Constructions Generalize over Lexical Strings

Cognitive linguistics in general and cognitive grammar in particular are closely re-lated to usage-based linguistics. Proponents of a usage-based approach to language argue that “structure emerges from usage, is immanent in usage, and is influenced by usage on an ongoing basis” (Langacker, 2010, p. 109). Language structure is seen as developing from experiences of language use in a gradual and piecemeal process.

Put differently, constructions, in particular, grammatical constructions emerge as a by-product of cross-modal cognitive abilities and processes working on structured linguistic input. In line with construction grammar, the outcome of the process is commonly described as a network of overlapping and interrelated constructions (Diessel, 2015, 2018; Langacker, 2000). Originally introduced by Langacker (1988), a usage-based model has been adopted in research on first language acquisition (Diessel, 2004, 2013, 2017; Tomasello, 2003, 2009), second language acquisition (De Bot, Lowie, & Verspoor, 2007; De Bot & Larsen-Freeman, 2011; N. C. Ellis, 2008;

N. C. Ellis & Cadierno, 2009; Verspoor & Behrens, 2011), language change (Bybee, 1998; Bybee & Hopper, 2001), and language evolution (MacWhinney, 2008), and is closely related to functionalist models of language comprehension and acquisition (E.

Bates & MacWhinney, 1989; MacWhinney, 2012), emergentist models of language (MacWhinney, 2015a; O’Grady, 2008), and research on learning in connectionist networks (Elman, 1993). From a usage-based view, first and second language learn-ers detect patterns in their linguistic input, memorize and schematize them, reuse them for processing similar input and producing novel output, adapt them to novel input, and forget patterns which have fallen out of use; on a historical time scale, members of language communities adhere to conventionalized patterns, streamline them in an erosionlike process, and slowly drift towards novel patterns; on an evolu-tionary time scale, the evolution and conspiracy of a range of nonlinguistic cognitive capacities is the prerequisite for the development of human language; in language processing, language users interpret incoming strings based on distribution-weighted cues; learning computational models adjust connection weights to their input to op-timize their predictions; and so on. Generally speaking, usage-based researchers in the different fields share the assumption that the respective evolving systems (e.g., human language learners, computational models, a language, Homo sapiens) de-velop linguistic patterns in response to structured linguistic input, environmental conditions, and functional constraints, recruit acquired patterns to deal with similar situations, and modify their patterns to meet the requirements of novel situations

2.2. USAGE-BASED LINGUISTICS 33 (Beckner et al., 2009). On this view, then, language learning, language change, and language evolution are inseparable from language use, they are essentially the effects of language use accumulating on different time scales (MacWhinney, 2005, 2014).

In the following, the usage-based model is outlined in more detail, with a focus on language learning, in particular, second language learning.

In a usage-based framework, grammatical constructions emerge as generaliza-tions over lexical strings which instantiate them. They are memory representa-tions of the commonalities of their instantiarepresenta-tions which are reinforced across usage events. While recurring commonalities become increasingly entrenched with each encounter, fine-grained differences between lexical strings are not reinforced and thus not stored in memory in the long run. As a consequence, the emerging con-structions are partially or wholly schematic and range from lexically specific strings to highly schematic templates, depending on the lexical variability of their instanti-ations (Langacker, 2010).

By way of illustration, consider once more the comparative correlative construc-tion The X-er the Y-er (6c). A sample of instantiaconstruc-tions of the construcconstruc-tion is given in Examples 11 and 12.

Example 11 (COCA)

a. Big Data, the bigger the better

b. “Will she sleep now?” “The longer the better,”

c. Virgins are prized, and the younger the better.

d. Dive-bar chicken wings – the spicier the better.

e. I love weird things, the uglier the better f. But most emphatically not the redder the hotter.

Example 12 (COCA)

a. The more information people have the better.

b. The harder they pulled the louder it became.

c. the more warriors she engulfs, the hungrier she gets

Upon hearing Examples 11a through 11d, a learner memorizes the parts which invariably recur across the lexical strings but generalizes over variant parts. As a consequence, a partially item-specific, partially schematic construction is acquired:

the Adj-er the better. The invariable parts of the strings – the, -er, and the better – are reinforced with each encounter and thus memorized “as is.” In contrast, fine-grained differences between the strings are not reinforced. The different adjectives

encountered at the position following the initial the in the string – big, long, young, and spicy – are therefore not stored in memory in detail. Only what is common to them on a lower level of specificity is reinforced, here, the occurrence of adjectives forming regular comparatives, which results in a schematic position Adj.

Once acquired, a construction is exploited to improve subsequent processing and to produce similar strings. The comparative correlative construction which has been acquired on the basis of Examples 11a through 11d contains a schematic po-sition Adj which may be filled with any adjective forming a regular comparative, for example, ugly, producing a novel string which was not part of the prior input (11e). Moreover, a construction continues to be modified by usage. Recognizing and processing the novel string in Example 11f as an instantiation of the comparative correlative construction forces the learner to expand the construction to cover ad-jectives other than better at the position following the second the. As a result, the position gains in schematicity: the Adj-er the Adj-er. The more item-specific con-struction the Adj-er the better is stored at a lower level of schematicity and remains intact. When encountering more varied strings, as in Example 12, the construction is expanded to comprise an even wider range of lexical strings and as a result be-comes more schematic. The resulting the X-er the Y-er construction represents the reinforced commonalities of all instantiations encountered up to this point and hence consists of two noun phrase- or clauselike components juxtaposed to express some comparative correlative relationship between an independent variable identified by the first part (the X-er) and a dependent variable identified by the second part (the Y-er) (Fillmore, Kay, & O’Connor, 1988; Goldberg, 2003).

Evidence for the usage-based emergence of schematic constructions comes from experimental and corpus studies on first language acquisition (Tomasello, 2003, pp. 113-126). At around 18 months of age, children begin to produce two-word utterances by stringing together two lexical items they have encountered in their input and learned to associate with a particular frame. Like the frames described above, the childhood precursors represent the typical aspects of different situations in a child’s lifeworld, yet, presumably on a more specific level and from a more con-strained view than in adulthood. For instance, in an attempt to convey the message that the quantity of something has increased or that they would wish so, English-speaking children combine words they have heard before in relevant contexts (e.g., Would you like more cookies?) to produce two-word utterances like in Example 13.

To give another example, when indicating that something has disappeared from the immediate scene, they reuse words from relevant contexts in their prior input (e.g., Daddy is gone.) to produce the utterances in Example 14.

2.2. USAGE-BASED LINGUISTICS 35

(Tomasello, 2003, pp. 114–116, adapted from Braine and Bowerman, 1976)

At this developmental stage, two-word utterances are “totally concrete in the sense that they are comprised only of concrete pieces of language, not categories” (Tomasello, 2003, p. 114). Their only categorical import is to partition the associated frame into two symbolizable component parts. From this, more schematic templates begin to emerge around this age which consist of a specific lexical anchor or pivot and a category which generalizes over the variant parts of the two-word utterances, here, moreX and X gone. While these pivot schemas are in part bound to specific lexical items, they are in part productive in the sense that they are extended to novel words.

For example, in an experimental study by Tomasello, Akhtar, Dodson, and Rekau (1997), children at 22 months of age were taught nonce words (e.g., they heard:

Look! The tam, the tam!, accompanied by a novel toy character) and in following elicitation tasks were able to combine the newly learned words with already acquired pivot schemas (e.g., when encountering two exemplars of the novel toy character and asked: What are these?, they replied: More tam.). Pivot schemas do not involve grammatical relations yet. Only later, children acquire item-specific constructions around specific verbs (“verb islands”, Tomasello, 1992) which mark grammatical relations, for example, by word order (e.g., Draw me! is not synonymous with Me draw! anymore).

In line with a usage-based model of language learning, this suggests that chil-dren acquire partially schematic constructions by generalizing over similar lexical strings in their language use. Tomasello argues that when constructing two-word utterances and linguistic pivot schemas from the language they hear, children rely on domain-general cognitive and social-cognitive abilities, such as planning ahead problem-solving or goal-directed behavior and applying a behavioral strategy in dif-ferent contexts. Informed by Langacker’s characterization of schematization as lack of reinforcement across usage events, Tomasello metaphorically likens the process of learning schematic constructions from item-specific strings to the stacking of

over-head transparencies, “each with a stored sequence such as an utterance printed on it” so that “repeated elements can be clearly recognized through the entire stack, but the variable elements are blurred because of their variability” (2003, p. 124).