THE IMPORTANCE OF FORMULAIC LANGUAGE

Before we turn to how our brains store and represent formulaic language,Footnote ¹ it is useful to ask why we study formulaic sequence processing at all. It is becoming increasingly clear that formulaic language is an important element of language learning and use. Perhaps the best evidence for this is simply its ubiquity. Normal discourse, both written and spoken, contains large (but not yet fully determined) percentages of formulaic language. Oppenheim (Reference Oppenheim and Riggenbach2000) counted the multiword stretches of talk that occurred identically in practice and final renderings of a short speech on the same topic and found between 48 percent and 80 percent (overall mean of 66 percent) consisted of identical strings. Sorhus (Reference Sorhus1977) calculated that speakers in her corpus of spontaneous English Canadian speech used an item of formulaic language once every five words. Erman and Warren (Reference Erman and Warren2000) calculated that 52–58 percent of the language they analyzed was formulaic, and Foster (Reference Foster, Bygate, Skehan and Swain2001) came up with a figure of 32 percent using different procedures and criteria. Biber, Johansson, Leech, Conrad, and Finegan (Reference Biber, Johansson, Leech, Conrad and Finegan1999) found that around 30 percent of the words in their conversation corpus consisted of lexical bundles, and about 21 percent of their academic prose corpus. Howarth (Reference Howarth and Cowie1998) looked at frequent verbs in a social science/academic corpus and found that they occurred in either restricted collocations or in idioms in 31–40 percent of the cases. Rayson (Reference Rayson2008) found that 15 percent of text is formulaic according to a Wmatrix analysis. Overall, these studies suggest that formulaic language makes up between one third and one half of discourse.

For formulaic language to be so widespread, it must provide some useful purpose in communication. Schmitt and Carter (Reference Schmitt, Carter and Schmitt2004) listed a number of these purposes:

• • Expressing a message or idea (The early bird gets the worm = do not procrastinate)

• • Realizing functions ([I'm] just looking [thanks] = declining an offer of assistance from a shopkeeper)

• • Expressing social solidarity (Yeah it is = expressing agreement)

• • Transacting specific information in a precise and understandable way (Cleared for takeoff = permission to enter a runway and commence takeoff)

• • Signaling discourse organization (on the other hand = conversely)

In addition to this list, it has been suggested that formulaic language has another benefit: It helps language users be more fluent. This idea was first outlined by Pawley and Syder (Reference Pawley, Syder, Richards and Schmidt1983; followed closely by Kuiper & Haggo, Reference Kuiper and Haggo1984). They argued that native speakers have cognitive limitations in how quickly they can process language, presenting evidence that the largest unit of novel discourse that native speakers can process is a single clause of 8–10 words. Nevertheless, these speakers are also able to produce language seemingly beyond this limitation, for example:

This phrase is clearly beyond the limit of 8–10 words, yet native speakers can say it without hesitation. This kind of evidence led to Pawley and Syder's argument that the storage of formulaic language in abundant or unconstrained long-term memory can compensate for limited working memory.

Some of the earliest empirical evidence supporting Pawley and Syder's (Reference Pawley, Syder, Richards and Schmidt1983) assertions came from Dechert (Reference Dechert, Faerch and Kasper1983), who found that the spoken output of a German learner of English was smoother and more fluent when she used formulaic language. These formulaic sequences were so useful in providing a platform for more fluent and accurate output that Dechert called them “islands of reliability,” suggesting that they may anchor the processes necessary for planning and executing speech in real time.

There is now converging evidence that formulaic language is processed both more quickly and potentially differently from nonformulaic language, which makes the processing of formulaic language an increasingly interesting topic. This article will survey recent research (largely drawn from the field of psychology) into how both native and nonnative speakers process various types of formulaic language. In particular, we will focus on literal and figurative meanings of formulaic language, processing of various types of formulaic language, the relationship between corpus-extracted formulaic language and its psycholinguistic processing, and the processing/storage of individual words versus formulaic language.

FORMULAIC LANGUAGE: A FOCUS ON IDIOMS

In the formulaic sequence literature, idioms have received a fair amount of attention. One of the reasons for this is that many idioms allow for two distinct interpretations: figurative and literal.Footnote ² Much of the research on idiom comprehension in native speakers has addressed the following issues: (a) activation of idioms’ figurative versus literal meanings and (b) processing of idiomatic expressions versus novel (nonformulaic) phrases. With regard to the former, a number of models have been proposed (e.g., Bobrow & Bell, Reference Bobrow and Bell1973; Cacciari & Tabossi, Reference Cacciari and Tabossi1988; Swinney & Cutler, Reference Swinney and Cutler1979). One of the influential models of idiom comprehension is the lexical representation hypothesis by Swinney and Cutler (Reference Swinney and Cutler1979). They propose that idioms are represented in the mental lexicon much like morphologically complex words are. They argue that the computation of the literal meaning and the retrieval of the figurative one are initiated simultaneously, as soon as the first word of the expression is encountered. However, because computation of the literal meaning is more time-consuming than the retrieval of the figurative one, the latter meaning should become activated first.

Another prominent theory of idiom processing puts forward the idea of an idiomatic key, which refers to the place where the expression becomes recognizable as idiomatic (Cacciari & Tabossi, Reference Cacciari and Tabossi1988). According to this theory, dubbed the configuration hypothesis, the individual words and their literal meanings are activated until the key has been reached. Once the idiomatic key is reached, the idiomatic configuration emerges and the figurative meaning is accessed, while the literal meaning is rejected as no longer viable. Cacciari and Tabossi (Reference Cacciari and Tabossi1988), as well as Tabossi and Zardon (Reference Tabossi, Zardon, Cacciari and Tabossi1993), pointed out that this is only true in the absence of a language context, which would prepare the reader for either figurative or literal rendering. With the aid of context, an idiomatic meaning may become available earlier.

In addition to exploring the activation of literal versus figurative interpretations of idioms, researchers have looked at the processing of idioms versus novel phrases. For example, Swinney and Cutler (Reference Swinney and Cutler1979) studied native speakers and found that idioms (break the ice) were processed more quickly than nonidiomatic phrases (break the cup). The findings of Gibbs (Reference Gibbs1980), Gibbs and Gonzales (Reference Gibbs and Gonzales1985), and Van Lancker, Canter, and Terbeek (Reference Van Lancker, Canter and Terbeek1981) also suggested that idioms enjoy faster processing (in comprehension, as well as production) than matched novel strings for native speakers. The idiom decomposition hypothesis (Gibbs, Nayak, & Cutting, Reference Gibbs, Nayak and Cutting1989) addresses when the figurative meaning of conventional language is activated in relation to novel language. This theory holds that idiom processing is highly dependent on whether an idiom is decomposable (the meanings of component idiom words are related to the overall figurative meaning) or nondecomposable (there is no obvious link between the meanings of words and the overall figurative meaning). Gibbs et al. argued that only in the case of decomposable idioms (pop the question) should idioms be faster to process than their novel control phrases (ask the question), because their individual components contribute to the idiom's figurative meaning. For nondecomposable idioms (kick the bucket) where no such link exists, no processing advantage should be observed for idioms over their novel matches (fill the bucket). However, this proposition is not supported by Tabossi, Fanari, and Wolf's (Reference Tabossi, Fanari and Wolf2009) results. In a semantic judgment task, native speakers responded more quickly to decomposable and nondecomposable idioms (as well as clichés) than to matched literal expressions. This suggests that idiomatic expressions are processed more quickly than compositional language; however, whether an idiom's constituents are, or are not, related to the idiom's overall figurative meaning does not seem to affect its processing.

Despite the differences that exist among idiom theories, the evidence seems to support the view that idioms are processed more quickly than nonformulaic language by native speakers. What remains an open question is whether for native speakers the literal or figurative meaning of an idiom is activated more quickly. Another important question is how idioms are processed by nonnative speakers. Similar to native speakers, one of the issues in the literature on nonnative speakers is whether there is difference in the processing of literal versus figurative meanings.

Van Lancker-Sidtis (Reference Van Lancker-Sidtis2003) looked at whether prosodic cues helped native and proficient nonnative speakers distinguish between the two idiom interpretations. Participants listened to sentences that contained idioms used either figuratively or literally and then had to identify the intended meaning. Results suggested that prosodic cues enabled native participants to successfully differentiate between idioms used figuratively and literally, whereas even highly proficient nonnatives were unable to do so.

In a study by Cieslicka (Reference Cieslicka2006), nonnative participants listened to nondefining sentences containing familiar idioms (George wanted to bury the hatchet soon after Susan left) and performed a lexical decision on one of four targets: a word related to the idiom's figurative meaning (forgive), its control (gesture), a word related to its literal meaning (axe), or its control (ace). Faster response times to targets related to the literal meaning than to ones related to the figurative one suggest that literal idiom interpretations are activated more strongly than figurative ones. Thus, according to Cieslicka, in nonnative idiom comprehension, the literal meaning enjoys a processing advantage over the figurative meaning. However, perhaps, it is not surprising that upon hearing the word hatchet, there is a strong facilitation for the word axe, since the two words are strongly semantically related.

In a direct exploration of the reading speed of formulaic language, Conklin and Schmitt (Reference Conklin and Schmitt2008) used a self-paced line-by-line reading paradigm with native and nonnative speakers. Lines with formulaic sequences (everything but the kitchen sink) and their matched nonformulaic control strings (everything but the kitchen sink) were embedded in story passages. Participants read the passages, one line at a time, by pushing a button to bring up each new line. By comparing the timing of the button-pushes for the two types of strings, Conklin and Schmitt found that their participants read formulaic sequence more quickly than the matching nonformulaic control strings. They interpreted this as evidence that formulaic language is more easily processed than nonformulaic language, and crucially, this effect applied to proficient nonnatives as well as natives.

The use of eye-movement apparatus is an even better methodology for studying reading, because it eliminates the behavioral interface (i.e., the physical button-pushing), which can add variability to the results. Another advantage is that the text can be displayed all at once rather than word-by-word or line-by-line, which is a closer approximation of the natural reading process. In eye-movement studies, participants merely look at a computer monitor, and the apparatus tracks the movement of the eyes and determines which part of the screen (e.g., which words) are being focused upon and for how long. Underwood, Schmitt, and Galpin (Reference Underwood, Schmitt, Galpin and Schmitt2004) were the first to use this methodology to explore the recognition of formulaic sequences in texts. They embedded idioms into reading passages and then measured how often and for what duration the final words in those idioms were fixated upon (e.g., met the deadline by the skin of his teeth) by native and nonnative speakers. The results were then compared to measurements of the same words in nonformulaic contexts (e.g., the dentist looked at his teeth). This tested the hypothesis that once an idiom is recognized from the first few words, the final word will require less attention, because it is already known from familiarity with the idiom. Underwood et al. found that native speakers indeed fixated less on the terminal words in formulaic than nonformulaic contexts, and for a shorter duration. This is evidence for a processing speed advantage for formulaic sequences versus creative language. Moreover, this advantage was partially shared by proficient nonnatives as well.

Finally, Siyanova-Chanturia, Conklin, and Schmitt (Reference Siyanova-Chanturia, Conklin and Schmitt2011) carried out an eye-movement version of the Conklin and Schmitt (Reference Conklin and Schmitt2008) study with native and nonnative participants. Stimuli similar to the Conklin and Schmitt passages were used, containing idioms (left a bad taste in my mouth) and matched control phrases (the bad taste left in his mouth). Whole passages were presented on the monitor screen, and the participants’ eye movements were tracked while they read these passages. The eye-movement analyses of first-pass reading time, total reading time, and number of fixations showed that the native-speaking participants processed the idioms significantly faster than the nonformulaic controls. For the nonnative participants, there was no evidence that the idioms were processed any faster than the matched controls; on the contrary, the figurative readings seemed to be read more slowly than literal readings.

Taken together, the research is mixed on whether nonnatives process idioms faster than matched novel strings. However, the findings reported here suggest that even for highly proficient nonnative speakers, processing may be slowed when idioms are used figuratively. Much of the research on formulaic sequence processing has been focused on idioms, and this is problematic for a number of reasons. First, idioms, while salient, tend not to be very frequent, and therefore nonnative speakers and first language (L1) children may not have much exposure to them. Thus they may not be the best test case for whether learners are sensitive to regularly occurring patterns in language. Second, the meaning of idioms can be transparent (or decomposable) to varying degrees. Sometimes the individual component words contribute overtly to the overall meaning of the expression (add fuel to the fire = make bigger) and sometimes not (kick the bucket ≠anything to do with kicking or buckets). Thus other kinds of consistently transparent formulaic sequences might provide a better test case for whether or not frequently occurring expressions that can be computed online actually are. Third, because many idioms have two possible senses—figurative and literal—they are ambiguous, and the processing system has to select from competing semantic representations that could slow processing. Thus, while idioms have been widely studied, other formulaic sequences may provide a more accurate picture about formulaic processing.

PROCESSING OF NONIDIOMATIC FORMULAIC LANGUAGE

The position that formulaic language has a processing speed advantage is now supported by a number of research approaches. For example, Kuiper (Reference Kuiper1996, Reference Kuiper and Schmitt2004) looked at actual language use in the real world and found that the speech of so-called smooth talkers (people who need to produce fluent speech under severe time pressure, such as auctioneers and sports announcers) was largely formulaic in nature. In another study, grammaticality judgments by L2 speakers for formulaic items were not only more accurate but also faster than judgments for matched nonformulaic control strings (Jiang & Nekrasova, Reference Jiang and Nekrasova2007).

Sosa and MacFarlane (Reference Sosa and MacFarlane2002) used an auditory word-monitoring task for the function word of in two-word combinations varying in frequency (sort of and kind of) with native speakers. They found that reaction times of of in high-frequency combinations were significantly slower than those in low-frequency ones, indicating that very frequent combinations are stored as wholes. The slowed responses observed for kind of (high-frequency) over piece of (low-frequency) could not simply be due to phonological reduction (i.e., kinda for kind of), as phonological reduction was equally prevalent across the frequency groups for their stimuli. Further, the number of correct responses produced was very low for the high-frequency collocations (37 percent compared to 60 percent of correct responses produced for the low-frequency group), which supports the view of holistic storage of frequently occurring multiword units. Sosa and MacFarlane argued that their results indicated that when such multiword units were used frequently, they became chunked and were subsequently stored as a unit.

Bod (Reference Bod2000, Reference Bod2001) investigated the processing of frequently occurring sentences by native speakers. Bod's participants read frequent three-word SVO (subject-verb-object) sentences (e.g., I like it) and low-frequency control sentences (e.g., I keep it) that were matched for lexical frequency, complexity, and surface structure. Participants responded faster to high-frequency sentences than low-frequency ones, suggesting that frequently occurring sentences are stored holistically in long-term memory. However, the processing cost found for the less frequent sentences may be a consequence of these phrases being less natural. Because of the tense and aspect of some of the low-frequency experimental items, they may have sounded less natural to the participants than high-frequency ones (e.g., I test it sounds less natural than I love it because it's encountered more often in the past and future tense and in present progressive; e.g., I tested it / I will test it / I'm testing it). Thus it is difficult to say whether the processing cost observed for the low-frequency phrases over the high-frequency ones is due to properties of the stimuli or to the holistic storage of the high-frequency sentences.

In a similar study, Arnon and Snider (Reference Arnon and Snider2010) investigated the role of frequency in the comprehension of compositional four-word phrases (i.e., the phrase is comprehensible from the meanings of the individual words, e.g., don't have to worry) with native speakers. They compared reading times for phrases, which differed in phrasal frequency but whose individual components were controlled for length and frequency. They found that the more frequent phrases were processed reliably faster than the less frequent ones. The authors concluded that language users appear to notice, learn, and subsequently store frequency information not only about words but also with regard to multiword phrases, even when they are entirely compositional. Although informative with respect to the role of phrasal frequency, Arnon and Snider's study is limited to highly compositional phrases that are rather different from highly familiar fixed or semifixed formulaic sequences, such as binomial expressions or collocations.

Bannard and Matthews (Reference Bannard and Matthews2008) compared monolingual children's production of phrases that differed in the frequency with which they appeared in child-directed speech (e.g., a drink of milk vs. a drink of tea). Young children (two–three years old) were found to be reliably faster and more accurate at repeating higher frequency phrases than lower frequency ones. This shows that children as young as two years are sensitive to the frequency with which multiword strings occur in their input, and that frequent phrases like a drink of milk are represented in young children's lexicon.

Tremblay and Baayen (Reference Tremblay, Baayen and Wood2010) used phrase recall and electrophysiological (ERP) measures to investigate the processing of four-word sequences (in the middle of) by native speakers. They found that the frequency of occurrence of the four-word sequences improved participants’ recall and that whole-string probability modulated P100 and N100 amplitudes,Footnote ³ which are usually associated with the perception of input and typically occur before the onset of semantic/syntactic processing. Although these results were taken to suggest that multiword forms are stored both as parts and wholes, it is unclear whether they support such a view or instead are indicative of more general attentional processing, as these early ERP components have been implicated in attention. Similarly, in a self-paced reading study, Tremblay, Derwing, Libben, and Westbury (Reference Tremblay, Derwing, Libben and Westbury2011) compared the processing of sentences containing lexical bundles (don't worry about it) and matched control phrases. They found that sentences containing lexical bundles were read faster than control sentences, and were more likely to be remembered and recalled correctly than sentences with novel phrases. Taken together, these findings may suggest that the more frequent lexical bundle is, the more likely it is to be represented as a chunk in memory, which eases the burden on working memory during initial processing and subsequent recall.

A recent eye-tracking study by Siyanova-Chanturia, Conklin, and van Heuven (Reference Siyanova-Chanturia, Conklin and van Heuven2011) investigated processing by native and nonnative English speakers of formulaic sequences imbedded in sentences that differ in phrasal frequency. Participants read sentences containing three-word binomial phrases (bride and groom) and their reversed forms (groom and bride), which are identical in syntax and meaning but that differ in phrasal frequency. Mixed-effects modeling revealed that native speakers and nonnative speakers, across a range of proficiencies, read more frequent formulaic sequences more quickly than less frequent ones. Furthermore, regardless of frequency, the typical binomial configurations were processed more quickly than the reversed forms, which indicates that something more than a pure frequency effect is influencing the processing. Crucially, this provides strong evidence that binomial phrases are entrenched in memory in some way.

Although they are not typically considered formulaic language, research on compounds (chalkboard) can inform the ongoing debate regarding the trade-off between the storage and computation of multimorphemic words. Badecker (Reference Badecker2001), Badecker and Allen (Reference Badecker and Allen2002), Juhasz (Reference Juhasz, van, Fischer, Murray and Hill2007), and Libben (Reference Libben1998) took a compositional approach to the processing of compounds. They argued that compounds are decomposed during their recognition. In contrast, Pollatsek, Hyona, and Bertram (Reference Pollatsek, Hyona and Bertram2000) proposed that individual words (e.g., blue and berry), as well as compounds (e.g., blueberry), are stored in the lexicon, and that access to a compound can occur via the individual words or via the holistic representation.

Likewise, Mondini, Jarema, Luzzatti, Burani, and Semenza (Reference Mondini, Jarema, Luzzatti, Burani and Semenza2002) and Mondini, Luzzatti, Saletta, Allamano, and Semenza (Reference Mondini, Luzzatti, Saletta, Allamano and Semenza2005) maintained that retrieval of a compound entails the activation of its individual components, as well as the whole form of the compound. Mondini et al. (Reference Mondini, Jarema, Luzzatti, Burani and Semenza2002) investigated processing of compounds of the type Adj + N and N + Adj (e.g., natura morta – “still life”) and matched novel combinations (e.g., natura bella—“beautiful nature”) by two nonfluent aphasic patients. In Italian, adjectives agree with the grammatical gender of the noun in both compounds and novel combinations. Mondini et al. (Reference Mondini, Jarema, Luzzatti, Burani and Semenza2002) hypothesized that if compounds are stored holistically, then participants should have difficulty making noun-adjective agreement for novel combinations (because of their language impairment), but not compounds (which can be retrieved whole from memory). They found that both participants performed significantly better on compounds than on novel noun-adjective combinations. This suggests that for novel combinations the participants retrieved the adjective and noun separately and then applied agreement rules. Compounds, on the other hand, were retrieved as wholes, and therefore, no morphosyntactic operations were necessary. Interestingly, one of the participants was also able to repeat compounds significantly more accurately than noncompounds. According to Mondini et al. (Reference Mondini, Jarema, Luzzatti, Burani and Semenza2002), this implies that compounds require less working memory, which provides further evidence that they are stored holistically. Such results suggest that compounds are stored and processed as wholes, rather than computed online. This may also be true of other multimorphemic combinations that are hypothesized to be stored and retrieved as lexical units such as collocations, binomials, and idioms. Later, we turn to studies that address the issue of multiword storage and representation in the mental lexicon.

Other studies with impaired participants have also contributed to our knowledge of formulaic sequence processing. Van Lancker and Kempler (Reference Van Lancker and Kempler1987) investigated the processing of familiar phrases (e.g., idioms, speech formulas, etc.) and novel phrases by left- and right-brain damaged participants using a picture-matching auditory comprehension task. The authors predicted a larger role of the right hemisphere in familiar phrase processing, as it has been implicated in the processing of idiomatic, nonliteral language. Their results revealed that despite impaired syntactic processing, the left-brain damaged group performed significantly better on familiar phrase recognition than the right-brain damaged group. The latter group, on the other hand, performed better in the novel phrase recognition task. The authors concluded that familiar phrases such as idioms, collocations, routines, and clichés are represented in the brain differently from the novel language. Following up on this finding, Van Lancker-Sidtis and Postman (Reference Van Lancker-Sidtis and Postman2006) examined occurrences of nonnovel expressions in the spontaneous speech of normal, right- and left-hemisphere damaged participants, respectively. They found that left-hemisphere damaged participants used significantly more fixed expressions than the normal control group, whereas right-hemisphere damaged participants produced fewer fixed expressions than the control group. This, the authors argued, provides support for the view that novel language is left-hemisphere lateralized, while fixed expressions are right-hemisphere lateralized.

THE RELATIONSHIP BETWEEN CORPUS-EXTRACTED FORMULAIC LANGUAGE AND ITS PSYCHOLINGUISTIC PROCESSING

Identification of the formulaic sequences used in many of the studies reviewed in this chapter have relied on corpus evidence. This makes the relationship between the two (formulaic sequences extracted from corpora and their psycholinguistic bases in the mind) a very interesting issue. Some scholars believe that collocation is an entirely textual phenomenon and is not indicative of how language is represented in the mind (e.g., Bley-Vroman, Reference Bley-Vroman2002). They believe that collocations arise spontaneously in text as an epiphenomenon of the meaningful use of language in context, rather than being linguistic patterns that can be learned and used. For example, the words dark night occur together simply because nights are dark, and so people will naturally use these words together when speaking about the nighttime. However, given all the evidence for the processing advantages of formulaic language in this chapter, it is difficult to believe that it does not somehow exist in the mind (for an alternative view, see Ellis Reference Ellis2002a, Reference Ellis2002b).

We found only one study that directly explored whether the formulaic sequences extracted from corpus data are psycholinguistically real. Schmitt, Grandage, and Adolphs (Reference Schmitt, Carter and Schmitt2004) identified a number of different types of formulaic sequence from corpus evidence and embedded them in a spoken dictation task with native and nonnative English speakers. Each burst of dictation was longer than short-term memory could hold (i.e., 20–24 words), so the respondents were not able to repeat a burst from rote memory. This meant they were forced to reconstruct the language. The researchers assumed that if the formulaic sequences in the bursts were stored holistically, then they would be reproduced intact, with no hesitation pauses or transformations. The results showed that many of the formulaic sequences in the dictation responses did meet this holistic criterion, but also that many did not. A sort of continuum of holisticness seemed to emerge. The authors concluded that many of the corpus formulaic sequences were not stored holistically, but that this varied from individual to individual. From this one study, it seems that any particular formulaic sequence extracted from a corpus may or may not be stored holistically by any particular person.

WHAT IS REPRESENTED IN THE LEXICON?

The body of research reviewed in this chapter points to the fact that adult native speakers, and most likely children and nonnative speakers who have had enough exposure to a language, appear to have representations not only for the words that make up formulaic sequences (fish, and, chips) but also for the sequence itself (fish and chips), which is in line with the view put forward by Wray (Reference Wray2008). Frequency seems to lead to a particular form being represented in the mental lexicon. However, if a form has not been encountered frequently enough, as in the case of lower proficiency nonnative speakers or very young children, it appears that it may not be well entrenched in memory.

The finding that the frequency of a formulaic sequence affects the ease of processing is of importance for models of language use and processing. In the words-and-rules approach, a distinction is made between the lexicon (a collection of memorized and stored forms) and grammar (a collection of rules that are applied to these forms; Pinker, Reference Pinker1999; Pinker & Ullman, Reference Pinker and Ullman2002). In line with this approach, frequency effects should only be observable in the processing of memorized forms (words). Researchers argue that frequency effects should not manifest themselves in the processing of compositional formulaic sequences. Thus, the words-and-rules approach is incompatible with the results reported in this chapter unless the lexicon is reconceived to include all formulaic sequences and not just idiomatic, nonliteral language.

However, usage-based (Bybee, Reference Bybee1998; Goldberg, Reference Goldberg2006; Tomasello, Reference Tomasello2003) and exemplar-based models (Abbot-Smith & Tomasello, Reference Abbot-Smith and Tomasello2006; Bod, Reference Bod2006; Pierrehumbert, Reference Pierrehumbert, Bybee and Hopper2001) propose that the basic unit of language acquisition is a construction and that the task of a language learner is to acquire a set of constructions that vary in size, complexity, and level of abstractness (Goldberg, Reference Goldberg2006; Tomasello, Reference Tomasello2003). These theories propose that all linguistic information is represented and processed in the same way, and thus it should be similarly affected by frequency. New experiences with a linguistic unit, that is, a word or a phrase, are not decoded and then discarded; rather, they determine memory representations (Bybee, Reference Bybee2006). As Bod (Reference Bod2006) noted, what is represented is based solely on statistics. Thus, language should be viewed not as a set of grammar rules, but as a statistical accumulation of experiences that changes every time a particular utterance is encountered (e.g., Ellis, Reference Ellis2002a). This view predicts faster processing for all frequent sequences—words and phrases—over less frequent ones and is compatible with connectionist approaches to language acquisition and processing, which emphasize statistical properties of the input in language learning (Christiansen & Chater, Reference Christiansen and Chater1999; Elman, Reference Elman1990; Rumelhart & McClelland, Reference Rumelhart, McClelland, Rumelhart and McClelland1986). In a connectionist approach, units do not exist in isolation; rather, they form and exist in relationships (networks) with each other. The frequency with which various linguistic exemplars occur together is a determining factor of the strength of the connections in the lexicon. It determines what and how speakers learn and eventually represent in their lexicon.

The results reviewed above indicate that formulaic sequences are represented in the lexicon and are processed faster than novel language. However, what it means to be represented in the lexicon and what underpins the processing advantage is unclear. One might argue that words that occur together frequently have strong connections. Thus when readers encounter fish and, activation quickly spreads to chips. Alternatively, faster processing of formulaic language could be explained probabilistically. In probabilistic models of language processing, information about word co-occurrences forms an integral part of speakers’ knowledge of language (e.g., Jurafsky, Reference Jurafsky1996; McDonald & Shillcock, Reference McDonald and Shillcock2003). The probability of chips occurring after fish and is higher than the probability of fish appearing after chips and. Because fish and chips is a frequent expression, whereas chips and fish is not, one might therefore expect to see chips after reading fish and, which should facilitate reading; no such expectations may exist for the reverse. Thus, the processing advantage for formulaic language may be due its predictability.

To specifically address the question of predictability, Siyanova-Chanturia, Conklin, and van Heuven (Reference Siyanova-Chanturia, Conklin and van Heuven2011) had participants perform a completion test for both binomials and their reversed forms (fish and . . . vs. chips and . . .). They then looked at whether scores on the completion test predicted reading times in their eye-tracking study. The completion test did not significantly add anything to their mixed-effects models of reading times. Importantly, the analyses revealed that predictability and phrase type (binomial vs. reversed form) were not entirely the same. Once test scores from phrase type were accounted for in the model, phrase type still had a significant effect. This shows rather convincingly that the processing advantage for familiar phrases extends beyond the “first word+and” (e.g., fish and) predicting the last one (e.g., chips). Rather, their results indicate the important contribution of phrasal frequency and entrenchment of a particular phrase in memory. Finally, it is worth pointing out that the predictability story per se does not go against a representational account: Each and every instance of a formulaic sequence (e.g., idiom, binomial, compound) is a highly predictable word combination in which subsequent words can be predicted from an initial one(s). Thus, being predictable is an intrinsic characteristic of a formulaic sequence. It appears that speakers are sensitive to the frequency with which formulaic sequences occur, and this leads to their entrenchment in memory.