Linguistic Design Principles
Design principles for DELPH-IN/MRS semantic representations identified during the Hankø meeting:
- Interface representation should include all information that is constrained by the grammar
- Interface representation should be highly normalized, abstracting away from details of surface syntax
- Close paraphrases should lead to comparable or identical structures in the interface representation
- Ex: Predicative copula does not contribute an EP (The fierce dog barks ~ The dog that is fierce barks)
- Ex: Dative alternation, passive (modulo discourse rel)
- Minimize ambiguity
- Corollary: Differentiate lexical predicates only if that distinction corresponds to morphosyntactic differences
- Decomposition is desirable, but not a goal in itself. If the predicates required for decomposition exist, then it is good.
cf. Copestake et al 2005 and "Slacker Semantics" (Copestake 2009)
Intrinsic Arguments on Adjectives and Adverbs
While the assumption of an intrinsic argument should be non-controversial for the predications associated with verbal and nominal phrases, more explanation is in order for extending this assumption to the predications introduced by adjectives and even adverbs.
One design decision for the ERG is that the same predicative copula be is used both for present participles as in the cat is sleeping and for predicative adjectives as in the cat is happy. Since we only want one EP for the verb phrase is sleeping, the copula be does not introduce an EP of its own, but rather constrains variable properties (tense and aspect) of the semantics introduced by its verbal complement. Hence predicative adjectives need to supply the eventuality that the copula will so constrain, and this variable is unsurprisingly the intrinsic argument of the EP introduced by the adjective.
A second design decision is that the same lexical entry is used both for predicative and attributive uses of adjectives, rather than either making distinct lexical entry pairs for most adjectives, or deriving the predicative entry from the attributive one via lexical rule. As a consequence, the EP introduced by this lexical entry has an intrinsic argument present even when the entry is used as an attributive adjective, where nothing else in the structure will constrain its variable properties. The grammar makes use of this ARG0 when the adjective itself is modified by a non-scopal word or phrase, as in extremely happy, with extremely introducing an EP which takes the ARG0 of the adjective it modifies as its ARG1. Without an intrinsic argument present in the EPs of attributive adjectives, all modifiers of adjectives would have to be treated as scopal, but this would create an unwanted asymmetry for consistently won and consistently happy, where we expect manner adverbs to be non-scopal modifiers of verbs.
A third design decision is that morphologically related adjective-adverb pairs introduce the same EP, with the syntactic category difference reflected in the semantics only in that the ARG1 of this predication will turn out to be an instance for adjectives but an eventuality for adverbs. Given the prior design decisions, even adverbs therefore have an intrinsic eventuality argument in their EPs, and this ARG0 is again employed by the ERG in composing the semantics of adverbial phrases containing a non-scopal modifier of the adverb, as in consistently victoriously. Here again, if victoriously lacked an intrinsic argument, then consistently would have to be treated as a scopal modifier taking the label of victoriously as its argument, leading to a similarly undesirable asymmetry between consistently won and consistently victoriously, assuming that the adverb modifying the verb here is non-scopal.
The presence of these intrinsic arguments in the predications for adverbs and attribute adjectives is thus motivated by the desires to minimize proliferation of lexical entries, to capture the morphological regularity of most adjective-adverb pairs in English, and to enable a consistent representation of non-scopal modifiers when they themselves modify modifiers.
The term ‘non-scopal modifiers’ includes all those which combine with their modifiee in such a way that quantifiers cannot take scope in between. This is a very broad phenomenon, including both modifiers of EPs with instance-type ARG0s and of EPs with event-type ARG0s. Among nominal (instance-type) non-scopal modifiers, we find that the semantic category includes subtypes that are syntactically quite distinct, viz., relative clauses (which involve a long-distance dependency) on the one hand, and non-relative clause modifiers (adjectives, PPs, etc.) on the other.
The key property of non-scopal modification is the sharing of labels between the modifier and modifiee. This in turn means that the two EPs will necessarily occupy the same node in the scope tree, and differentiates non-scopal modification from other predicate-argument relations. However, label sharing between two predications is not a sufficient condition to determine non-scopal modification: although it is rare in current ERG analyses, it is possible for a predicate to take a non-scopal, unquantified argument, as is the case, for example, in the 1214 analysis of a complex temporal expression like [at] four thirty:
h:numbered_hour(4)[ARG0 x, ARG1 i] h:card(30)[ARG0 i]
A similar configuration would arise if nominal arguments like pronouns or proper names were analyzed as unquantified (and the rules of MRS composition mandate the label sharing between the semantic head and its argument in such cases).
The shared scopal position (of non-scopal modifiers as well as non-scopal, unquantified arguments) is illustrated by the contrast between the following examples:
- All white cats are deaf.
- All deaf cats are white.
Though the ARG0 of _cat_n_1 is the ARG1 of both _deaf_a_1 and _white_a_1 in both of these examples, they do not mean the same thing. This, in turn, is because the modifier, but not the predicate, shares its label with _cat_n_1 and thus ends up in the same place in the scope tree (inside the restriction of the quantifier, as opposed to its body). A similar contrast can be built with relative clauses:
- All cats that are white are deaf.
- All cats that are deaf are white.
The shared scopal position between non-scopal modifiers and modified verbs can be observed with respect to quantifier scope. A modifier like loudly does not provide any additional slots for quantifiers to scope into, such that the following example is unambiguous:
- Every dog barked loudly.
This contrasts with scopal modifiers, which do interact with quantifiers, giving two readings in cases such as:
- Every dog probably barked.
Non-scopal modification is not a semantic phenomenon in the practical sense that we use in this documentation, i.e. there is no set of semantic fingerprints that uniquely picks out sentences with this property. Rather, non-scopal modification is a property of several phenomena discussed in this documentation, including at least:
Apposition (pending reanalysis in terms of ICONS)
Compounding: ‘compound’, ‘compound_name’, including titles and N-ed (J-N_CRD-T_C cuts across compounding and coordination)
Implicit Locatives: (‘today’, ‘every time he arrives’): ‘loc_nonsp’
Instrumental Relatives: ‘with_p_rel’
- Extraposed relative clauses (create page for these?)
Some previous descriptions of ERS have used the term ‘intersective modifiers’ for this phenomenon, but here we prefer ‘non-scopal modifiers’ since we give the same treatment in ERS to intersective, subsective and privative adjectives (Kamp & Partee 1993). At the level of MRS representations, we argue that it is appropriate to treat these equivalently, as they all share the property of disallowing intervening quantifiers.
There is a semantic distinction within non-scopal modifiers that cross-cuts the relative clause/other modifier distinction, namely, whether or not the head being modified plays a role in the predication sharing its label. For example in The dog Abrams bought barked, _buy_v_1 shares its label with _dog_n_1 and the ARG0 of _dog_n_1 is also the ARG2 of _buy_v_1. However, in The dog Kim told Abrams to buy barked, _dog_n_1, despite sharing its label with _tell_v_1 has no role to play in that predication. For non-relative clause modifiers of nouns, we find an analogous case with the so-called tough adjectives, as in Abrams is an easy editor to impress, where _easy_a_for shares a label with _editor_n_1 but does not take its ARG0 as an argument.
Note, too, that certain elements can be syntactic complements but semantic modifiers. This happens, for example, in the analysis of adverbial-complement taking verbs such as (certain senses of) behave, do and go, as well as verbs of motion taking directional complements:
- Kim behaved badly.
- Kim did well.
- It went well.
- Kim walked to the store.
Similarly, degree specifiers such as very in Kim is very happy are syntactically combined as specifiers (not modifiers) but appear in the ERS as non-scopal modifiers.
Kamp, H., & Partee, B. (1995). Prototype theory and compositionality. Cognition, 57(2), 129-191.