Sentence-level constructions: A demonstration in support of Ambridge (2020)

Abstract

Ambridge reviews and augments an impressive body of research demonstrating both the advantages and the necessity of an exemplar-based model of knowledge of one’s language. He cites three computational models that have been applied successfully to issues of phonology and morphology. Focusing on Ambridge’s discussion of sentence-level constructions, this commentary cites additional research in support of his exemplar hypothesis. It then provides an informal demonstration of how Skousen’s (1989) Analogical Model might be extended to the processing of sentence-level constructions.

Keywords

Analogical modeling construction grammar exemplar linguistics instance-based syntax

In his target article, Ambridge (2020) argues that novel instances of linguistic usage are produced and comprehended ‘on the fly’ by analogy with stored exemplars of previous usage accumulated in one’s long-term memory rather than by the application of resident generalizations that have been abstracted away from one’s experience with a language. In support of this thesis, Ambridge reviews an impressive body of research which demonstrates how both generative rule-based models and network abstraction (i.e., connectionist) models of linguistic representation are empirically inadequate and why any adequate model of abstract representations must include access to specific exemplars of previous usage. He further argues that such reference to remembered exemplars alone is sufficient to account for the linguistic behaviors said to motivate those theoretical abstractions in the first place.

As Ambridge summarizes in his section on inflectional morphology, there are today at least three computational models of exemplar approaches that have been applied with considerable success to issues of phonological, morphological, and lexical usage, Nosofsky’s Generalized Context Model (Nosofsky, 1990), Daelemans’s Tilburg Memory Based Learning Model (Daelemans et al., 2004), and Skousen’s Analogical Model (Skousen, 1989). These models have allowed linguists to test the claims and implications of exemplar approaches explicitly against both observationally and experimentally obtained data, often achieving 90–95% accuracy (as reported in studies by Chandler, 2010; Eddington, 2002). For a variety of both theoretical and practical reasons, however, these models have not yet been applied to issues of syntax such as those discussed by Ambridge in his section on sentence-level constructions. Nonetheless, as Ambridge notes, the performance of the models must capture and account for our intuitions and behaviors that indicate that an often invisible – in the sense that there often are no overt structural markers – yet clearly psychologically real constituent structure is assigned to the strings of words that we produce and comprehend. This creates a conundrum because those psychologically real constituent structures are abstract in the sense just noted, yet we argue that they do not arise through the operation of any resident abstract syntactic generalizations.

Within Cognitive Linguistics today, construction grammars are the predominate approach to modeling linguistic structure (Hoffmann, 2017). At their highest level of abstraction, constructions are represented as prototype-like structural generalizations said to arise through a process of abstraction away from repeated experiences with specific exemplars of usage. As I have argued elsewhere, however (Chandler, 2017), and as Ambridge elaborates on in his article, positing such prototype-based constructions is empirically inadequate as a model of cognitive representations. Experimental data such as those derived by Stefanowitsch and Gries (see Gries, 2011), for example, provide additional strong evidence that people actually do retain memories for the specific exemplars from which such constructions are said to arise.

In addition to the evidence reviewed by Ambridge, exemplar-based models of syntactic processing must also accommodate at least four additional sets of facts about the online processing of sentences. First, Swinney (1979) reported evidence that as soon as the brain recognizes a word it begins to activate all of the possible meanings that it has associated with that word and that only after a short time lapse does it begin to settle on the meaning consistent with the developing interpretation of the input sentence. Second, Gahl et al. (2004) have demonstrated that the brain’s anticipation of what possible alternative structures may follow a given verb reflects closely the relative frequency of those alternatives in large corpora of actual usages. For example, the verb remember is followed by a simple direct object noun phrase about 53% of the time and by a that-complement about 25% of the time. A third line of facts, discussed in Elman (2009), provides additional strong evidence that knowledge of specific events also plays a significant role in sentence interpretation beginning early in the sentence. For example, reading The surgeon cuts . . . begins construction of a mental image and linguistic expectations that are very different from those triggered by The lumberjack cuts . . . . Finally, Sato et al. (2013) showed that the developing mental representations for those evolving interpretations change incrementally during sentence processing and that in a verb-final language such as Japanese an unexpected final verb can force an immediate change in the overall mental image. All of these facts imply an exemplar-based process driving the dynamics of sentence comprehension.

None of the three exemplar models cited above has yet been applied to the incremental syntactic interpretation of word strings. Nonetheless, Skousen’s Analogical Model (AM) (Skousen, 1989) seems especially well suited to accommodating the facts summarized above. Its central component compares supracontexts (subsets of the word strings) in the input sentence with the matching supracontexts in each remembered exemplar. As each new word is recognized, the revised string is reinterpreted using AM’s algorithm (a process which, by the way, accommodates the n-gram phenomena also discussed by Ambridge). The evolving syntactic structure of the new sentence emerges on the fly as it is compared to the previously interpreted exemplars. For example, applying AM incrementally to the ambiguous sentence They fed her dog biscuits and using the Corpus of Contemporary American English (COCA) (Davies, 2008) as the data set of exemplars yields the following informal analyses: They fed . . . is interpreted as a subject-verb construction 100% of the time and anticipates an intransitive continuation 21% of the time, a ditransitive continuation 29% of the time, and a direct-object continuation 40% of the time. They fed her . . . changes those predictions to 67% direct-object, and 33% ditransitive. Given They fed her dog . . ., her dog alone is interpreted as possessive + N 100%, but the overall string is predicted to be direct-object 60% of the time and ditransitive 40%. Given the full sentence, the string dog biscuits appears in COCA only as a compound noun, but applying AM to the overall string predicts the ditransitive reading 100% of the time with her [dog biscuits] 4% of the time and [her dog] biscuits 95%. There are still many unanswered questions about this approach, but it appears that we shall also soon be able to test this aspect of Ambridge’s hypothesis empirically.

Footnotes

ORCID iD

Steve Chandler

References

Ambridge

(2020). Against stored abstractions: A radical exemplar model of language acquisition. First Language 40(5-6): 509–559.

Chandler

(2010). The English past tense: Analogy redux. Cognitive Linguistics, 21(3), 371–417. https://doi.org/10.1515/cogl.2010.014

Chandler

. (2017). The analogical modeling of linguistic categories. Language and Cognition, 9(1), 52–87. https://doi.org/10.1017/langcog.2015.24

Daelemans

Zavrel

van den Bosch

(2004). TiMBL: Tilburg memory-based learner, version 5.1, reference guide (ILK Technical Report 04-02). ILK Research Group, Tilburg University.

Davies

(2008). The Corpus of Contemporary American English (COCA). https://www.english-corpora.org/coca/

Eddington

(2002). A comparison of two analogical models, Tilburg memory-based learner versus analogical modeling. In Skousen

Lonsdale

Parkinson

D. B.

(Eds.), Analogical modeling, an exemplar-based approach to language (pp. 141–155). John Benjamins Publishing Co.

Elman

J. L.

(2009). On the meaning of words and dinosaur bones: Lexical knowledge without a lexicon. Cognitive Science, 33, 547–582. https://doi.org/10.1111/j.1551-6709.01023.x

Gahl

Jurapsky

Roland

(2004). Verb subcategorization frequencies: American English corpus data, methodological studies, and cross-corpus comparisons. Behavior Research Methods, Instruments, & Computers, 36(3), 432–443.

Gries

S. T.

(2011). Data in construction grammar. In Hoffmann

Trousdale

(Eds.), The Oxford handbook of construction grammar (pp. 93–110). Oxford University Press.

10.

Hoffmann

(2017). Construction grammars. In Dancygier

(Ed.), The Cambridge handbook of cognitive linguistics (pp. 310–329). Cambridge University Press.

11.

Nosofsky

R. M.

(1990). Relations between exemplar similarity and likelihood models of classification. Journal of Mathematical Psychology, 34, 393–418.

12.

Sato

Schafer

A. J.

Bergen

B. K.

(2013). One word at a time: Mental representations of object shape change incrementally during sentence processing. Language and Cognition, 54(4), 345–373. https://doi.org/10.1515/langcog-2013-0022

13.

Skousen

(1989). Analogical modeling of language. Springer.

14.

Swinney

(1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 15, 545–569.