Projecting Grammatical Features in Nominals: Computational Implementation

The computational implementation of a linguistic theory of the representation and projection of grammatical features in nominals is described. The processing of nominals is part of a larger model of language comprehension implemented as a computational cognitive model within the ACT-R cognitive architecture. The model combines a serial, pseudo-deterministic processing mechanism for building linguistic representations—implemented within ACT-R’s production system—with a parallel, activation and selection mechanism for choosing between alternatives—implemented as an interaction between ACT-R’s procedural (production) and declarative memory (DM) systems. 1. Linguistic Theory The underlying linguistic theory is an adaptation of X-Bar Theory (Chomsky, 1970; Jackendoff, 1977) called BiPolar Theory (Ball, 2007). In Bi-Polar Theory, there are four primary phrase internal grammatical functions: head, specifier, complement, and modifier. With respect to nominals or noun phrases (NPs), the typical head is a noun like “pilot” and the typical specifier is a determiner like “the” as in “the pilot”. We reject the functional head hypothesis (Abney, 1987) which treats “the” as the head and “pilot” as a complement. The specifier and head—the most basic elements of a nominal—constitute the two poles of Bi-Polar Theory. At a minimum, a nominal will contain a specifier, a head, or both. The typical modifier— which is not required—is either an adjective like “old” which occurs between the specifier and head as in “the old pilot” or a prepositional phrase like “in the airplane” which occurs after the head as in “the pilot in the airplane”. There are few true complements in nominals and they will not be considered in this paper. We prefer the term nominal (or object referring expression) to NP, since the head of a nominal is not necessarily a noun—the head may be empty (e.g. “the red” in “I like the red” in reference to a red object) or it may contain a word or phrase that is not a noun (e.g. “running” in “the running of the bull” or “giving to the poor” in “his giving to the poor is nice”). It is a key claim of this research that words and phrases functioning as specifiers and modifiers—in addition to heads—may project grammatical features to encompassing nominals. Grammatical features may be redundantly encoded in words and phrases fulfilling different grammatical functions. At the level of the nominal, the projected grammatical features are collected into a set without duplicates. Redundantly encoded grammatical features may occasionally conflict or a grammatical feature may be unspecified—without the expression being ungrammatical—necessitating mechanisms for handling conflicts and accommodating unspecified features. The primary grammatical features include definiteness, number, animacy, gender, person and case. The definiteness feature is most closely associated with determiners like “the” and “a”, demonstrative pronouns like “this” and “that” and quantifiers like “all” and “some”. There are at least three possible values: universal (e.g. “all” in “all books”), definite (e.g. “the” in “the book”) and indefinite (e.g. “a” in “a book”), and perhaps negative “e.g. “no books”). The number, animacy and gender features are most closely associated with nouns. The possible values for number are singular, mass (a subtype of singular) and plural. The possible values for animacy are human (a subtype of animate), animate and inanimate. The possible values for gender are male and female. There is no neuter gender in English. With a few exceptions, only human (or animate) nouns are encoded for gender. Plural and mass nouns, but not singular count nouns, are also indefinite. For example, the singular count noun “man” is singular, human and male; the plural count noun “rocks” is indefinite, plural and inanimate; and the singular mass noun “rice” is indefinite, singular and inanimate. To be grammatical, a nominal normally requires an indication of definiteness, typically provided by the specifier, and an indication of number, typically provided by the head. For example, in “the book”, “the” is definite and “book” is singular. Since pronouns, proper nouns, and plural and mass nouns also provide an indication of definiteness, they can occur alone as nominals (e.g. “he” is definite and singular, “John” is definite and singular, “books” is indefinite and plural as in “books are fun to read”). On the other hand, singular count nouns do not provide an indication of definiteness and do not normally occur alone in nominals (e.g. “*book is fun to read”). The grammatical features person and case are only associated with a small number of personal, possessive and reflexive pronouns (e.g., “I” is first person, subjective case; “me” is first person, objective case; “he” is third person subjective case; “him” is third person, objective case). All reflexive pronouns are objective case (e.g. “myself” is first person objective, “himself” is third person, objective) and all possessive pronouns are genitive case (e.g. “my” is first person, genitive, “hers” is third person, genitive). In alignment with basic principles of Cognitive Linguistics (cf. Langacker, 1987, 1991), we assume that grammatical features are semantically motivated and not purely syntactic. We also assume that grammatical features are identified and used by humans during language comprehension. That is, humans recognize the different categories of grammatical features even if they cannot provide explicit names for the categories. A key aspect of language comprehension is determining the referents of nominals. The set of grammatical features projected to the nominal provides the grammatical basis for determining the referent, and is especially important for determining co-reference. For example, given the input “The man kicked the ball. She ran to first base.” the nominal “the man” indicates that an object of type man is being referred to that is somehow salient in the context of the utterance. This salience is indicated by the definite feature of “the”. Likewise for “the ball”. On the other hand the occurrence of “she” is problematic. Pronouns normally indicate co-reference to a previously introduced referent. However, the female gender of “she” is inconsistent with the male gender of “the man” and the human animacy of “she” is inconsistent with the inanimate feature of “the ball”. There is no previously mentioned referent to which the pronoun can co-refer. Besides their importance for reference determination, grammatical features facilitate language comprehension in other ways. For example, interpreting the classic “flying planes are dangerous” vs. “flying planes is dangerous” depends on number agreement between the subject “flying planes” and the auxiliary verb “is” vs “are” with “flying planes” being ambiguous between a reading in which the head “planes” projects the feature plural, and a reading in which the head “flying” leads to construal of the expression as singular. Likewise, determining the meaning of “the book I gave the man” and “the man I gave the book” hinges on the animacy of “book” and “man”, interacting with the ditransitive verb “give” which prefers an animate indirect object. Although grammatical features can be extremely useful for language comprehension, they are only useful to the extent that there is grammatical evidence that they exist. It makes little sense to treat common nouns as having case or person features since there is no grammatical evidence (i.e. marking) for these features in English. For example, “the man” can occur as the subject or object as in “the man kicked the ball” and “the horse kicked the man”. Including a case feature for common nouns simply introduces an ambiguity that must be resolved by the context in which the noun occurs—the noun itself provides no such indication. With respect to person, all common nouns could be treated as third person by analogy with third person pronouns which are grammatically distinct, coupled with claims that subject-verb agreement in English is based on both number and person. However, Ball (submitted) argues that subject-verb agreement in English is based strictly on number, with the exception of the first person pronoun “I” and present tense verbs (e.g. “I am hungry”), making a third person feature for common nouns grammatically unnecessary. We adhere to the basic principle that where there is no grammatical distinction, there is no grammatical feature. Without grammatical evidence, there is simply no basis for learners of English to learn the feature. Although most pronouns are marked for case and person in English, common nouns are not. Insisting that all nouns have case and person features to capture a (universal) generalization over nouns and pronouns, is counter-productive—the grammatical generalization introduces unnecessary ambiguity which does not facilitate comprehension. The search for ever more abstract linguistic generalizations in generative grammar does not facilitate progress in building language comprehension systems. Knowledge of language involves representations or constructions at multiple levels of abstraction, with the most specific constructions that match a given linguistic input carrying most of the weight for language comprehension. 2. Psycholinguistic Theory There is extensive psycholinguistic evidence that human language processing is essentially incremental and interactive (Gibson & Pearlmutter, 1998; Altmann, 1998; Tanenhaus et al., 1995; Altmann & Steedman, 1988). Garden-path effects, although infrequent, strongly suggest that processing is essentially serial at the level of phrasal and clausal analysis (Bever, 1970). Lower level processes of word recognition suggest parallel, activation-based processing mechanisms (McClelland & Rumelhart, 1981; Paap et al., 1982). At the level of phrasal and clausal analysis, humans appear to deterministically pursue a single analysis which is only occasionally disrupted, requiring reanalysis. One of the great challenges of psycholinguistic research is to explain how humans can process language effortlessly and accurately given the complexity and ambiguity that is attested (Crocker, 2005). As Boden (2006, p. 407) notes, deterministic processing “would explain the introspective ease and speed of speech understanding”, but a purely deterministic, incremental processing mechanism would more frequently make incorrect local choices requiring reanalysis than is evident in human language processing. Marcus (1980) proposed a lookahead mechanism to improve the performance of a deterministic, yet monotonic, processor, bringing it into closer alignment with human performance. However, there is considerable evidence that humans immediately determine the meaning of linguistic inputs (cf. Tanenhaus et al., 1995; Altmann & Mirkovic, 2009) which is inconsistent with extensive lookahead, delay or underspecification—the primary serial and monotonic mechanisms for dealing with ambiguity. As Altmann & Mirkovic (2009, p. 605) note “The view we are left with is a comprehension system that is „maximally incremental‟; it develops the fullest interpretation of a sentence fragment at each moment of the fragment‟s unfolding”. Not only is there not extensive lookahead, delay or underspecification, the human language processor engages in “thinkahead”, predicting what will come next rather than waiting until the succeeding input is available before deciding on the current input. To capture the essentially incremental nature of human language processing, we adopt a serial, pseudo-deterministic processor that builds linguistic representations by integrating compatible elements, relying on a non-monotonic mechanism of context accommodation to handle cases where some incompatibility that complicates integration manifests itself. Context accommodation makes use of the full context to make modest adjustments to the evolving representation or to construe the current input in a way that allows for its integration into the representation. Context accommodation need not be computationally expensive (i.e., a single production may effect the accommodation, just as a single production may effect integration without accommodation). In this respect, context accommodation is not a reanalysis mechanism that disrupts normal processing; rather, it is part and parcel of normal processing. Reanalysis mechanisms need only kick in when context accommodation fails and larger adjustment is needed. Further, as will be shown below, context accommodation can give the appearance of parallel processing in a serial processing mechanism, blurring the distinction between serial and parallel processing. The mechanism of context accommodation is most closely related to the limited repair parsing of Lewis (1998). Context accommodation may be viewed as a very modest form of repair, but since context accommodation need not take any more time than normal processing, the use of the term repair is somewhat inappropriate. According to Lewis (1998, p. 262) “The putative theoretical advantage of repair parsers depends in large part on finding simple candidate repair operations”. The mechanism of context accommodation provides evidence for this theoretical advantage. To capture the essentially interactive nature of human language processing, we propose a probabilistic, contextsensitive mechanism for activating alternatives in parallel and selecting the most highly activated alternative. This parallel, probabilistic mechanism selects between competing alternatives, but does not build any structure—building structure is the function of the incremental integration mechanism. At each choice point, the parallel, probabilistic mechanism uses all available information to activate and select the preferred alternative, and the serial, pseudodeterministic mechanism integrates the preferred alternative into the evolving representation. Use of the full local context supports selection of alternatives that are likely to be correct, allowing the serial integration mechanism to be largely deterministic. However, the local context is not always consistent with the global context and locally preferred choices sometimes turn out to be globally dispreferred. The mechanism of context accommodation allows the processor to adjust the evolving representation to accommodate the subsequent context, without lookahead, backtracking or reanalysis. Only when the context accommodation mechanism breaks down do more disruptive reanalysis processes become necessary. The use of the term pseudodeterministic to describe the basic processing mechanism reflects the integration of parallel, probabilistic activation and selection mechanisms and context accommodation with what is otherwise a serial, deterministic processor.