Of Minds and Language (52 page)

c. Sentence behavior is instant and “horizontal” – speakers believe that they comprehend and produce meaningful sentences simultaneously with their serial input or output. Comprehension does not start only at the end of each sentence: production does not wait until a sentence is entirely formulated.

These three observations set a conundrum:

(5) a. Sentence processing involves computation of syntax with whole sentences as domain – it is
vertical
.

b. Language behavior proceeds serially and incrementally– it is horizontal.

Recently, Dave Townsend and I rehabilitated the classic comprehension model of Analysis by Synthesis (A×S) that provides a solution to the conundrum (following Halle and Stevens 1962, Townsend and Bever 2001). On this view, people understand everything twice: once based on the perceptual templates; once by the assignment of syntactic derivations. In the A×S architecture the two processes are almost simultaneous. First, the perceptual templates assign likely interpretations to sentences, using a pattern completion system in which initial parts of a serial string automatically trigger a complete template. Typical templates of this kind are:

(6) a. Det … X → np[Det…N]np

b. NP V(agreeing with NP) (optional NP) ! Agent/Experiencer Predicate (object/adjunct)

Second, the initially assigned potential meaning triggers (and constrains) a syntactic derivation. The two ways of accessing meaning and structure converge, roughly at the ends of major syntactic units. That is, as we put it,
we understand everything twice
. The model has several unusual features (Town-send and Bever 2001). First, the model assigns a complete correct syntax
after
accessing an initial meaning representation. Second, that meaning is sometimes
developed from an incorrect syntactic analysis. For example, syntactic passives (7a) are initially understood via the variant of the canonical sentence template in (6b) that applies correctly to lexical passives (7b); raising constructions (7c) are understood initially via the same kind of misanalysis.

(7) a. Syntactic passive: Bill was hit

b. Lexical passive: Bill was hurt

c. Raising: Bill seemed happy

d. Control: Bill became happy

The schema in (6b) initially misassigns “hit” as an adjective within a predicate phrase. That analysis is sufficient to access semantic information modeled on the interpretation template for lexical passive adjectives – a syntactic misanalysis. This analysis is then corrected by accessing the correct derivation. This sequence of operations also explains the fact that the experimental evidence for the trace appears in syntactic passives and raising constructions only after a short time has passed (Bever and McElree 1988, Bever et al. 1990, Bever and Sanz 1997). This model also explains a number of simple and well-known facts. Consider the following examples:

(8) a. The horse raced past the barn fell

b. More people have been to Russia than I have

Each of these cases exemplifies a different aspect of the A×S model. The first reflects the power of the canonical form strategy in English (6b), which initially treats the first six words as a separate sentence (Bever 1970). Native speakers judge this sentence as ungrammatical, often even after they see parallel sentences with transparent structure:

(9) a. The horse ridden past the barn fell

b. The horse that was raced past the barn fell

c. The horse racing past the barn fell

The example is pernicious because recovering from the misanalysis is itself vexed: the correct analysis includes the garden-pathing proposition that “the horse raced” (i.e., was caused to race): thus, every time the comprehender arrives at the correct interpretation she is led back up the garden path.

Example (8b) (due to Mario Montalbetti) is the obverse of the first example. The comprehender thinks at first that the sentence is coherent and meaningful, and then realizes that it does not have a correct syntactic analysis. The initial perceptual organization assigns it a schema based on a general comparative template of two canonical sentence forms – “more X than Y,” reinforced by the apparent parallel Verb Phrase structure in X and Y (“ …have been to
Russia … I have”). On the A×S model, this superficial initial analysis triggers the derivational parse system, which ultimately fails to generate a derivation.

I do not expect to have convinced the reader of our model via such simplified examples: in our book, we organize a range of often surprising experimental and neurological facts supporting an early stage of comprehension resting on frequent statistically valid templates, followed by a structurally correct syntactic derivation (Townsend and Bever 2001,
Chapters 5
–
8
; see Friederici, this volume, for imaging data consistent with this bi-phasic model of comprehension).

This model requires languages to have certain universal features. Most important is the otherwise unmotivated fact that actual languages have a characteristic set of statistically grounded structural patterns at each level of representation (phonological, morpho-lexical, syntactic). It further requires that complex constructions be functionally homonymous with simpler constructions in ways that allow the simpler constructional analysis to convey the more complex meaning at an initial pre-derivational stage of processing. The model is inelegant in that it solves the conundrum (5) by fiat – sentence processing is both fast and complex because it is simultaneously handled by two systems, one fast and sometimes wrong, one slower but ultimately correct. This is an inelegant solution to the conundrum, but shows that humans may solve it, albeit inelegantly.

Two historically competing principles about the mind have alternately dominated the cognitive sciences:

(10) a. Everything we do is based on habits.

b. Everything (important) we do is based on creative rules.

The A×S model architecture shows how the two insights might be integrated together in adult behavior. A corresponding model holds for the acquisition of language. On that model, the child alternates (logically) between formulating statistical generalizations about the language, and assembling derivational operations that account for those generalizations. Many researchers are demonstrating that child-directed speech in fact has statistical regularities that might guide the infant and child towards language (e.g., Curtin et al. 2005, for segmentation, Golinkoff 1975; Brent 1997; Cartwright and Brent 1997; Gerken 1996; Golinkoff et al. 2005; Mintz 2002, 2003, 2006; Redington et al. 1998). At the same time, infants are quite good at detecting statistical patterns from serial strings with specific kinds of structure (Gomez and Gerken 1999;
Marcus et al. 1999; Saffran 2001, 2003; Saffran et al. 1996); older children also show statistical sensitivity in developing grammatical and lexical ability (Bates and MacWhinney 1987, Gillette et al. 1999, Moerk 2000, Naigles and Hoff-Ginsburg 1998, Yang 2006). If one component of syntax acquisition is the compilation of relevant generalizations, this model requires that the child be presented with some statistical regularities in the language he hears. This requirement explains several computationally eccentric facts about attested languages:

(11) a. Each language has a canonical surface form: in English this is schematically as presented in the left side of the expression in (6b).

b. Statistically, the canonical form has a dominant assignment of seman tic relations: in English this is the template we found explanatory for much adult language behavior (6b).

c. The canonical semantic interpretation is violated in a set of minority constructions: in English, this includes passives, raising, unaccusatives, middle constructions.

d. The minority constructions that violate the form can nonetheless be approximately correctly interpreted by application of the canonical form interpretation. (This is exemplified in the initial stages of com prehending syntactic passives and raising constructions, discussed above in examples (7).)

None of these properties follows from the computational architecture of any of the last fifty years of generative grammar. Yet they are characteristic of attested languages. In English, the first property has been noted as the result of rule “conspiracies,” which guarantee that sentences have the same surface form regardless of their thematic relations and derivation. The vast majority of sentences and clauses have a canonical form with a subject preceding a correspondingly inflected verb:

(12) a. The boy hits the ball

b. The ball was hit by the boy

c. It is the boy who hits the ball

d. The boy was happy

e. The boy seems happy

f. The boy was eager to push

g. The boy was easy to push

h. It was easy to push the boy

i. The boy pushes easily

j. Who pushed the boy?

k. Who does the boy push?

The notion of such conspiracies is not novel, be it in syntax or phonology (cf. Ross 1972, 1973a, b). In traditional derivational terms, it reflects constraints on derivations such that they have the same general surface form regardless of differences in logical form or semantic relations. This is despite the fact that each underlying form could be reflected in a unique surface sequence or signaled by a specific marker. On our interpretation, such computationally possible languages would be allowed by generative architectures, but are not learnable: they would make it hard for the language-learning child to develop a statistically based pattern that it can internalize and use for further stages of acquisition.

The canonical form (11a) facilitates the discovery of a surface template based on statistical dominance of the pattern. The semantic schema (11b) above relates the canonical form to a standard interpretation – although a majority of individual constructions may not conform to that schema, the vast majority of actual utterances in corpora do so – another fact about languages unexplained by generative architectures. The third fact (11c) – that some cases violate the canonical semantic interpretation of the canonical form – is particularly important if the child is eventually to discover that there are actual derivations in which a given surface form expresses different patterns of thematic relations. Finally, the interpretability of the schema-violating constructions via misanalysis and homonymy with simpler constructions (11d) – contributes to the child's ability to interpret sentence types for which it does not yet have a syntactic analysis. I summarize the set of these conditions as the “Canonical Form Constraint (CFC).”

The CFC suggests a way in which the child can transcend the “poverty of the stimulus.” First, the child can create and then analyze his own set of form/ meaning pairs going beyond the actual sentences he hears, based on these generalizations. Second, this solves an important problem for any learning scheme – how do children remember and understand sentences for which they do not yet have a correct syntactic analysis? (Valian 1999). It would not work for the child to maintain a list of grammatically unresolved sentences: any given list is heterogenous without prior structural ordering. The A×S model suggests that children can rely on statistical patterns and occasional false analyses to generate an internal bank of meaning/form pairs and maintain an internalized data bank to evaluate candidate derivational analyses. This reduces the need for children to access positive and negative feedback as guides to their emerging syntactic abilities. On this view, the child can attempt derivation of a construction based on a subset of sentences of a given general pattern, and then “test” the derivational structure on other sentences of a similar pattern. (For related ideas, see Chouinard and Clark 2003, Dale and Christiansen 2004,
Golinkoff et al. 2005, Lieven 1994, Moerk 2000, Morgan et al. 1995, Saxton 1997, Valian 1999). These facts and considerations offer an explanation for the CFC – peculiar in the sense that the computational architecture of syntax does not in itself require the CFC. It is reflected in attested languages because it makes them learnable, using a hypothesis formation procedure.

If this picture is correct, children should show evidence of actually learning perceptual strategies, based on statistical frequency of preponderant features of their surrounding language. We and others have found evidence supporting this (Bever 1970, Maratsos 1974, Slobin and Bever 1982). The original finding was based on having children act out simple sentences with puppets. (Typical data are summarized in Table 18.1). 2-year-old children use a simple strategy that focuses primarily on the exact sequence NounPhrase + Verb, interpreting that as Agent + Verb. Thus, at age 2, children interpret declarative and object cleft sentences, along with semantically unlikely sentences, above chance: in these constructions, the noun immediately before the verb is the agent. By age 3–4, they rely both on a more elaborated analysis of word order and semantic strategies:

(13) a. #N… = Agent

b. Animate nouns are agents, inanimate nouns are patients

(13a) represents a shift from assigning the noun immediately before the verb as agent, to assigning the first noun in the overall sequence as agent. This produces correct performance on simple declarative sentences, but a decrease in performance on sentence types in which the first noun phrase is not the agent (object clefts and passives).

The emergence of the two kinds of strategies accounts for the decrease in performance on semantically reversible sentences that violate the CFC
(the giraffe was kicked by the dog). The emergence of reliance on semantic information accounts for the increase in performance on sensible sentences (the dog ate the cookie), and the decrease in performance on semantically odd sentences (the cookie ate the dog). The reliance on semantic factors at age 4 also can override the word-order strategy, leading to correct performance on irreversible passives (the cookie got eaten by the dog).

Table 18.1 Percentage correct interpretations of simple sentences by children
^a