP), or one contained in a relative clause attached to the definite.</Paragraph> <Paragraph position="3"> A quantified possessive in a definite NP always distributes over the NP, preventing it from acting like a constant, as in Every man's mother loves him. It can also bind any pronouns the definite NP c-commands. Quantified NPs contained in a PP attached to a definite NP can also distribute over the definite, as in The head of every public authority in New York is rich, though the meaning of the definite NP is ambiguous. If the universal distributes over the head of every public authority in New York, then its denotation depends on which public authority is considered. But if the universal does not distribute over the definite, then there is one particular person who heads all of the public authorities. Our initial representation must be compatible with either possibility.</Paragraph> <Paragraph position="4"> Not all embedded quantified NPs can distribute over a definite. Quantified NPs embedded in relative clauses attached to a definite NP are unable to distribute over the definite. This constraint prevents every boy from quantifying over the child who cares for every man, so the definite can only denote one particular child. Universal NPs that cannot distribute over a definite NP are also unable to bind a pronoun outside that phrase, as noted by May (1985) and Roberts (1987).</Paragraph> <Paragraph position="5"> We must also consider the behavior of definite NPs in VPE. The meaning of a definite NP is ambiguous whenever it contains a pronoun whose antecedent is the subject of the sentence, 6 as in the following example.</Paragraph> <Paragraph position="6"> Example 18 The postman/saw hisi dog.</Paragraph> <Paragraph position="7"> The policemanj did too.</Paragraph> <Paragraph position="8"> Possible Meanings: 1. The policeman saw the postman's dog.</Paragraph> <Paragraph position="9"> 2. The policeman saw his own dog.</Paragraph> <Paragraph position="10"> (strict reading) (sloppy reading) We must also provide a good representation for a definite subject, one that will account for the differences between universal and definite subjects in VPE (compare Example 19 with 18).</Paragraph> <Paragraph position="11"> Example 19 Every postman/saw hisi dog.</Paragraph> <Paragraph position="12"> Every policemanj did too.</Paragraph> <Paragraph position="13"> Possible Meanings: Every policeman saw his own dog. (sloppy reading only) Universal quantifiers cannot bind across sentences, so the only possible meaning for the elided sentence in 19 is the sloppy one. However, definite subjects support both sloppy and strict meanings (as shown in 18). If we choose a quantified variable to represent a definite subject, we would have to allow its quantifier to bind across sentences.</Paragraph> <Paragraph position="14"> The meaning of a definite NP is affected by its structure and the meanings of embedded NPs, as well as its potential anaphoric use. In the remainder of this section, we introduce our LF representation for definites, describe ways to update this Mary P. Harper Ambiguous Noun Phrases in Logical Form representation once ambiguity is resolved, and discuss how the representation is used in VPE.</Paragraph> </Section> <Section position="2" start_page="432" end_page="434" type="sub_section"> <SectionTitle> 4.2 Definite NPs: An Initial Representation </SectionTitle> <Paragraph position="0"> We represent definite NPs as functions of all of the variables that can affect their meanings. This representation satisfies our constraints by combining the advantages of definite descriptions (Russell 1971) (discussed in Section 7.2.1) with the functional notation we introduced to represent pronouns. A definite function is assigned a unique name (i.e., def with a unique integer subscript) to distinguish two occurrences of the same definite NP, has a restriction consisting of a single predicate or a conjunction of predicates derived from information contained in the NP, and has a (possibly empty) list of arguments containing: * any variables associated with lambda operators that have scope over it, * any variables associated with nonsubject quantified NPs that could bind a pronoun in that position, * any quantified variables associated with embedded quantified NPs that are not also embedded in a relative clause. 7 To illustrate our representation for definite NPs, consider the initial representation of the following sentence.</Paragraph> <Paragraph position="1"> Example 20 Every man showed a boy his picture.</Paragraph> <Paragraph position="2"> Vx: (man x) x, My)(show y ((defl y z) I (and (picture (defl y z)) (possess (his2 y z) (defl y z)))) \[3z: (boy z) z\]) This representation is very similar to Example 7 except for the definite NP, his picture, which is represented as the function (defl y z). The definite function's argument list consists of the variables y and z, just like the pronoun his. As in the representation of pronouns, we omit the variable x from the argument list because the lambda operator for y abstracts x, so y is the more general argument. Anything that affects the meaning of the pronoun also affects the meaning of the definite. The function's restriction is the conjunction of statements following the vertical bar. The vertical bar in the function serves two purposes: it is used to distinguish the function's definition (on the left) from references to it (on the right), and it indicates that the function's restriction should be expanded just like the restriction on an existential operator.</Paragraph> <Paragraph position="3"> This representation for definite NPs accounts for quantified NPs embedded in a definite. There are three cases to consider. The first is exemplified by Every man's mother loves him. The NP every man's mother does not denote a single mother; every man distributes over the definite noun phrase. A possessive quantified NP embedded in a definite NP always distributes over the definite. Hence, its variable must be included in the argument list of the definite function plus any other functions that the quantifier can affect (e.g., the pronoun function for him). Additionally, its quantifier must be moved to indicate that it has scope over the function, as shown below.</Paragraph> <Paragraph position="4"> 7 We should also add that a sententially attached PP with a quantified object can quantify over a definite as well (e.g., In every car, the driver turned the steering wheel, in which the universal distributes over both definites). Vx: (man x) ((defl x) I (and (mother (defl x)) (possess x'(defl x)))), &(y)(love y (him2 x y)) The second case concerns quantified NPs contained in a relative clause attached to a definite NP. They cannot have scope over the definite so their variables are automatically excluded from the definite's argument list. Furthermore, they cannot affect the meaning of any other NP outside of the relative clause. The following example uses these facts to represent The child who cares for every man visits him. s Example 22 The child who cares for every man visits him.</Paragraph> <Paragraph position="5"> ((defl) I (and (child (defl)) ((defl), &(x)(care x (for \[Vy: (man y) y\]))))), &(z)(visit z (him2 z)) The third case concerns quantified objects of PPs attached to a definite NPo These quantified NPs optionally have scope over the definite, and so our representation must be consistent with both possibilities. To avoid making decisions about whether a quantified object of a PP attached to a definite distributes over the definite, we include its variable in the function's argument list but leave the quantifier inside the function's restriction. This representation is shown below: Example 23 The head of every public authority in New York is rich.</Paragraph> <Paragraph position="6"> ((defl x) I (head-of \[Vx: (and (public-authority x) (in x New York)) x\] (defl x))), ,\(y)(rich y) Possible Meanings: 1. ((defl) I (head-of Vx: (and (public-authority x) (in x New York)) x (defl x))), )~(y)(rich y) 2. Vx: (and (public-authority x) (in x New York)) ((defl x) I (head-of x (defl x))), )~(y)(rich y)) Notice that the quantifier is placed inside the restriction of defl, and the variable x is placed in the argument list (for the semantics of such a function, see Appendix A). Later, after we decide whether or not the quantifier distributes over the definite, the initial representation will be updated, as discussed in the next section.</Paragraph> <Paragraph position="7"> The decision about whether a quantified object of a PP attached to a definite distributes over it cannot be made at the level of LF, though it has an impact on the quantifier's ability to bind pronouns (or anaphoric definites) in the sentence. Quantified objects of prepositions attached to a definite NP can bind pronouns in the sentence only when they distribute over the definite (Roberts 19879; May 1985). For example, in The secretary of every spy keeps an eye on him, the NP every spy can bind the pronoun him only when it has scope over the definite NP, giving it a distributive reading. However, we cannot make our representation of the pronoun him contingent on quantifier scoping decisions. Hence, we must include the variable in the argument list of the pronoun, and add a constraint to the pronoun resolution module preventing a pronoun function from being bound by a quantifier unless it distributes over the NP containing it. Because a definite function is initially a composite representation for all possible meanings of a definite NP, as appropriate information becomes available, we repeat8 We do not provide an explicit representation for who; instead we represent it by borrowing the relative head's representation. In 22, who is represented as a definite function. If the relative head was quantified, we would have represented the relative pronoun using the quantified variable. 9 Roberts modifies the definition of c-command to allow a PP-attached quantified NP to optionally c-command the same NPs as the containing NP.</Paragraph> <Paragraph position="8"> Mary P. Harper Ambiguous Noun Phrases in Logical Form edly update the function, refining its range of possible meanings. This process continues until there is no longer any ambiguity in the intended meaning of the definite. In the next section, we will discuss two methods for achieving this.</Paragraph> </Section> <Section position="3" start_page="434" end_page="436" type="sub_section"> <SectionTitle> 4.3 Definite NPs: Two Ways to Update the Initial Representation </SectionTitle> <Paragraph position="0"> with some value depending on its antecedent (as in the case of the pronoun function).</Paragraph> <Paragraph position="1"> For example, if the antecedent for a definite noun phrase occurs in another sentence and they are compatible in number and gender, then we would equate the definite function with the antecedent's discourse entity. Otherwise, the update is not allowed.</Paragraph> <Paragraph position="2"> Antecedents for definite NPs can also be found in the same sentence. For example, consider the initial representation of the following sentence.</Paragraph> <Paragraph position="3"> Example 24 The owner of every dog is afraid of the animal.</Paragraph> <Paragraph position="4"> ((defl x) \[ (and (owner (defl x)) (of (defl x) \[Vx: (dog x) x\]))), A(y)(afraid y (of ((def2 x y) \[ (animal (def2 x y))))) Because the representation of every dog is formally consistent with the definite function and additional definite constraints hold (e.g., number and gender agreement holds, the antecedent does not c-command the anaphoric definite, and the universal quantifier distributes over the subject so it can bind the anaphoric definite), the definite function can be equated with the antecedent's representation, as shown in 25.</Paragraph> <Paragraph position="5"> Example 25 (The owner of every dog/)/is afraid of the animal/ Vx: (dog x) ((defl x) \[ (and (owner (defl x)) (of (defl x) x))), A(y)(and (afraid y (of ((def2 x y) \[ (animal (def2 x y)))) (= (def2 x y) x))) This example would be problematic for approaches using either definite descriptions or definite quantifiers, which provide no mechanism for handling bound variable, anaphoric definites.</Paragraph> <Paragraph position="6"> turally complex definite NP (i.e., an NP containing embedded pronouns and quantified NPs), we must: 1) determine the meanings for all embedded NPs and 2) decide whether quantifiers corresponding to embedded quantified NPs not contained in relative clauses distribute over the definite. Given this information, we can refine the meaning of a definite function using the behavior of definite descriptions (Russell 1971) as our model. Any definite description that does not contain variables bound by outside quantifiers acts like a constant because of the uniqueness assumption (see Section 7.2.1 for a discussion of definite descriptions). On the other hand, if a quantifier has scope over the definite description (either because an embedded quantified NP distributes over it or the antecedent for an embedded pronoun is quantified), it denotes different individuals depending on the instantiation of that variable. Once we determine the meanings of the NPs contained in a definite function's restriction and decide whether any extractable embedded quantifiers should distribute over it, we examine the function's restriction to determine whether it contains any necessary arguments, i.e., variables bound by operators outside of the restriction. When the necessary arguments for a definite function are determined, its meaning can be refined in two different ways. A definite function can be anaphoric only if it does not Computational Linguistics Volume 18, Number 4 contain any necessary arguments. Hence, if a definite function contains no necessary arguments and is anaphoric, then it is equated with its antecedent (as in the previous section). However, if the definite function contains any necessary arguments or if it contains no necessary arguments but is nonanaphoric, then we limit the argument list to precisely the necessary arguments. By equating the original function with a new function over the necessary arguments, a process which we call argument simplification, we limit the initial composite representation of a nonanaphoric definite NP to its final meaning.</Paragraph> <Paragraph position="7"> For example, consider the initial representation of the sentence in Example 20. Notice that defl is defined as a function of all of the variables that could potentially cause it to change. Given that the antecedent for his is a boy, the LF is updated as follows.</Paragraph> <Paragraph position="8"> Example 26 Every man showed a boyi hisi picture.</Paragraph> <Paragraph position="9"> Vx: (man x) x, My) (show y ((defl y z) I (and (picture (defl y)) (possess (his2 y z) (def~ y z)) (= (his2 y z) z))) \[~z: (boy z) z\]) After (his2 y z) is replaced with the variable z, the only necessary argument for (defl y z) is z. Since the restriction of the function is bound by an outside operator, the definite cannot be anaphoric. Hence, to provide the final meaning of the definite, we apply argument simplification to replace the function (defl y z) with a more precise function of z, as shown in Example 27. Because of the meanings of equality and the vertical bar in the restriction of the function, this representation is simplified, as shown below. Example 27 Every man showed a boy/hisi picture.</Paragraph> <Paragraph position="10"> Vx: (man x) x, )~(y)(and (show y ((defl y z) I (and (picture (defl y z)) (possess (his2 y z) (defl y z))</Paragraph> <Paragraph position="12"> Simplification: Vx: (man x) x, ),(y)(and (show y (def3 z) \[3z: (boy z) z\]) (picture (def3 z)) (possess z (def3 z))) Also consider how we update the initial representation of the sentence in Example 23. If every public authority in New York distributes over the definite function, then the universal quantifier Vx is extracted from the restriction prior to applying argument simplification, as shown below.</Paragraph> <Paragraph position="13"> Example 28 The head of every public authority in New York is rich.</Paragraph> <Paragraph position="14"> Vx: (and (public-authority x) (in x New York)) ((defl x) I (head-of x (defl x))), ,~(y)(rich y) Given this scoping decision, the variable x is free in the restriction of (defl x) and so the function must retain the argument. On the other hand, if every public authority in New York does not distribute over the function, then the quantifier remains in the restriction, as shown in 29.</Paragraph> <Paragraph position="15"> The head of every public authority in New York is rich.</Paragraph> <Paragraph position="16"> ((defl x) I (head-of \[Vx: (and (public-authority x) (in x New York)) x\] (deh x))), A(y)(rich y) Because the restriction of (defl x) contains no free variables, we must decide whether the definite is anaphoric or not. Assuming it is not, argument simplification is applied as shown below.</Paragraph> <Paragraph position="17"> Example 30 The head of every public authority in New York is rich.</Paragraph> <Paragraph position="19"> Hence, we can systematically derive both of the possible meanings for the definite NP the head of every public authority.</Paragraph> <Paragraph position="20"> The availability of a definite NP as an antecedent for a pronoun depends on its intended meaning, which cannot be determined using only syntactic information. Hence, c-command does not always correctly predict when definites are accessible as antecedents for anaphoric expressions. To determine the intended meaning of a definite, we must determine the meanings of all embedded NPs and decide whether any embedded quantified NPs distribute over the definite. Consider Example 31.</Paragraph> <Paragraph position="21"> What are the legal antecedents for her in this sentence? Certainly, the teacher is a candidate, but consider his mother. We cannot immediately determine whether his mother is a legal antecedent for her because (her6 x w) is not immediately compatible with the representation for his mother (i.e., (def3 x y z)). We must first determine the meaning of his mother by selecting the antecedent for his. Depending on the outcome, the final meaning of his mother may or may not be accessible to the pronoun. If the antecedent for his is Fred or the teacher, then his mother can be the antecedent for her (following argument simplification). However, if the antecedent is every student, then his mother cannot be the antecedent for her.</Paragraph> </Section> <Section position="4" start_page="436" end_page="439" type="sub_section"> <SectionTitle> 4.4 Definites and Verb Phrase Ellipsis </SectionTitle> <Paragraph position="0"> To handle VPE, we first determine the meanings of definite functions contained in a trigger VP before providing the meaning of an elided sentence. Consider the following example in which the dog must denote the same dog in the trigger and elided sentences.</Paragraph> <Paragraph position="1"> Fred saw the dog.</Paragraph> <Paragraph position="2"> ((defl) I (name (defl) Fred)), &(x)(see x ((def2 x) I (dog (def2 x))) Before deriving the meaning of the elided sentence from this representation of the trigger, we must apply argument simplification to the definite function (def2 x) (assuming that it is nonanaphoric).</Paragraph> <Paragraph position="3"> Mary P. Harper Ambiguous Noun Phrases in Logical Form Depending on the meaning selected for the pronoun his, there are two different readings for the elided sentence.</Paragraph> <Paragraph position="4"> If his refers indirectly to Fred, the intended meaning for the trigger sentence is shown in 38.</Paragraph> <Paragraph position="5"> Notice that the function def2 denotes a different individual in the trigger and elided sentences, depending on the value of x.</Paragraph> <Paragraph position="6"> On the other hand, suppose that his refers directly to Fred. Then the intended meaning of the trigger sentence is shown in 40.</Paragraph> <Paragraph position="7"> Notice that once the pronoun function is replaced by (defl), the restriction of (def2 x) contains no free variables except those in the argument list of the function itself. Hence we update the LF as follows.</Paragraph> <Paragraph position="8"> Hence, our general LF representation of definite NPs allows us to derive both the sloppy and strict readings of the elided sentence.</Paragraph> <Paragraph position="9"> We have introduced a composite representation for definite NPs along with a way to update its meaning as more information becomes available. Our approach has several strengths: It is consistent with the three computational constraints discussed in Section 1; it handles a variety of definite behaviors with one mechanism; and it provides useful constraints on intrasentential antecedents for definites (in addition to traditional constraints like number and gender agreement) and a more flexible mechanism than c-command for determining whether a definite noun phrase is a possible antecedent for an anaphoric expression.</Paragraph> </Section> </Section> <Section position="8" start_page="439" end_page="447" type="metho"> <SectionTitle> 5. Indefinites in Logical Form </SectionTitle> <Paragraph position="0"> In this section, we develop an initial LF representation for singular indefinite NPs and provide a mechanism for updating it once additional information is processed.</Paragraph> <Section position="1" start_page="439" end_page="440" type="sub_section"> <SectionTitle> 5.1 Indefinites: Linguistic Behavior and Initial Representation </SectionTitle> <Paragraph position="0"> Singular indefinite NPs share many behaviors with singular definites, including the fact that the final meanings of both are affected by the meanings of embedded NPs.</Paragraph> <Paragraph position="1"> However, unlike definites, the meanings of indefinites are affected by negation, and so they cannot be initially represented as functions in LE For example: Example 43 Fred did not see a woman.</Paragraph> <Paragraph position="2"> Possible Meanings: a. ~x: (woman x) Not(see Fred x) b. Not 3x: (woman x) (see Fred x) = Vx: (woman x) Not(see Fred x) Whenever there is negation in a sentence containing an indefinite, two meanings of the sentence are possible. If the negation does not have scope over the indefinite, Mary P. Harper Ambiguous Noun Phrases in Logical Form then the indefinite is represented as an existential outside the scope of the negation as shown in 43a (and could be represented as a function). In contrast, if the negation has scope over the indefinite, then it has scope over the existential operator making it equivalent to a universal (as shown in 43b). If we represent the indefinite in 43 as a function before deciding whether the negation has scope over it, then the second reading could not be expressed. Thus, we represent indefinites initially as existentially quantified and restricted variables, as in Example 44.</Paragraph> <Paragraph position="3"> Example 44 Fred saw a dog.</Paragraph> <Paragraph position="4"> ((defl) I (name (defl) Fred)), ;~(x)(saw x \[3y: (dog y) y\]) To provide a scope-neutral form, we place the quantified term \[3y : (dog y) y\] directly into the predicate-argument structure (see Appendix A). This initial representation of the indefinite is provided using only syntactic information and knowledge about how to map arguments into the predicate-argument structure, obeying the modularity constraint. Once quantifier scoping information is available, we update the sentence's LF, using a mechanism that also allows us to account for several interesting indefinite behaviors.</Paragraph> </Section> <Section position="2" start_page="440" end_page="445" type="sub_section"> <SectionTitle> 5.2 Indefinites: Updating the Initial Representation </SectionTitle> <Paragraph position="0"> An approach that models indefinites solely as existentially quantified variables cannot account for the variety of behaviors of indefinite NPs, since quantifiers cannot have scope over variables in other sentences. This accounts for the fact that the quantified NP cannot be the antecedent for the pronoun she in: The boy kissed everygirl. She slapped him. However, in The boy kissed a girl. She slapped him, the antecedent for she can be a girl.</Paragraph> <Paragraph position="1"> A similar problem arises in VPE. If the trigger VP contains a pronoun whose antecedent is an indefinite subject, two possible meanings for the elided sentence are possible, as in Example 45.</Paragraph> <Paragraph position="2"> Example 45 A postman/saw hisi dog.</Paragraph> <Paragraph position="3"> A policemanj did too.</Paragraph> <Paragraph position="4"> Possible Meanings: a. A policeman saw the postman's dog.</Paragraph> <Paragraph position="5"> b. A policeman saw his own dog.</Paragraph> <Paragraph position="6"> (strict reading) (sloppy reading) When the antecedent for his is a postman, the elided sentence has two possible meanings, but a quantified representation for the indefinite can only account for one of them. 1deg Given that the antecedent for his is a dog, there are two possible representations for the trigger VP. The pronoun function is either replaced with the lambda or the existential variable corresponding to the subject. If we use the first representation of the pronoun, we are able to derive the sloppy reading. But, if we use the second, then the meaning for the elided sentence would contain an unbound variable because existential quantifiers do not have scope across sentences.</Paragraph> <Paragraph position="7"> Donkey sentences (originally noticed by Geach \[1962\]) suffer a similar difficulty. A typical donkey sentence is Every miner who owns a donkeyi beats iti. Though the existential 10 This example is in sharp contrast with Every postmani saw his i dog. Every policeman did too. The elided sentence can only mean Every policeman saw his own dog. Computational Linguistics Volume 18, Number 4 operator corresponding to a donkey cannot have scope over the pronoun it, the NP a donkey can be its antecedent. In contrast, in Every miner who brushed every donkey beat it, the NP every donkey cannot be the antecedent for it.</Paragraph> <Paragraph position="8"> These examples indicate that our initial representation of a singular indefinite is insufficient for modeling the variety of linguistic behaviors exhibited. However, the existential operator is only necessary until we can determine that the indefinite is not in the scope of a negation and can decide what has scope over the indefinite. After this information is available, it is desirable to transform the initial representation into a form more compatible with the behaviors of indefinites, especially if that transformation refines the meaning of the indefinite (conforming with the formal consistency constraint). Our solution is to eliminate the existential operator and replace the existential's variables with skolem functions.</Paragraph> <Paragraph position="9"> Once scoping is specified, each existentially quantified variable is replaced by a function whose argument list consists of all of the universally quantified variables that have scope over the existential operator. To demonstrate how existential variables are replaced by functions during skolemization, consider Example 46.</Paragraph> <Paragraph position="10"> la. 3x: (man x) Vy: (woman y) (see x y) b. Vy (and (man (indef34)) (if (woman y) (see (indef34) y))) 2a. Vy: (woman y) 3x: (man x) (see x y) b. Vy (if (woman y) (and (man (indefg5 y)) (see (indef35 y) y))) There are two meanings for the sentence in 46, indicated in la and 2a, and these meanings are preserved when the existential variables are replaced by skolem functions, as shown in lb and 2b.</Paragraph> <Paragraph position="11"> When an indefinite is represented as an existentially quantified variable and is not in the scope of negation, it can be replaced by an indefinite function. However, additional information about the indefinite noun phrase must be gathered before such a transformation is performed. We need to:</Paragraph> <Paragraph position="13"> Determine the antecedents of embedded pronouns and anaphoric definites.</Paragraph> <Paragraph position="14"> Determine the meanings of embedded definite and indefinite NPs. Determine whether any embedded universally quantified NPs, not contained in a relative clause, distribute over the indefinite. (The operators that bind the variables contained in the restriction of the existential operator, but that are not themselves contained in the restriction, necessarily have scope over the existential operator, as in the case of definite functions).</Paragraph> <Paragraph position="15"> Determine whether quantifiers within the sentence have scope over the indefinite, even if those quantifiers do not bind a variable contained in the indefinite's restriction. (The final meaning of an indefinite is affected by quantifiers &quot;that could never affect the meaning of a definite. Consider the sentence, His mother saw every boy. Despite the fact that his needs an antecedent, syntactic constraints eliminate every boy from the list of Mary P. Harper Ambiguous Noun Phrases in Logical Form .</Paragraph> <Paragraph position="16"> possible candidates. Hence, his mother acts as a constant in the sentence. Compare this sentence with A friend of his saw every boy. Despite the fact that the antecedent for his cannot be every boy, a friend of his could still be in the scope of every boy).</Paragraph> <Paragraph position="17"> Determine whether lambda operators have scope over the indefinite, even if they do not bind a variable contained in the indefinite's restriction. (We discuss why lambda operator scope over an indefinite is an issue for handling VPE in the next section).</Paragraph> <Paragraph position="18"> To demonstrate how much information is necessary to determine the final meaning of an indefinite NP, consider the initial representation for the sentence Every man showed every boy a picture of his mother.</Paragraph> <Paragraph position="19"> Example 47 Every man showed every boy a picture of his mother.</Paragraph> <Paragraph position="20"> Vx: (man x) x, ~(y)(show y \[3w: (and (picture w) (of w ((defl w y z) I (and (mother (defl w y z)) (possess (his2 w y z) (defl w y z)))))) w\] \[Vz: (boy z) z\]) Before determining the final meaning of the indefinite, we must determine the meaning of the definite NP his mother and decide whether Vx, Vz, or &(y) has scope over it. If the antecedent for his is every boy, the updated LF is shown in Example 48. Because z is unbound in the restriction of the existential, Vz must have scope over 3w; however, we must still determine whether Vx or &(y) also have scope over the existential. If they do not, we update the LF by replacing the existential variable w with a function of z.</Paragraph> <Paragraph position="21"> By replacing the existential variables with a function of z, we indicate that only Vz has scope over the existential. This final meaning is compatible with the initial representation of the indefinite in 47, but we have constrained the initial meaning with additional information.</Paragraph> <Paragraph position="22"> There are several advantages gained by replacing existential variables by functions. First, it provides a way to indicate quantifier scoping in a representation containing only universal and existential quantifiers. Second, the method of indicating scope is similar to Allen's (1987) method since we are not limited to expressing scope as a linear string of operators. Third, universal variables cannot be replaced with functions. Hence, skolemization may be useful for modeling the differences between universals and indefinites in English. Fourth, once quantifier scoping information is available, the replacement of existential variables with functions is a meaning preserving operation as required by the formal consistency constraint. Finally, a functional representation for an indefinite allows us to account for several behaviors that are poorly modeled using an existential variable representation of indefinites alone.</Paragraph> <Paragraph position="23"> The functional representation for an indefinite provides a mechanism for determining whether a singular indefinite can be the antecedent for a singular pronoun in a subsequent sentence. Consider the example: Every woman saw a dog. It bit the tallest woman. The antecedent for the pronoun it can be a dog only if the universal operator corresponding to every woman does not have scope over the existential. Consider the initial representation of the first sentence.</Paragraph> <Paragraph position="24"> Example 51 Every woman saw a dog.</Paragraph> <Paragraph position="25"> Vx: (woman x) x, A(y)(see y \[3z: (dog z) z\]) Now, suppose that the universal has scope over the indefinite, then the LF would be updated as follows.</Paragraph> <Paragraph position="26"> Example 52 Vx: (woman x) x, A(y)(see y ((indef34 x) I (dog (indef34 x)))) A consequence of this scoping decision is that a dog cannot be the antecedent for it, without causing a violation of the formal consistency constraint, n On the other hand, 11 Even if we construct a discourse entity for a dog, following Webber (1978), the discourse entity for (indef34 x) would denote a set of dogs, and the pronoun resolution module would not allow a plural entity to be the antecedent for a singular pronoun.</Paragraph> <Paragraph position="27"> Mary P. Harper Ambiguous Noun Phrases in Logical Form if the universal does not have scope over a dog, then the LF for the first sentence (shown in 51) is updated as shown in 53.</Paragraph> <Paragraph position="28"> Example 53 Vx: (woman x) x, %(y)(see y ((indef37) I (dog (indef37)))) Because a dog is represented as a function with no arguments, it is compatible with the pronoun function representing it. ~2 The representation is also useful for handling the donkey sentence Every miner who owns a donkeyi beats iti. The antecedent for it is a donkey, yet in English, a quantified NP contained in a relative clause attached to an NP cannot bind a pronoun outside of that clause. Hence, it cannot be bound by the existential quantifier corresponding to a donkey. However, once the meaning of the indefinite has been determined it may become formally consistent with the pronoun function. Consider the initial representation of this sentence.</Paragraph> <Paragraph position="29"> Example 54 Every miner who owns a donkey beats it.</Paragraph> <Paragraph position="30"> Vx: (and (miner x) x, A(y)(own y \[3z: (donkey z) z\])) x, A(w)(beat w (itss w)) The pronoun it is represented as a function of w only and cannot be equated with z. However, by replacing the existential term in the relative clause with a functional term, we will be able to assert that the antecedent for it is a donkey. To replace the variables corresponding to the existential operator with a function, we must determine whether the existential quantifier is in the scope of negation. Assuming that the negation introduced by the restriction on the universal operator does not have scope over a donkey, we replace the existential term with a function whose arguments are the variables corresponding to operators that have scope over it. If only Vx has scope over the existential (i.e., A(y) does not have scope), we can assert the anaphoric relationship between it and a donkey.</Paragraph> <Paragraph position="31"> x, A(w)(and (beat w (itss w)) (= (itss w) (indef22 x)))) Since (it58 w) is consistent with a function of x (because A(w) abstracts the variable x), we can assert the anaphoric relationship, as shown above. 13 It is important to note that the initial type of a quantified NP determines whether skolemization can make it accessible to a pronoun function whose argument list does not contain that NP's variable. If an NP is initially represented as a universal, then unless the universal variable is included in the argument list of the pronoun, it cannot be the antecedent for that pronoun even if it is in the scope of negation. For example, 12 It could also be used to create a singular discourse entity compatible with the singular pronoun. 13 Our solution has much in common with Webber's (1978) parameterized individuals. Webber introduces a parameterized individual (which looks much like an indefinite function) as the antecedent for it. However, she does not modify the initial representation of the indefinite. Computational Linguistics Volume 18, Number 4 in Every miner who did not see every donkey beat it, the antecedent for it cannot be not every donkey. In contrast, so long as an indefinite remains an existential, even if it cannot bind the pronoun, it may become accessible to the pronoun once we determine its precise behavior and convert it into a functional term.</Paragraph> </Section> <Section position="3" start_page="445" end_page="447" type="sub_section"> <SectionTitle> 5.3 Indefinites and Verb Phrase Ellipsis </SectionTitle> <Paragraph position="0"> The representation of an indefinite as a function is also very useful for modeling indefinite subjects in VPE. Consider Example 45 again. By converting existentially quantified variables into functions, we are able to provide the strict meaning for the elided sentence without creating an ill-formed representation; we simply replace the subject's existential variables with a skolem constant (assuming it is not in the scope of another quantifier).</Paragraph> <Paragraph position="1"> To properly model indefinite NPs contained in a trigger VP for VPE, we must consider whether the lambda operators in a VP representation have scope over indefinites represented as existentially quantified variables. Consider Example 56.</Paragraph> <Paragraph position="2"> Example 56 Fred saw a dog.</Paragraph> <Paragraph position="3"> George did too.</Paragraph> <Paragraph position="4"> Possible Meanings: 1. George saw the same dog that Fred saw.</Paragraph> <Paragraph position="5"> 2. George saw a different dog than Fred saw.</Paragraph> <Paragraph position="6"> When an indefinite NP occurs in the trigger VP, the elided sentence is ambiguous. This is in contrast to Example 32 discussed in Section 4.4: Fred saw the dog. George did too, in which the elided sentence can only mean George saw the same dog that Fred saw. By ignoring lambda operators when converting existential variables to skolem functions, we would be unable to provide the second meaning for the elided sentence in this example.</Paragraph> <Paragraph position="7"> To illustrate this point further, consider how we determine the three meanings of the elided sentence in Example 57.</Paragraph> <Paragraph position="8"> Mary P. Harper Ambiguous Noun Phrases in Logical Form Now, before converting the existential into its functional form, we must determine which meaning of the pronoun is intended.</Paragraph> <Paragraph position="9"> Suppose that the pronoun his refers indirectly to the subject; then, the trigger LF is refined as shown in 60.</Paragraph> <Paragraph position="11"> Since ,Ux) must have scope over the existential to bind the variable x in its restriction, the existential must be a function of that variable and can be used to provide the third meaning of the elided sentence in 57, as shown in 61.</Paragraph> <Paragraph position="12"> Example 61 Trigger: Fred/saw a friend of his/.</Paragraph> <Paragraph position="14"> Ellipsis: George clid too. (George saw a friend of George's.)</Paragraph> <Paragraph position="16"> In contrast, assume that the pronoun his refers directly to Fred. This choice is reflected in the LF shown in 62.</Paragraph> <Paragraph position="18"> We must still determine whether &(x) has scope over the existential. If it does, we replace the existential variables with a function of x as shown in 63, allowing us to provide the second reading of the elided sentence in 57.</Paragraph> <Paragraph position="19"> Example 63 Trigger: Fred/saw a friend of his/.</Paragraph> <Paragraph position="21"> Ellipsis: George did too. (George saw a different friend of Fred's.)</Paragraph> <Paragraph position="23"> Finally, if ,~(x) does not have scope over the existential, then we replace the existential variable with a skolem constant, allowing us to provide the first reading of the elided sentence in 57.</Paragraph> <Paragraph position="25"> Hence, we are able to provide all three readings for the elided sentence in 57.</Paragraph> <Paragraph position="26"> We have described an initial representation for indefinite NPs along with a way to update the representation after more information becomes available. Our approach is consistent with the three computational constraints discussed in Section 1 and models a variety of indefinite behaviors.</Paragraph> </Section> </Section> <Section position="9" start_page="447" end_page="448" type="metho"> <SectionTitle> 6. Implementation </SectionTitle> <Paragraph position="0"> In this section, we describe the operation of a program to generate the meanings for sentences that contain intersentential VPE. This implementation demonstrates that the LF described in this paper can be automatically generated during sentence parsing and that the intended meaning of the LF can be determined through a series of machine guided steps (see Figure 3).</Paragraph> <Paragraph position="1"> The program parses the sentences in the following example, starting with the trigger sentence.</Paragraph> </Section> <Section position="10" start_page="448" end_page="450" type="metho"> <SectionTitle> 2. George saw George's mother's picture. </SectionTitle> <Paragraph position="0"> If there are multiple parses for a sentence, there will be multiple LFs with each LF corresponding to a different parse tree. In such a case, the user is prompted to select the intended parse (and hence, the intended LF). The trigger sentence in Example 65 has a single parse tree and the parser produces the single LF shown in 66.</Paragraph> <Paragraph position="2"> In this program, we label the roles of the noun phrases to keep track of logical subjects and objects in order to prevent a passive voice sentence from becoming the trigger for an active voice sentence (or vice versa). We label the subject's lambda variable as the logical subject for a sentence with active voice or as the logical object for a sentence with passive voice. We do not specify case roles like agent in our LF because their determination may require contextual information. Also, the case role of a subject need not be the same in the trigger and elided sentences, e.g., Fred hit the window. The hammer did too. The subject in the first sentence is probably filling the role of agent, whereas the subject of the second fills the role of instrument.</Paragraph> <Paragraph position="3"> Because our program is designed to provide meanings of sentences with VPE, it examines each LF provided by the parser to see if it contains an elided VP. If it does not contain one, then additional processing of the potential trigger sentence is put off until the final meaning of the sentence is needed. Ideally, processing should be done as information becomes available; however, for the purpose of this implementation, it is conceptually simpler to refine the trigger LF only when its meaning must be determined to disambiguate an elided VP. Hence, the LF in 66 is saved on a stack of recently processed sentences.</Paragraph> <Paragraph position="4"> To determine the meaning of an elided sentence (a process depicted in Figure 4), the program must first locate the trigger sentence. (The fact that the representation contains an elided VP is indicated by the dummy predicate in 67.) Example 67 George did too.</Paragraph> <Paragraph position="5"> ((defs) I (name (defs) George)), &(w)(dummy33 (subject w)) To accomplish this, the program has the user select the LF for the trigger sentence from the LFs stored on the stack (assume the user selects the LF in 66). Since the meaning of the trigger sentence is underspecified, the meaning of the elided sentence is also ambiguous. Thus, the program must next determine the trigger sentence's meaning.</Paragraph> <Paragraph position="6"> To do this, the program must find antecedents for all of the pronoun functions. Once the pronouns have antecedents, it handles all definites, and then all quantified NPs. This sequence occurs unless the antecedent for a pronoun is an NP in the sentence Evaluation of the elided sentence for Example 65.</Paragraph> <Paragraph position="7"> that is not immediately compatible with the pronoun, in which case, the program attempts to refine the NP's meaning to determine whether it is formally consistent with the pronoun. In our example, the program begins disambiguating the trigger sentence by searching for antecedents of the pronoun function (his2 y). The user is asked to determine whether the antecedent occurs in another sentence. Assuming it doesn't, the pronoun compatibility routine prompts the user to select an antecedent NP from the sentence, using information about the type and location of the pronoun to eliminate impossibilities. 14 If a possible antecedent is a definite or indefinite, even if it is not formally consistent with a pronoun function, it may still become compatible with the pronoun once its meaning is determined.</Paragraph> <Paragraph position="8"> To refine the meaning of definite noun phrases with embedded NPs, the program must first determine the meanings of embedded pronouns, definites, and indefinites, and then determine whether any embedded quantifiers have scope over the definite.</Paragraph> <Paragraph position="9"> For example, to determine the meaning of (def3 y), the meaning of (his2 y) must be determined. If the user selects Fred as the antecedent, then the program also prompts the user to select either the lambda variable y or (defl). If the user picks the lambda variable (as shown in Figure 4) then the variable replaces the pronoun function in LF. Given this choice, y is a necessary argument for (def3 y), so the definite cannot be anaphoric and cannot be further simplified. Since y is a necessary argument for (def4 y), there is no modification of that function.</Paragraph> <Paragraph position="10"> The program continues processing the meanings of the noun phrases in the sentence until the meanings of all constituent NPs are determined (or until an error occurs). Assume, in our example, that the program produces the following refined LF for the trigger sentence in Example 68.</Paragraph> <Paragraph position="11"> 14 If the pronoun is reflexive, its antecedent must occur in the same clause and cannot be embedded in an NP. On the other hand, if the pronoun is not reflexive, its antecedent cannot be in the same clause unless the pronoun or its antecedent is embedded in another noun phrase. Finally, potential antecedents must be formally consistent with the pronoun function.</Paragraph> <Paragraph position="12"> In the event that the trigger sentence contains more than one VP, the program prompts the user to select the trigger VP. Since there is only one VP in 68, the program checks it to ensure that it contains no free variables, 15 is compatible with the voice of the elided VP, and that a trigger sentence with more than one pronoun whose antecedent is the same syntactic subject obeys the multiple pronoun constraint discussed in Harper (1990). Notice that the VP in 68 contains no free variables and is compatible in voice to the elided VP in 67. Hence, the program replaces the elided VP with the trigger VP as shown in 69, and the meaning of the elided sentence is determined.</Paragraph> </Section> <Section position="11" start_page="450" end_page="458" type="metho"> <SectionTitle> 7. Related Work </SectionTitle> <Paragraph position="0"> Other researchers have developed an intermediate representation for a sentence from syntactic information (Pollack and Pereira 1988; Alshawi and van Eijck 1989). These approaches agree that in order to determine the meaning of a sentence, it is useful to build a partial meaning that is augmented once contextual information becomes available. These approaches, however, use a different scheme for indicating the final meaning of a sentence and do not handle VPE.</Paragraph> <Paragraph position="1"> In the rest of this section, we review past representations of pronouns, definite NPs, and indefinite NPs. We emphasize VPE research because it considers not only the representation of sentences in general, but also the representation of trigger sentences. Each approach is examined in the light of its modeling capability and our computational constraints.</Paragraph> <Section position="1" start_page="450" end_page="452" type="sub_section"> <SectionTitle> 7.1 Verb Phrase Ellipsis and Models of Pronouns </SectionTitle> <Paragraph position="0"> Pronouns are often classified as either bound variable or referential pronouns (Sag 1976; Webber 1978; Reinhart 1983; Partee and Bach 1981), but the adequacy of this dichotomy is questionable (as discussed in Section 3.1). Models of VPE must consider not only pronouns in normal sentences but also pronouns in trigger sentences, and must account for the ambiguity that arises when a pronoun's antecedent is the syntactic subject of a trigger sentence. Sag (1976) and Webber (1978) handle this ambiguity by introducing a rule to replace a pronoun whose antecedent is known to be the syntactic 15 This program only deals with intersentential VPE. Hence, all variables in the trigger VP must be bound in the VP, otherwise the elided sentence cannot receive a meaning. If we augment our approach to handle antecedent-contained ellipsis, we would have to allow variables bound by an operator outside of the VP but inside the meaning of the sentence. Computational Linguistics Volume 18, Number 4 subject of a trigger sentence with the larnbda variable corresponding to that subject.</Paragraph> <Paragraph position="1"> They also assume that a pronoun whose antecedent is a nonsubject definite is necessarily referential. But a problem arises when a pronoun's antecedent is a nonsubject, nonreferential definite, as in Example 70.</Paragraph> <Paragraph position="2"> 1. George showed Fred's mother Fred's mother's dog.</Paragraph> <Paragraph position="3"> 2. George showed George's mother George's mother's dog.</Paragraph> <Paragraph position="4"> Impossible Readings: 3. *George showed George's mother Fred's mother's dog.</Paragraph> <Paragraph position="5"> 4. *George showed Fred's mother George's mother's dog.</Paragraph> <Paragraph position="6"> Given the indices on the NPs, the elided sentence has meanings 1 and 2, shown in 70. Sag's and Webber's models correctly allow meaning 1 and incorrectly allow meaning 3 at the expense of meaning 2 (because Fred's mother can be the only meaning for her given their approach).</Paragraph> <Paragraph position="7"> Reinhart (1983) also indicates that pronouns are either bound variables or referential, providing a syntactic rule for determining when a pronoun can be bound by its antecedent: a pronoun can be bound by an NP if and only if it c-commands the pronoun. Reinhart does not represent pronouns or definites as quantified terms, yet she claims that when a pronoun's antecedent is a definite NP or a pronoun that c-commands the pronoun, then the pronoun is bound by a lambda operator abstracting the antecedent. At first glance, the idea of binding a pronoun with the lambda operator of its antecedent (given that the NP c-commands the pronoun) seems promising; it can be used to handle Example 70. However, in English, a nonreferential definite can be a pronoun's antecedent even if it does not c-command the pronoun, as in Every mani gave the psychiatrist who cares for (hisi mother)j herj diary. Reinhart can only provide the pronoun her with a referential meaning, which is inappropriate in this case. Also, Reinhart assumes that the lambda variable is the only nonreferential representation for a pronoun whose antecedent is a definite NP in the same sentence. If this is correct, then there should only be one meaning for a pronoun whose antecedent is a nonreferential definite subject. However, consider Example 71 (which was inspired by an example in Sells, Zaenen, and Zec \[1989\]).</Paragraph> <Paragraph position="8"> Reinhart's approach can only provide the first meaning of the elided sentence in 71. This example suggests that pronouns can refer to definite subjects in two nonreferential ways; lambda abstraction accounts for only one of them.</Paragraph> <Paragraph position="9"> Partee and Bach (1981) attempt to dispense with LF in translating from syntax to final interpretation, building on Montague's (1970) general theory (with a few modifications to get around the strict compositionality of that approach). All of the possible representations for ambiguous sentences are simultaneously generated, avoiding the Mary P. Harper Ambiguous Noun Phrases in Logical Form need for an intermediate level of representation. They directly provide model-theoretic interpretations for sentences containing pronouns and elided VPs. In their approach, null or elided VPs and pronouns are initially represented as variables. Pronouns are represented as variables that are either bound by some operator or remain free within the representation of the sentence. If a pronoun variable is unbound, it is assigned some value by a context assignment function, that is, a function that maps the variable to the individual that the pronoun denotes. In other words, pronouns are either bound variables in this model or they are referential. An elided VP is represented as a free property variable, typed to receive a value corresponding to a VP already in discourse. It receives its interpretation in much the same way as an unbound pronoun variable, with the exception that its antecedent must be available in linguistic context. Once the value of the null VP is specified, the meaning of the elided sentence is determined.</Paragraph> <Paragraph position="10"> Bach and Partee point out a variety of examples for which their approach fails. Because there is no mechanism for ensuring that a pronoun bound in the trigger sentence is bound by the same operator in the elided sentence, their approach provides a host of impossible interpretations for elided sentences. For example, their approach provides an impossible interpretation for the elided sentence in No man believes that Mary loves him. But she does, given that no man is the antecedent for him and the meaning of the null VP is loves him. Since the variable for him is unbound in the elided interpretation, it must be assigned a value (e.g., Fred) by the context assignment function. Partee and Bach also discuss examples in which an elided sentence receives an impossible interpretation when a free pronoun variable in the trigger VP becomes accidentally bound by a quantifier in the elided sentence. In addition to the problems pointed out by Partee and Bach, others arise if we assume that definite NPs are quantified.</Paragraph> <Paragraph position="11"> The above approaches also do not adhere to all of our computational constraints, which are essential in any computer model. For example, Sag's (1976), Webber's (1978), and Reinhart's (1983) approaches do not conform with the formal consistency constraint (because they replace pronoun strings with a variable to account for bound variable meanings of a pronoun) and Partee and Bach's approach (1981) does not conform with the compactness constraint.</Paragraph> </Section> <Section position="2" start_page="452" end_page="458" type="sub_section"> <SectionTitle> 7.2 Past Representations of Definites and Indefinites </SectionTitle> <Paragraph position="0"> In this section, we examine previous representations of definite NPs. In particular, we review definite descriptions and definite quantifiers. We also examine some recent work that departs from traditional representations of definite NPs (e.g., Heim \[1982\]; Roberts \[1987\]; Kamp \[1981\]; Klein \[1987\]).</Paragraph> <Paragraph position="1"> to represent definite NPs using definite descriptions or definite quantifiers. Russell (1971) introduced definite descriptions to capture the meaning of definite NPs like the dog in Example 72.</Paragraph> <Paragraph position="2"> Example 72 The dog barked.: (barked (~x)(dog x)) which means:</Paragraph> <Paragraph position="4"> Computational Linguistics Volume 18, Number 4 The definite description, (~x)(dog x), which stands for the object x such that the property (dog x) is true names a unique object, and hence, is translated into the formula, 3x (and (dog x) Vy ((dog y) *-4 (= x y))). Notice three important features of the meaning of the sentence in Example 72: the dog described by the definite NP is assumed to exist, is assumed to be unique, and fills some role in the sentence.</Paragraph> <Paragraph position="5"> Definite descriptions suffer from several problems. First, there is no role specified for the effect of context on the uniqueness statement (a problem noted by many people, including Allen \[1987\] and Hintikka and Kulas \[1985\]). For example, the dog in 72 is described as the-one-and-only the dog, regardless of context. Second, definite descriptions do not adequately model anaphoric definites (as noted by Hintikka and Kulas \[1985\]), which need not be unique and seem to adapt to the behavior dictated by their antecedents, as in Every boyi saw (hisi dog)/before the beast/saw himi. To cover this example, the definite description for the beast could be replaced by some value consistent with the representation of its antecedent, but not without violating the formal consistency constraint, or another representation for anaphoric definites could be devised, but this might violate our compactness constraint. Another difficulty involves the representation of Bach-Peters sentences, like (The boy who wrote her/)i kissed (the girl who loved himi)j, which cannot be represented without infinite recursion (as noted by Hintikka and Kulas \[1985\]).</Paragraph> <Paragraph position="6"> Other researchers have represented definites using the quantificational meaning of a definite description directly (e.g., Webber \[1983\] and Montague \[1970\]). While an in-place definite description simply fills an argument slot in a predicate-argument structure representing the sentence, a quantifier scopes an open sentence. For example, the sentence, The dog barked, could be represented as shown in 73 (along with a short-hand notation, where the quantifier 3! reads there exists a unique).</Paragraph> <Paragraph position="7"> Example 73 The dog barked.</Paragraph> <Paragraph position="8"> 3x: (and (dog x) (Vy ((dog y) *-* x=y))) (barked x) Short hand notation: 3!x: (dog x) (barked x) A quantificational representation for definites suffers from several problems. First, as one might guess, it suffers from the same uniqueness problem that in-place definite descriptions have. 16 Second, definite noun phrases do not exhibit the same type of quantifier scope ambiguity that other quantified NPs have. Compare the two sentences below.</Paragraph> <Paragraph position="9"> Example 74 a. Every man loves a woman.</Paragraph> <Paragraph position="10"> 1. Vx: (man x) 3y: (woman y) (loves x y) 2. 3y: (woman y) Vx: (man x) (loves x y) b. Every man loves the woman.</Paragraph> <Paragraph position="11"> 1. Vx: (man x) 3!y: (woman y) (loves x y) 2. 3!y: (woman y) Vx: (man x) (loves x y) 16 Dowty, Wall, and Peters (1981) suggest a nice way to fix the uniqueness problem. They eliminate the one-and-only aspect of a definite quantifier by relativizing uniqueness to a context of utterance (much as the domain of a universal NP must be relativized to a context of utterance). Though their solution improves definite quantifiers, it does not eliminate the problem with definite anaphora. Mary P. Harper Ambiguous Noun Phrases in Logical Form While the sentence in 74a has two different meanings and two representations, the sentence in 74b expresses one meaning, but has two representations. Definite descriptions, on the other hand, provide only a single representation for Every man loves the woman while providing multiple representations for sentences with definite scope ambiguity, as in The mechanic adjusted the steering wheel in each car. Third, as we have already discussed in Section 4.1, the strict reading of the elided sentence in Example 18 is difficult for a quantified definite representation to account for. Finally, pronoun references to definites are not constrained in the way that pronoun references to other quantified NPs are. For example, when a quantifier is embedded in a relative clause attached to a noun phrase, it cannot bind a pronoun outside of that clause, as in Fred gave the psychiatrist who cares for every woman her diary. In contrast, the pronoun her in Every man gave the psychiatrist who cares for his mother her diary can have his mother as its antecedent. One might, like Hornstein (1984), assume that definites are quantified but have different properties than universal quantifiers. However, this assumption does not correct some of the problems of definite quantifiers, such as uniqueness or their inability to model anaphoric definites.</Paragraph> <Paragraph position="12"> Rather than modifying definite quantifiers, it seems appropriate to represent definites as functions and gain four immediate benefits. First, the representation of a definite NP as a uniquely named function provides a mechanism for handling anaphoric definite NPs. Second, because of this naming convention, definite functions are distinct unless an equality is asserted between them. Third, since we represent each definite as a function with a restriction and the restriction provides us with a mechanism for determining the final meaning of the definite, we are able to capture the properties of a definite description without neglecting anaphoric definites. Fourth, while definite quantifiers can violate the compactness and formal consistency constraints, 17 our approach does not.</Paragraph> <Paragraph position="13"> 7.2.2 Heim (1982) and Discourse Representation Theory. Another approach to modeling definites was developed by Heim (1982) and Discourse Representation Theory (DRT) researchers Kamp (1981), Roberts (1987), Klein (1987). A hallmark of these approaches is their ability to handle anaphoric definites in a reasonable way and the commitment to modeling the meaning of a series of sentences in discourse, not just individual sentences.</Paragraph> <Paragraph position="14"> Heim (1982) treats definites and indefinites very similarly in her theory since both can be referred to across sentence boundaries, unlike universal NPs. To provide an interpretation for a sentence, Heim first determines the logical form for a sentence. Her logical form is essentially a parse tree with quantifier scoping information indicated, though it is not a logical representation for the meaning of the sentence. Once the logical form for a sentence is constructed, she provides a file change semantics for the sentence using felicity conditions to distinguish definites from indefinites. For example, the novelty-familiarity felicity condition states that indefinites should always cause a new discourse referent (or file card) to be created in the discourse model but definite NPs should not introduce a new discourse referent (or file card). Clearly, Heim's approach emphasizes the anaphoric aspect of the definite NP, which both definite descriptions and definite quantifiers fail to handle well.</Paragraph> <Paragraph position="15"> Heim's model handles anaphoric definites, but she must introduce accommodation to cover nonanaphoric definites. Consider the sentence Every mani loves hisi mother. 17 Formal consistency is violated when a wide scope definite is replaced with something that cannot be described as a constant and compactness is violated when a different representation is introduced to handle anaphoric definites. Discourse representations for Someone loves everyone.</Paragraph> <Paragraph position="16"> Accommodation, a concept introduced by Lewis (1979), allows the introduction of a new file card for a definite NP if and only if that NP is related to a previous file card. In the example, his mother is nonanaphoric; however, given that the antecedent for his is every man, accommodation allows the introduction of a new discourse referent. However, to provide a file card for his mother, the model must know that his refers to every man; accommodation requires more than syntax and sentence-level information to provide the representation for some definite NPs. Additionally, Heim's model would have trouble handling any definite NPs without an accommodation link to a previous NP in the discourse model.</Paragraph> <Paragraph position="17"> Heim's approach requires a considerable amount of information before a definite or indefinite NP is represented in file change semantics (including quantifier scoping information and pronoun antecedents). Hence, the process of representing sentences in file change semantics does not seem to comply with our modularity constraint.</Paragraph> <Paragraph position="18"> Kamp (1981) introduces a discourse theory similar to Heim's, called Discourse Representation Theory (DRT), providing a model-theoretic interpretation for discourse models. Both theories are motivated by the fact that pronouns in one sentence can have definite and indefinite antecedents in another sentence, while universals cannot bind pronouns in other sentences. Kamp's approach has been extended by several researchers (e.g., Klein \[1987\]; Roberts \[1987\]). Since Kamp does not discuss quantifier scope ambiguities, we introduce DRT as discussed by Roberts.</Paragraph> <Paragraph position="19"> In DRT, a set of construction rules converts natural language into discourse structures. To do so, however, quantifier scoping information must be specified. Consider the two DRT representations for Someone loves everyone. The first, shown in Figure 5a, corresponds to the reading in which someone has scope over everyone. The discourse referents for everyone and someone are xl and x2, respectively. The universal NP causes the creation of the antecedent-consequent box. Because x2 is defined outside of xl's antecedent-consequent box, it acts like a constant. The second, shown in Figure 5b, corresponds to the reading in which everyone has scope over someone. Because the discourse referent for someone is created in the consequent box of the universal, its denotation depends on xl, the discourse referent for everyone. Each quantifier scoping requires a different discourse representation for the meaning of the sentence.</Paragraph> <Paragraph position="20"> Mapping into a discourse representation is a top-down process that reduces the original sentence to a structure with a discourse referent for each noun phrase, with predicates indicating restrictions on the discourse referents as well as relations between discourse referents. As we already pointed out, a universal is represented by placing its discourse referent and restriction into an antecedent box, with additional sentence information placed in the consequent box (resulting in a meaning like a universal in Discourse representation for Every miner who owns a donkey beats it.</Paragraph> <Paragraph position="21"> predicate calculus). Indefinites and definites are represented by placing their discourse referents and restriction information in the box corresponding to the current level in the model. An accessibility relation determines when a pronoun can have a particular discourse referent as its antecedent. A pronoun's antecedent can be any discourse referent defined in the box where the pronoun is instantiated or in any box containing that box. Additionally, a pronoun in a consequent box can also refer to anything in the antecedent box (unless the antecedent is embedded in another box contained in the antecedent box).</Paragraph> <Paragraph position="22"> Roberts (1987) combines DRT with c-command to distinguish two types of binding, c-command binding and discourse binding. C-command binding occurs when the best way to represent the anaphoric NP is by replacing it with the variable associated with the operator of the NP that c-commands it. On the other hand, discourse binding is needed to handle anaphoric dependencies on things that don't c-command a pronoun or anaphoric definite. For example, consider the sentence Every miner who owns a donkey beats it. In Robert's approach, if a donkey had c-commanded it, then no discourse referent would be created for it; instead, the pronoun would be represented using the discourse referent of its antecedent. However, because a donkey does not c-command it, the sentence is handled as shown in Figure 6. Notice that the pronoun is represented as a discourse referent x3, which is equated with the discourse referent for a donkey (i.e., x2). Pronouns that haven't already been replaced by a discourse referent must be equated with some accessible discourse referent.</Paragraph> <Paragraph position="23"> Klein (1987) has augmented DRT to handle VPE by introducing a concept that is very similar to lambda abstraction. He is able to represent VPs and abstract the syntactic subject of the sentence. Consider how Example 3 is handled.</Paragraph> <Paragraph position="24"> The discourse representation for the two readings of the elided sentence are shown in Figure 7. Klein represents the trigger VP as a boxed structure named P. Within this box is a distinguished variable x2 (distinguished variables are marked with brackets), which corresponds to the abstracted subject. The trigger sentence is represented as P(xO), which is very similar to applying the discourse referent for the subject to a lambda function named P. The discourse referent for his is x4, which can either be equated with the distinguished discourse referent (i.e., x2) or with something outside of the VP box. To get the sloppy reading, it is equated with the distinguished discourse Discourse representations for the readings of Example 3.</Paragraph> <Paragraph position="25"> referent, as shown in Figure 7a. The elided sentence is represented initially as Q(xl), where xl is the discourse referent for the subject of the elided sentence. The sloppy reading for the sentence is provided when Q is equated with P.</Paragraph> <Paragraph position="26"> On the other hand, to derive the strict reading of the elided sentence, the discourse referent for the pronoun in the VP is equated with the discourse referent for the subject, namely xO (as shown in Figure 7b). Again the meaning of the elided sentence is derived by equating Q with P, but in this case the pronoun's discourse referent is equated with the discourse referent for the subject. Hence, Klein derives the two expected readings for the elided sentence in 3.</Paragraph> <Paragraph position="27"> This approach to VPE is similar to ours, except we introduce explicit differences between definite and indefinite NPs. To see why this is an issue, compare Examples 75 and 57.</Paragraph> <Paragraph position="28"> Because definites and indefinites are represented in the same way in DRT, the sloppy readings for both of these examples are represented in precisely the same way (shown in Figure 8a), as are the strict readings (shown in Figure 8b). However, the meaning of his friend is quite different from the meaning of a friend of his in VPE. The elided sentence in 75 cannot mean George saw a different friend of Fred's, in contrast to the elided Discourse representations for the readings of Examples 57 and 75.</Paragraph> <Paragraph position="29"> sentence in 57. One more interpretation is available for the elided sentence in 57 than for the elided sentence in 75, but in each case, only two representations are provided. In our approach, definites and indefinites are treated quite differently, not simply in how they are initially represented, but also in how they are processed to determine final meanings. It is therefore easier for us to explain the differences between examples 57 and 75. In Section 5.3, we demonstrated how our approach is able to provide the three readings for the elided sentence in 57. In contrast, our approach provides only two readings for the elided sentence in 75, as shown with a similar example (i.e., Example 3) in Section 4.4.</Paragraph> <Paragraph position="30"> In Klein's (1987) approach to VPE, the discourse referents for definites and indefinites are provided only after quantifier scoping information is available. Hence, Klein's discourse representation for these NPs is similar to our refined representations for definites and indefinites (i.e., definite functions following argument simplification and indefinite functions), where a discourse referent that is created in a box introduced by a universal corresponds to a function of a universal variable. However, Klein's approach does not determine whether lambda operators (i.e., distinguished variables) have scope over discourse referents in a VP box. To handle Examples 57 and 75 in the framework of DRT, Klein must determine whether a discourse referent should be defined inside or outside of the VP box, stipulating that a discourse referent is created inside a box if and only if the operator responsible for introducing that box has scope over the NP. In contrast to our approach, the information required to make this decision must be known before building the discourse model.</Paragraph> </Section> </Section> class="xml-element"></Paper>