A newer approach, called connectionism, also relies on inductive inferencing to model learners’ performance, but it rejects an account that claims that it is rules that are being induced. Connectionists have developed computer models of net
works, which are held to be analogous to the neural networks in the brain. The net
works consist of interconnected nodes. The nodes are taken to represent neurons, which are connected with one another through synapses. The computer models are
“trained” by receiving massive amounts of target language input. As the language data are taken in, certain connections in the networks are strengthened. Connection weights are thus tunable; they fluctuate from second to second. At any given time, the weights are settling into or moving away from certain states. At any point in the “training” of these computer networks, the distribution of weighted connec
tions represents the network’s current map of the structure of the target language.
The networks are self-organizing, meaning they organize themselves in response to positive evidence, that is, the patterns in the input. However, some connectionist models use other learning algorithms, such as back propagation, in which connections are weakened when incorrect outputs are produced (i.e., they are “corrected”). In any event, although there is no conscious hypothesis formation of rules occurring, the networks model bottom-up inductive learning, mapping patterns that are present in the input and increasingly approximating the target in response to more and more input.
Another fascinating characteristic of such networks is that sometimes addition
al strengthening of connections results in output containing overgeneralization errors of the eated sort, even though such forms are not present in the input. In other words, the computer simulations appear to be producing rule-governed behavior even though they do not follow rules— that is, they are not programmed to follow rules. Connectionists have even been able to model the U-shaped learn
ing curve, known to exist for English past tense formation, whereby learners’ per
formance on both regular and irregular verbs is initially accurate, then reaches its nadir when learners overgeneralize the regular -ed ending to irregular verbs, and finally is restored to accuracy as the learners incorporate the irregular verb forms into their interlanguages. Plunkett and Marchman (1993) have pointed out that the U-shaped function reflects a dynamic competition between regular and irregular past tense verb endings in English.
As the number of verbs in the competition pool expands across the course of learning, there are shifts in the relative strengths of regular and irregular forms. The U-shaped dip in the learning curve occurs around the point in development in which there is a change in the proportional strength of regular -ed compared to other mappings.
Thus, sharp changes in behavior can be due to the natural evolution
of a nonlinear system even when the external forces are constant (Elman et al., 1996: 2 0 2 -2 0 3 ).
Other dynamic systems also experience this dynamic instability or bifurcation point. Referred to in Chaos/Complexity Theory as the “camel’s back phenome
non,” at some point in time “the last straw” is placed and the system undergoes a perturbation. Since such systems are self-organizing, the chaos subsides and new order emerges; the interlanguage has been restructured. A speaker’s gram
mar is thus seen not as a fixed body of rules but rather as “a statistical ensemble of language experiences” (N. Ellis, 2002) that changes every time a new utter
ance is processed— usually slightly, but on some occasions dramatically.
But connectionism does not merely help to model emergent approximations to the target language. Unlike behaviorists, connectionists are interested in cog
nitive processes, not just responses to stimuli (Gasser, 1990; McCarthy, 2001).
For instance, it is known that “in connectionist networks, items of new infor
mation are more easily incorporated when analyzed as variations on known information; new patterns are automatically assimilated to old patterns as much as possible” (Goldberg, 1995: 71). Shirai (1992) therefore suggests that con
nectionism can illuminate crosslinguistic transfer. When new languages are encountered, the existing representations of L I or other previously learned lan
guages are activated to reshape the incoming L2 data.
As attractive as connectionist models are, they clearly do not explain all human acquisitional experience. No computer can be programmed to reflect human agency or intentionality. Computers are basically passive; they are not goal-directed. There is no computer program that only selectively attends because it is daydreaming about the upcoming football game and not focusing on the lan
guage input. While they provide good models of implicit learning, they do not take into account attention. Because of this, they are slow to learn. Nevertheless, the results so far are intriguing and provide support for a claim I made long ago that frequency in the input is an important factor in second language acquisition (Larsen-Freeman, 1976). It pays to stick around! By the way, none of this redeems the practice of merely subjecting students to abundant comprehensible input, for the reason I gave earlier. It is still our professional responsibility to seek the most efficacious way to acquire a language, and merely providing learners with comprehensible input is not likely to be it (see #7 that follows).
3 . Wh a t a b o u t p a t t e r n e d s e q u e n c e s o r l e x i c o g r a m m a t i c a l u n i t s?
Connectionist modeling may be very useful in accounting for the acquisition of multiword strings/sequences or lexicogrammatical patterns, especially if Bolinger (1975) is correct that language, rather than being subserved by a rule system, is produced on the basis of “a large, capacious and redundantly struc
tured memory system” (Skehan, 1994: 181).
Well before the advent of computer-driven corpus linguistics, it had become increasingly clear that native speakers of a given language control thousands and thousands of fixed and semi-fixed patterned sequences that behave as single lexical units. Fixed expressions in English, such as “at any rate” and semi-fixed open expres
• Te a c h i n g La n g u a g e: Fr o m Gr a m m a r t o Gr a m m a r i n g
sions such as “I’m not at all sure that.. . ” have been credited with contributing a great deal of fluency to English native speaker speech (Pawley and Syder, 1983). In other words, according to Pawley and Syder, every time we speak, our utterances are not created anew by the application of rules, but are at least partly composed of these meaningful, unanalyzed chunks of language, which are retrieved holistically from memory, saving time in planning and carrying out syntactic operations.
Of course, while retrieving patterned sequences from memory might explain real
time language processing, not everyone would accept that acquisition of fixed pat
terned sequences accounts for all language acquisition. Surely, for example, there must be some generative mechanism that supports linguistic innovation. Although the posi
tion is somewhat controversial, some SLA researchers contend that a likely scenario is that learners acquire a stock of fixed and semi-fixed chunks of language, which they later analyze to discover generative grammatical rules (Wong Fillmore, 1976). In the case of first language acquisition, it is possible that the stock of patterned sequences becomes the material on which universal grammar operates (Peters, 1983).
In other words, grammar acquisition may be first characterized as a period of lexicalization, in which learners use prefabricated sequences or chunks of lan
guage, followed by a period of syntacticization, in which learners are able to infer a creative rule-governed system. The sequence may conclude with a period of relexicalization, in which learners, like native speakers, use patterned sequences to produce accurate and fluent speech (Skehan, 1994). This sequence may not characterize the learning of all L2 learners, though. It is possible, for example, that some second language learners, having satisfied their commu
nicative needs, will stop at the lexicalization stage. Then, too, more analytically inclined learners may push quickly into the syntacticization phase, while more memory-oriented learners may tend to treat language more in terms of chunks.
This is not to suggest that these processes are mutually exclusive in language users. In fact, some believe that both memorization and rule-governed processes operate. For instance, positron emission tomography (PET) scans have shown dif
ferent patterns of brain activation for human subjects asked to produce past tense of English regular and irregular verbs (Jaeger et al., 1996). From this finding, the researchers draw support for Pinker and Prince’s (1994) dual-systems hypothesis, which proposes that regular past tense is computed by rule and past tense forms of irregular verbs are computed by activating some aspect of lexical memory.
However some believe that a dual mechanism account is unnecessary and argue against its application to SLA (e.g., Murphy, unpublished manuscript). The fact that different areas of the brain are activated offers no insight into functional differences.
Besides, Jaeger et al.’s methodology is flawed, it is claimed. And even if there is a pro
cessing difference between regular and irregular forms, it may have less to do with the regular-irregular difference and more to do with their different frequencies of occurrence in the input (Seidenberg and Hoeffnei; 1998). Indeed, research by N. Ellis and Schmidt (1998) suggests that both regular and irregular forms can be account
ed for by associative memory using a simple connectionist model.
In sum, there may be a lot of truth to the statement that what we humans do is “push old language into new” (Becker, 1983), or retrieve chunks of language from our memories of discourse and reconfigure them in novel, principled ways.
How we do this, of course, is the big question.
4 . H O W ARE PATTERNED SEQUENCES RECONFIGURED TO PRODUCE