5

USAGE-BASED APPROACHES
TO SLA1
Nick C. Ellis and Stefanie Wulff

The Theory and Its Constructs

Various approaches to second language acquisition (SLA) can be labeled as
“usage-based.” What unites these approaches is their commitment to two work-
ing hypotheses:

(1) Language learning is primarily based on learners’ exposure to their second

language (L2) in use, that is, the linguistic input they receive.
(2) Learners induce the rules of their L2 from the input by employing cognitive
mechanisms that are not exclusive to language learning, but that are general
cognitive mechanisms at work in any kind of learning, including language
learning.

In the following, we will look at the following major constructs of usage-based

approaches to SLA in more detail:

• Constructions: language learning is the learning of constructions, pairings

of form and meaning or function. Constructions range from simple mor-
phemes like -ing to complex and abstract syntactic frames such as Subject-
Verb-Object-Object (as in Nick made Steffi a sandwich).
• Associative language learning: learning constructions means learning the asso-
ciation between form and meaning or function. The more reliable the associa-
tion between a form and its meaning or function, the easier it is to learn. For
example, the sound sequence /ˈsæn(d)wɪtʃ/ is reliably associated with a par-
ticular meaning (“slices of meet and/or cheese between two slices of bread”).
The form -ing, in contrast, has different meaning/functions in different con-
texts, making it comparatively harder to learn.
76 Ellis and Wulff

• Rational cognitive processing: language learning is rational such that a learner’s

knowledge of a given form–meaning pair at any point in their language devel-
opment is a reflection of how often and in what specific contexts the learner
has encountered that form–meaning pair.
• Exemplar-based learning: language learning is in large parts implicit in the
sense of taking place without the learner being consciously aware of it. The
learner’s brain engages simple learning mechanisms in distributional analyses
of the exemplars of a given form–meaning pair that take various characteristics
of the exemplar into consideration, including how frequent it is, what kind of
words and phrases and larger contexts it occurs with, and so on.
• Emergent relations and patterns: language learning is a gradual process in
which language emerges as a complex and adaptive (in the sense of continu-
ously fine-tuning) system from the interaction of simple cognitive learning
mechanisms with the input (and in interaction with other speakers in various
social settings).

Constructions
The basic units of language representation are constructions. Constructions are
pairings of form and meaning or function. By that definition, we know that
simple words like, say, squirrel, must be constructions: a form—that is, a particular
sequence of letters or sounds—is conventionally associated with a meaning (in the
case of squirrel, something like ‘agile, bushy-tailed, tree-dwelling rodent that feeds
on nuts and seeds’). In Construction Grammar, constructions are not restricted
to the level of words (Goldberg, 2006). Instead, these form–function pairings
are assumed to pervade all layers of language. Simple morphemes such as -licious
(roughly meaning ‘delightful or extremely attractive’) are constructions. Idiomatic
expressions such as I can’t wrap my head around this (meaning ‘I do not fully com-
prehend this’) are constructions. Even abstract syntactic frames are constructions:
sentences like Nick gave the squirrel a nut, Steffi gave Nick a hug, or Bill baked Jessica
a cake all have a particular form (Subject-Verb-Object-Object) that, regardless of
the specific words that realize its form, share at least one stable aspect of meaning:
something is being transferred (nuts, hugs, and cakes). Some constructions do not
have a meaning in the traditional sense, but serve more functional purposes; passive
constructions, for example, serve to shift what is in attentional focus by defocusing
the agent of the action (compare an active sentence such as Bill baked Jessica a cake
with its passive counterpart A cake was baked for Jessica).
Constructions can be simultaneously represented and stored in multiple forms
and at various levels of abstraction: table + s = tables; [Noun] + (morpheme -s) =
‘plural things’). Ultimately, constructions blur the traditional distinction between
lexicon and grammar. A sentence is not viewed as the application of grammatical
rules to put a number of words obtained from the lexicon in the right order; a sen-
tence is instead seen as a combination of constructions, some of which are simple
 Usage-Based Approaches to SLA 77

and concrete while others are quite complex and abstract. For example, What did
Nick give the squirrel? comprises the following constructions:

• Nick, squirrel, give, what, do constructions

• VP, NP constructions
• Subject-Verb-Object-Object construction
• Subject-Auxiliary inversion construction

We can therefore see the language knowledge of an adult as a huge warehouse

of constructions. Constructions vary in their degree of complexity and abstrac-
tion. Some of them can be combined with one another while others cannot; their
combinability largely depends on whether their meanings/functions are compat-
ible, or can at least be coerced into compatibility, given the specific context and
situation in which a speaker may want to use them together. The more often a
speaker encounters a particular construction, or combination of constructions,
in the input, the more entrenched that (arrangement of) constructions becomes.

Associative Learning Theory

Constructions that are frequent in the input are processed more readily than rare
constructions are. This empirical fact is compatible with the idea that we learn
language from usage in an associative manner. Let’s stick to words for now, though
the same is true for letters, morphemes, syntactic patterns, and all other types of
constructions. Through experience, a learner’s perceptual system becomes tuned
to expect constructions according to their probability of occurrence in the input,
with words like one or won occurring more frequently than words like seventeen
or synecdoche.
When a learner notices a word in the input for the first time, a memory is
formed that binds its features into a unitary representation, such as the phonologi-
cal sequence /wʌn/ or the orthographic sequence one. Alongside this represen-
tation, a so-called detector unit is added to the learner’s perceptual system. The
job of the detector unit is to signal the word’s presence whenever its features are
present in the input. Every detector unit has a set resting level of activation and
some threshold level which, when exceeded, will cause the detector to fire. When
the component features are present in the environment, they send activation to the
detector that adds to its resting level, increasing it; if this increase is sufficient to
bring the level above threshold, the detector fires. With each firing of the detec-
tor, the new resting level is slightly higher than the previous one—the detector is
primed. This means it will need less activation from the environment in order to
reach threshold and fire the next time. Priming events sum to lifespan-practice
effects: features that occur frequently acquire chronically high resting levels. Their
resting level of activation is heightened by the memory of repeated prior acti-
vations. Thus our pattern-recognition units for higher-frequency words require
78 Ellis and Wulff

less evidence from the sensory data before they reach the threshold necessary for
firing.
The same is true for the strength of the mappings from form to interpretation.
Each time /wʌn/ is properly interpreted as one, the strength of this connection
is incremented. Each time /wʌn/ signals won, this is tallied too, as are the less
frequent occasions when it forewarns of wonderland. Thus the strengths of form–
meaning associations are summed over experience. The resultant network of asso-
ciations, a semantic network comprising the structured inventory of a speaker’s
knowledge of language, is tuned such that the spread of activation upon hearing
the formal cue /wʌn/ reflects prior probabilities of its different interpretations.
Many additional factors qualify this simple picture. First, the relationship
between frequency of usage and activation threshold is not linear but follows a
curvilinear “power law of practice” whereby the effects of practice are greatest at
early stages of learning, but eventually reach asymptote (see Chapter 6). Second,
the amount of learning induced from an experience of a construction depends
upon the salience of the form (i.e., how much it stands out relative to its context)
and the importance of understanding it correctly. Third, the learning of a con-
struction is interfered with if the learner already knows another form that cues
that interpretation, or conversely, if the learner knows another interpretation for
that form. Fourth, a construction may provide a partial specification of the struc-
ture of an utterance, and hence an utterance’s structure is specified by a number
of distinct constructions which must be collectively interpreted. Some cues are
much more reliable signals of an interpretation than others, and it is not just first-
order probabilities that are important—sequential probabilities matter a great deal
as well, because context qualifies interpretation. For example, the interpretation of
/wʌn/ in the context Alice in wun . . . is already clear after the learner has heard
Alice in . . .; in other words, Alice in and /wʌn/ are highly reliably associated with
each other. If a sentence starts out with I /wʌn/ . . ., in contrast, several competing
interpretations are co-activated (I wonder . . ., I won . . ., I once . . ., etc.) because
the first person pronoun I is a much less reliable cue for the interpretation of /
wʌn/ than Alice.

Rational Language Processing

These associative underpinnings allow language users to be rational in the sense of
having a mental model of their language that is custom-fit to their linguistic expe-
rience at any given time (Ellis, 2006a). The words that they are likely to hear next,
the most likely senses of these words, the linguistic constructions they are most
likely to utter next, the syllables they are likely to hear next, the graphemes they are
likely to read next, the interpretations that are most relevant, and the rest of what’s
coming . . .? (next) across all levels of language representation, are made readily
available to them by their language processing systems. Their unconscious language
representation systems are adaptively tuned to predict the linguistic constructions
 Usage-Based Approaches to SLA 79

that are most likely to be relevant in the ongoing discourse context, optimally pre-
paring them for comprehension and production. As a field of research, the ratio-
nal analysis of cognition is guided by the principle that human psychology can
be understood in terms of the operation of a mechanism that is optimally adapted
to its environment in the sense that the behavior of the mechanism is as efficient
as it conceivably could be, given the structure of the problem space and the cue-
interpretation mappings it must solve (Anderson, 1989).

Exemplar-Based Learning
Much of our language use is formulaic, that is, we recycle phrasal constructions
that we have memorized from prior use (Wulff, 2008). However, we are obviously
not limited to these constructions in our language processing. Some construc-
tions are a little more open in scope, like the slot-and-frame greeting pattern
[Good + (time-of-day)] which generates examples like Good morning and Good
afternoon. Others still are abstract, broad-ranging, and generative, such as the sche-
mata that represent more complex morphological (e.g., [NounStem-PL]), syn-
tactic (e.g., [Adj Noun]), and rhetorical (e.g., the iterative listing structure, [the
( ), the ( ), the ( ), . . ., together they . . .]) patterns. Usage-based theories investi-
gate how the acquisition of these productive patterns, generative schema, and
other rule-like regularities of language is based on exemplars. Every time the
language learner encounters an exemplar of a construction, the language system
compares this exemplar with memories of previous encounters of either the same
or a sufficiently similar exemplar to retrieve the correct interpretation. According
to exemplar theory, constructions such as Good + (time of day), [Adj Noun], or
[NounStem-PL] all gradually emerge over time as the learner’s language system,
processing exemplar after exemplar, identifies the regularities that exemplars share
and makes the corresponding abstractions.

The Associative Bases of Abstraction

Prototypes, the exemplars that are most typical of their categories, are those that
are similar to many members of their category but not similar to members of other
categories. People more quickly classify sparrows as birds (or other average sized,
average colored, average beaked, average featured specimens) than they do birds
with less common features or feature combinations, like geese or albatrosses. They
do so on the basis of an unconscious frequency analysis of the birds they have
known (their usage history) with the prototype that reflects the central tenden-
cies of the distributions of the relevant features of these memorized exemplars.
We don’t walk around consciously counting these features, but yet we have very
accurate knowledge of the underlying distributions and their most usual settings.
We are really good at this. Research in cognitive psychology demonstrates that
such implicit tallying is the raw basis of human pattern recognition, categorization,
80 Ellis and Wulff

and rational cognition. As the world is classified, so language is classified. As for

the birds, so for their plural forms. In fact, world and language categorization
go hand in hand: Psycholinguistic research demonstrates that people are faster at
generating plurals for the prototype or default case that is exemplified by many
types, and are slower and less accurate at generating “irregular” plurals, the ones
that go against the central tendency and that are exemplified by fewer types, such
as [plural + ‘NounStems’ = ‘NounStems-es’] or, worse still, [plural + moose = ?],
[plural + noose = ?], [plural + goose = ?].
These examples make it clear that there are no 1:1 mappings between cues
and their outcome interpretations. Associative learning theory demonstrates that
the more reliable the mapping between a cue and its outcome, the more readily it
is learned. Consider an ESL learner trying to learn from naturalistic input what
-s at the ends of words might signify. This particular form has several potential
interpretations: It could be the plural (squirrels), it could indicate possession (Nick’s
hat), it could mark third person singular present (Steffi sleeps), and so on. Therefore,
if we evaluate -s as a cue for any one of these outcomes, it is clear that the cue
will be abundantly frequent in learners’ input, yet neither of the cues are reliably
associated with their interpretation or outcome. A similar picture emerges when
we reverse the directionality of our thinking: plural -s, third person singular pres-
ent -s, and possessive -s all have variant expression as the allomorphs [s], [z], and
[ɨz].Thus if we evaluate just one of these, say, [ɨz], as a cue for one particular
outcome, say, plurality, then it is clear that there are many instances of that out-
come in the absence of the cue. Such contingency analysis of the reliabilities of
the cue-interpretation associations suggests that they will not be readily learnable.
High-frequency grammatical functors are often highly ambiguous in their inter-
pretations (Goldschneider & DeKeyser, 2001).
Connectionism is one strand of research in SLA that seeks to investigate how
simple associative learning mechanisms such as the kind of contingency analysis
mentioned earlier meets the complex language evidence available to a learner in
their input and output. The term “connectionist” reflects the idea that mental
and behavioral models are in essence interconnected networks of simple units.
Connectionist models are typically run as computer simulations. The simulations
are data-rich and process-light: Massively parallel systems of artificial neurons use
simple learning processes to statistically generalize over masses of input data. It is
important that the input data is representative of learners’ usage history, which is
why connectionist and other input-influenced research rests heavily on large-scale,
maximally representative digital collections of authentic language (these are often
called databanks or corpora). Connectionist simulations show how prototypes
emerge as the prominent underlying structural regularity in the whole problem
space, and how minority subpatterns of inflection regularity, such as the English
plural subpatterns discussed earlier (or the much richer varieties of the Ger-
man plural system, for example), also emerge as smaller, less powerful attractors.
Connectionism provides the computational framework for testing usage-based
 Usage-Based Approaches to SLA 81

theories as simulations, for investigating how patterns appear from the interactions
of many language parts.

Emergent Relations and Patterns

Complex systems are those that involve the interactions of many different parts,
such as ecosystems, economies, and societies. All complex systems share the key
aspect that many of their systematicities are emergent: They develop over time
in complex, sometimes surprising, dynamic, and adaptive ways. Complexity arises
from the interactions of learners and problems too. Consider the path of an ant
making its homeward journey on a pebbled beach. The path seems complicated
as the ant probes, doubles back, circumnavigates, and zigzags. But these actions
are not deep and mysterious manifestations of intellectual power. Instead the con-
trol decisions are simple and few in number. An environment-driven problem
solver often produces behavior that is complex because it relates to a complex
environment.
Language is a complex adaptive system (Beckner et al., 2009; Ellis & Larsen-
Freeman, 2009; see also Chapter 12). It comprises the interactions of many players:
People who want to communicate and a world to be talked about. It operates
across many different levels (neurons, brains, and bodies; phonemes, constructions,
interactions, and discourses), different human conglomerations (individuals, social
groups, networks, and cultures), and different timescales (evolutionary, epigenetic,
ontogenetic, interactional, neurosynchronic, diachronic). “Emergentists believe
that simple learning mechanisms, operating in and across the human systems for
perception, motor-action and cognition as they are exposed to language data as
part of a communicatively-rich human social environment by an organism eager
to exploit the functionality of language, suffice to drive the emergence of complex
language representations” (Ellis, 1998, p. 657).

Two Languages and Language Transfer

Our neural apparatus is highly plastic in its initial state. It is not entirely an empty
slate, since there are broad genetic constraints on the usual networks of system-
level connections and on the broad timetable of maturation. Nevertheless, the cor-
tex of the brain is broadly equipotent in terms of the types of information it can
represent (Elman et al., 1996). From this starting point, the brain quickly responds
to the input patterns it receives, and through associative learning, it optimizes its
representations to model the particular world of an individual’s experience. The
term “neural plasticity” summarizes the fact that the brain is tuned by experience.
Our neural endowment provides a general purpose cognitive apparatus that, con-
strained by the makeup of our human bodies, filters and determines our experi-
ences. In the first few years of life, the human learning mechanism optimizes its
representations of the first language (L1) being learned. Thousands of hours of L1
82 Ellis and Wulff

processing tunes the system to the cues of the L1 and automatizes its recognition
and production. It is impressive how rapidly we start tuning into our ambient lan-
guage and disregarding cues that are not relevant to them (Kuhl, 2004). One result
of this process is that the initial state for SLA is no longer a plastic system; it is one
that is already tuned and committed to the L1. Our later experience is shaded by
prior associations; it is perceived through the memories of what has gone before.
Since the optimal representations for the L2 do not match those of the L1, SLA is
impacted by various types of L1 interference. Transfer phenomena pervade SLA
(Flege, 2002; Jarvis & Pavlenko, 2008; Lado, 1957; MacWhinney, 1997; Odlin,
1989; Weinreich, 1953).

Associative Aspects of Transfer: Learned Attention

and Interference
Associative learning provides the rational mechanisms for L1 acquisition from
input—analysis and usage—allowing just about every human being to acquire
fluency in their native tongue. Yet although L2 learners too are surrounded by
language, not all of it “goes in,” and SLA is typically limited in success. This is
Corder’s distinction between input, the available target language, and intake, that
subset of input that actually gets in and that the learner utilizes in some way
(Corder, 1967). Does this mean that SLA cannot be understood according to the
general principles of associative learning? If L1 acquisition is rational, is SLA fun-
damentally irrational? No. Paradoxically perhaps, it is the very achievements of
L1 acquisition that limit the input analysis of the L2. Associative learning theory
explains these limitations too, because associative learning in animals and humans
alike is affected by what is called learned attention.
We can consider just one example of learned attention here. Many gram-
matical form–meaning relationships are both low in salience and redundant in
the understanding of the meaning of an utterance. It is often unnecessary, for
instance, to interpret inflections that mark grammatical meanings such as tense
because they are usually accompanied by adverbs that indicate the temporal
reference: “if the learner knows the French word for ‘yesterday,’ then in the
utterance Hier nous sommes allés au cinéma (Yesterday we went to the movies)
both the auxiliary and past participle are redundant past markers” (Terrell, 1991,
p. 59). This redundancy is much more influential in SLA than L1 acquisition.
Children learning their native language only acquire the meanings of temporal
adverbs quite late in development. But L2 learners already know about adverbs
from their L1 experience, and adverbs are both salient and reliable in their com-
municative functions, while tense markers are neither (see Chapter 7). Thus,
the L2 expression of temporal reference begins with a phase where reference
is established by adverbials alone, and the grammatical expression of tense and
aspect thereafter emerges only slowly if at all (Bardovi-Harlig, 2000; see also
Chapter 4).
 Usage-Based Approaches to SLA 83

This is an example of the associative learning phenomenon of “blocking,”

where redundant cues are overshadowed because the learners’ L1 experience leads
them to look elsewhere for the cues to interpretation (Ellis, 2006b). Under normal
L1 circumstances, usage optimally tunes the language system to the input; under
these circumstances of low salience of L2 form and blocking, however, all the
extra input in the world can sum to nothing, with interlanguage sometimes being
described as having “fossilized” (Han & Odlin, 2006). Untutored adult associative
L2 learning from naturalistic usage can thus stabilize at a “Basic Variety” of inter-
language which, although sufficient for everyday communicative purposes, pre-
dominantly comprises just nouns, verbs, and adverbs, with little or no functional
inflection and with closed-class items, in particular determiners, subordinating
elements, and prepositions, being rare or not present at all (Klein, 1998).
The usual social-interactional or pedagogical reactions to such nonnative-like
utterances involve an interaction partner (Long, 1983; Mackey, Abbuhl, & Gass,
2011; see also Chapter 10) or instructor (Doughty & Williams, 1998) who inten-
tionally brings additional evidence to the learner’s attention by some means of
attentional focus that helps the learner to “notice” the cue (Schmidt, 2001). This
way, SLA can be freed from the bounds of L1-induced selective attention: a focus
on form is provided in social interaction (Tarone, 1997; see also Chapter 11) that
recruits the learner’s explicit conscious processing. We might say that the input
to the associative network is “socially gated” (Kuhl, 2007).

What Counts as Evidence?

Like other enterprises in cognitive science and cognitive neuroscience, usage-
based approaches are not restricted to one specific research methodology or evi-
dential source. Indeed, different approaches require different methods, and often
a combination of different qualitative and quantitative methods. As mentioned
earlier, many usage-based analyses employ data from large digitized collections
of language, so-called corpora; computational modeling is at the heart of rational
cognition analysis, exemplar theory, and emergentist analyses alike. Other relevant
research methods include classroom field research, psycholinguistic studies of pro-
cessing, and dense longitudinal recording.
Corpus-based analysis constitutes a particularly growing trend across usage-based
paradigms (McEnery & Hardie, 2012; Sinclair, 1991). If language learning is in the
social-cognitive linguistic moment of usage, we need to capture all these moments
so that we can objectively study them. We need large, dense, longitudinal corpora
of language use, with audio, video, transcriptions, and multiple layers of annotation,
for data sharing in open archives. We need these in sufficient dense mass so that we
can chart learners’ usage history and their development (Tomasello & Stahl, 2004).
We need them in sufficient detail that we can engage in detailed analyses of the
processes of interaction (Kasper & Wagner, 2011). MacWhinney has long been
working toward these ends, first with CHILDES (MacWhinney, 1991), a corpus
84 Ellis and Wulff

of L1 acquisition data, and later with Talkbank (MacWhinney, 2007), a corpus that
also covers language data from L2 learners. Alongside these and other corpora, a
growing number of computer tools are becoming available that assist the researcher
in analyzing corpus data. These corpus tools can help researchers interested in the
most diverse areas of SLA by covering the full range from qualitative data analysis,
such as a fine-grained conversation analysis of individual corpus files (say, a tran-
scribed conversation between a student and an ESL teacher), to semi-quantitative
analysis of a representative sample of attestations of a particular phenomenon (such
as the use of the -ing morpheme by English language learners), to large-scale quan-
titative analysis of distributional patterns (e.g., the association strength between
verbs and the larger constructions they occur in; see the exemplary study in this
chapter or Gries & Wulff, 2005, 2009, for examples).

What Are Some Common Misunderstandings

about the Theory?
Broad frameworks, particularly those that revive elements of no-longer-fashionable
theories such as behaviorism or structuralist approaches to linguistics, open the
potential for misunderstanding. Common misconceptions include that connec-
tionism is the new behaviorism; that connectionist models cannot explain creativity
and have no regard for internal representation; and that cognitive approaches deny
influence of social factors, motivational aspects, and other individual differences
between learners. At the heart of most of these misunderstandings is the idea that
usage-based analyses only do number-crunching, with too much of a focus on
the effects that the frequency of constructions and other cues play in the learning
process. While it is true that most usage-based approaches will discuss frequency
as one of several factors, no usage-based theorist would claim that frequency is
the only factor impacting SLA. In fact, there is a lively debate among usage-based
theorists about the exact role frequency effects play in what is conceived of as a
complex network of factors that can mute and amplify each other in complex
ways (Ellis & Larsen-Freeman, 2006). At an even more fundamental level, what
constitutes a frequency effect in the first place is a question we are far from hav-
ing answered. Without going into too much detail here, there is ample empiri-
cal evidence, for instance, that we cannot always define a frequency effect by the
rule “the more frequent, the more salient/important/relevant”—by that rationale,
English articles and prepositions, which are the most frequent words in the English
language, should not pose such an obstinate challenge to the average language
learner! Instead, it seems that frequency effects come in different kinds (as absolute
frequencies, ratios, association strengths, and other distributional patterns), and they
will have differently weighted impacts depending on the target structure under
examination, and, crucially, depending on the state of the learner’s language devel-
opment. An emergentist/complex systems approach views SLA as a dynamic pro-
cess in which regularities and systems emerge from many of the processes covered
 Usage-Based Approaches to SLA 85

in this volume—from the interaction of people, brains, selves, societies, and cultures
using languages in the world (Beckner et al., 2009; Ellis, 2008)—while at the same
time investigating component processes in a rigorous fashion.

An Exemplary Study: Ellis, O’Donnell, and Römer (2014a)

Research Questions
While previous studies were able to demonstrate that frequency, prototypicality,
and contingency are factors that impact L2 learners’ constructional knowledge,
most of these studies have considered only one of these factors at a time. This study
wanted to determine whether and how these factors jointly affect L2 learners’
constructional knowledge. The specific kind of constructions this study focused on
are so-called verb-argument constructions (VACs). VACs are semi-abstract patterns
that comprise verbs and the arguments they occur with, such as ‘V across N’ or ‘V
of N’; in this study, the authors examined VACs that another team of researchers
previously identified using corpus analysis (Francis, Hunston, & Manning, 1996).

Methods
One hundred thirty-one German, 131 Spanish, and 131 Czech advanced L2
learners of English as well as 131 native speakers of English were engaged in a
free association task: They were shown 40 VAC frames such as ‘V across N or ‘it V
of the N’ and asked to fill in the verb slot with the first word that came to mind.
The learners’ responses were compared with results obtained from two native
speaker databases. To get an impression of the frequencies with which different
verbs occur in the VACs examined, and to calculate how strongly each verb is
associated with the individual VACs, the authors consulted the British National
Corpus (BNC). The BNC is a 100 million word corpus of British English that
strives to be representative of language use across different registers and genres.
To obtain the verb type frequencies, one can simply run a search for the VACs in
the BNC and count how often each verb type occurs. To calculate the associa-
tion strength between each verb type and each VAC, the authors used a specific
association measure called DeltaP (for more information on how DeltaP works,
see Ellis, 2006a). To see how prototypical the verbs selected by the participants
would be for each VAC, the authors consulted a second database called WordNet
(Miller, 2009). WordNet is a lexical database, so unlike the BNC, it is not a col-
lection of cohesive and complete texts and dialogues, but rather a thesaurus-like
database that groups words together based on their meanings. Using sophisticated
computational techniques, the authors used this information to generate semantic
networks for each of the VACs examined. For the ‘V across N’ VAC, for instance,
the verbs in the center of the network are go, move, run, and travel, while verbs like
shout, splash, and echo constitute less prototypical verbs in that VAC.
86 Ellis and Wulff

Main Findings
A multifactorial analysis (i.e., a statistic that can gauge the impact of more than
one factor on a specific outcome at a time; in our example, it measured the poten-
tial impact of frequency, prototypicality, and contingency on speakers’ associations)
revealed that for all of the VACs examined, each factor made an independent
contribution to learners’ and native speakers’ associations:

1. The more frequently a particular verb occurred in a specific VAC in the

native speaker corpus data, the more likely it was elicited as a response for that
VAC in the word association experiment.
2. The more strongly a verb and a VAC were associated with each other as
expressed in their DeltaP association scores, the more likely that verb was
elicited as a response for that VAC in the word association experiment.
3. The more prototypical a verb was for the VAC as indicated by its position in
the semantic networks the authors generated for each VAC, the more likely it
was elicited as a response for that VAC in the word association experiment.

Theoretical Implications
Based on the statistical analyses, the authors concluded that advanced L2 learn-
ers’ knowledge of VACs involves rich associations that are very similar in kind
and strength to those of native speakers (Ellis, O’Donnell, & Römer, 2014). The
word associations generated in the experiment testified to learners having rich
associations for VACs that are tuned by verb frequency, verb prototypicality, and
verb-VAC contingency alike—factors that, in combination, interface across syntax
and semantics.

Why/How This Theory Provides an Adequate Explanation of

Observable Phenomena in SLA
Observation 1: Exposure to input is necessary for SLA. Usage based approaches are
input-driven, emphasizing the associative learning of constructions from input.
As with other statistical estimations, a large and representative sample of language
is required for the learner to abstract a rational model that is a good fit to the
language data. Usage is necessary, and it is sufficient for successful L1 acquisition
though not for SLA. This is because the initial state for SLA is knowledge of an
L1, and the learner’s representations, processing routines, and attention to language
are tuned and committed to the L1.
Observation 2: A good deal of SLA happens incidentally. The majority of language
learning is implicit. Implicit tallying is the raw basis of human pattern recogni-
tion, categorization, and rational cognition. All of the counting that underpins
the setting of thresholds and the tuning of the system to the probabilities of the
 Usage-Based Approaches to SLA 87

input evidence is unconscious. So also is the emergence of structural regulari-

ties, prototypes, attractors, and other system regularities. At any one point we are
conscious of one particular communicative meaning, yet meanwhile the cogni-
tive operations involved in each of these usages are tuning the system without
us being aware of it (Ellis, 2002). We know (or can be shown to be sensitive
to in our processing) far too many linguistic regularities for us to have explic-
itly learned them. Usage-based approaches maintain that incidental associative
learning provides the rational mechanisms and is sufficient for L1 acquisition
from input-analysis and usage, allowing just about every human being to acquire
fluency in their native tongue. They do not suffice for SLA because of learned
attention.
Observation 3: Learners come to know more than what they have been exposed to in the
input. The study of implicit human cognition shows us to know far more about
the world than we have been exposed to or have been explicitly taught. Prototype
effects are one clear and ubiquitous example of this: learners who have never been
exposed to the prototype of a category nevertheless classify it faster and more
accurately than examples further from the central tendency, and name it with the
category label with great facility. The same is true for language, where learners go
beyond the input in producing U-shaped learning, with novel errors (like goed
instead of went) and other systematicities of stages of interlanguage development
in L2 acquisition, for example of negation or question formation. These creations
demonstrate that the learners’ language system is constantly engaged in making
generalizations and finding abstractions of systematicities.
Observation 4: Learners’ output (speech) often follows predictable paths with predictable
stages in the acquisition of a given structure. As in L1 acquisition, SLA is characterized
not by complete idiosyncrasy or variability but rather by predictable errors and
stages during the course of development: interlanguage is systematic. Usage-based
approaches hold that these systematicities arise from regularities in the input: For
example, constructions that are much more frequent, that are consistent in their
mappings and exhibit high contingency, that have many friends (constructions
that behave in a similar way) of like-type, and that are salient, are likely to be
acquired earlier than those that do not have these features (Ellis, 2007).
Observation 6: Second language learning is variable across linguistic subsystems. The
learners’ mental lexicon is diverse in its contents, spanning lexical, morphological,
syntactic, phonological, pragmatic, and sociolinguistic knowledge. Within any of
these areas of language, learners may master some structures before they acquire
others. Such variability is a natural consequence of input factors such as exem-
plar type and token frequency, recency, context, salience, contingency, regular-
ity, and reliability, along with the various other associative learning factors that
affect the emergence of attractors in the problem space. Some aspects of these
problem spaces are simpler than others. Second language learning is a piecemeal
development from a database of exemplars with patterns of regularity emerging
dynamically.
88 Ellis and Wulff

Observation 7: There are limits on the effects of frequency on SLA. This is explic-
itly addressed above under Associative aspects of transfer: learned attention and
interference.
Observation 8: There are limits on the effect of a learner’s first language on SLA. The
effect of a learner’s L1 is no longer considered the exclusive determinant of SLA
as proposed in the Contrastive Analysis Hypothesis. Usage-based accounts see
the major driving force of language acquisition to be the constructions of the
target language and the learner’s experience of these constructions. However, the
significance of transfer from L1 in the L2 learning process is uncontroversial. As
we explain earlier under “Two Languages and Language Transfer,” at every level
of language, there is L1 influence, both negative and positive. The various cross-
linguistic phenomena of learned selective attention, overshadowing and blocking,
latent inhibition, perceptual learning, interference, and other effects of salience,
transfer, and inhibition all filter and color the perception of the L2. So usage-based
accounts of L2 acquisition look at the effects of both L1 and L2 usage upon L2A
(Robinson & Ellis, 2008).
Observation 9: There are limits on the effects of instruction on SLA. L1-tuned learned
attention limits the amount of intake from L2 input, thus restricting the endstate
of SLA. Attention to language form is sometimes necessary to allow learners to
notice some blocked, overshadowed, or otherwise nonsalient aspect of the lan-
guage form. Reviews of the empirical studies of instruction demonstrate that social
recruitment of learners’ conscious, explicit learning processes can be effective.
However, instruction is not always effective. Any classroom teacher can provide
anecdotal evidence that what is taught is not always learned. But this observation
can be made for all aspects of the curriculum, not just language. Explicit knowl-
edge about language is of a different stuff from that of the implicit representational
systems, and it need not impact upon acquisition for a large variety of reasons.
Explicit instruction can be ill-timed and out of synchrony with development
(Pienemann, 1998; see also Chapter 9); it can be confusing; it can be easily forgot-
ten; it can be dissociated from usage, lacking in transfer-appropriateness and thus
never brought to bear so as to tune attention to the relevant input features during
usage; it can be unmotivating; it can fail in so many ways.

The Explicit/Implicit Debate

Learning a new symbol, for example, the French word for the symbol , étoile,
initially involves explicit learning: you are consciously aware of the fact that you
did not know the French word for ‘star’ before, and that now you do (Ellis, 1994).
Some facts about how to use étoile properly you may not know yet, such as its
proper pronunciation, its grammatical gender, synonymous forms, words, phrases,
and idioms that étoile is associated with. Some of these facts you will learn by
making a conscious effort, that is, via explicit learning; other facts you will not
consciously figure out but rather learn implicitly. Without you being aware of it,
 Usage-Based Approaches to SLA 89

your language system is hard at work, upon each subsequent encounter of étoile,
to fill in these knowledge gaps and fine-tune the mental representation you have
for this construction.
Despite that many of us go to great lengths to engage in explicit language
learning, the bulk of language acquisition is implicit learning from usage. Most
knowledge is tacit knowledge; most learning is implicit; the vast majority of our
cognitive processing is unconscious. Implicit learning supplies a distributional
analysis of the problem space: our language system implicitly figures out how
likely a given construction is in particular contexts, how often they instantiate one
sense or another, how these senses are in turn associated with different features of
the context, and so on. To the extent that these distributional analyses are con-
firmed time and again through continuous exposure to more input, generaliza-
tions and abstractions are formed that are also largely implicit.
Implicit learning would not do the job alone. Some aspects of an L2 are
unlearnable—or at best are acquired very slowly—from implicit processes alone.
In cases where linguistic form lacks perceptual salience and so goes unnoticed by
learners, or where the L2 semantic/pragmatic concepts to be mapped onto the
L2 forms are unfamiliar, additional attention is necessary in order for the relevant
associations to be learned. To counteract the L1 attentional biases to allow implicit
estimation procedures to optimize induction, all of the L2 input needs to be made
to count (as it does in L1 acquisition), not just the restricted sample typical of the
biased intake of L2 acquisition.
Ellis (2005) reviews the instructional, psychological, social, and neurological
dynamics of the interface by which explicit knowledge of form–meaning associa-
tions impacts upon implicit language learning:

The interface is dynamic: It happens transiently during conscious processing,

but the influence upon implicit cognition endures thereafter. Explicit mem-
ories can also guide the conscious building of novel linguistic utterances
through processes of analogy. Patterned practice and declarative pedagogi-
cal grammar rules both contribute to the conscious creation of utterances
whose subsequent usage promotes implicit learning and proceduralization.
Flawed output can prompt focused feedback by way of recasts that present
learners with psycholinguistic data ready for explicit analysis. (p. 305)

Once a construction has been represented in this way, its use in subsequent implicit
processing can update the statistical tallying of its frequency of usage and prob-
abilities of form–function mapping.
So we believe that learners’ language systematicity emerges from their history of
interactions of implicit and explicit language learning, from the statistical abstrac-
tion of patterns latent within and across form and function in language usage. The
complex adaptive system of interactions within and across form and function is
far richer than that emergent from implicit or explicit learning alone (Ellis, 2014).