It is now widely accepted that language is largely formulaic in nature, and that phraseological competence is an important part of nativelike, fluent, and idiomatic language use (Hoey, Reference Hoey2005; Pawley & Syder, Reference Pawley, Syder, Richards and Schmidt1983; Wray & Perkins, Reference Wray and Perkins2000). It is also becoming increasingly apparent that phraseology plays an important role in foreign language learning and teaching (Ellis, Reference Ellis, Doughty and Long2003; Howarth, Reference Howarth1998; Lewis, Reference Lewis1993; Nattinger & DeCarrico, Reference Nattinger and DeCarrico1992). Just as “phraseology binds words, grammar, semantics, and social usage” (Ellis, Reference Ellis, Meunier and Granger2008, p. 5), the use of phraseological units such as collocations, phrasal verbs, compounds, and idioms may impact positively or negatively on the three dimensions of language proficiency—complexity, accuracy, and fluency (Housen & Kuiken, Reference Housen and Kuiken2009). Osborne (Reference Osborne, Meunier and Granger2008), for example, identified some phraseological effects that trigger some categories of persistent learner errors such as the omission of third-person –s or the pluralization of adjectives.
The last few years have seen a proliferation of phraseological studies based on learner corpora, which can be roughly defined as electronic collections of texts produced by foreign or second language (L2) learners. As learner language is highly heterogeneous—there are many types of learners and learning situations—learner corpus data are compiled on the basis of strict design criteria, pertaining to the learner (e.g., age, gender, mother-tongue background, number of years spent studying the foreign language, or time spent in a country where the foreign language is spoken) and the task (e.g., medium, topic, timing, or the use of reference tools). Learner corpora have two distinguishing characteristics that make them an ideal source of data to study the learner phrasicon, that is, the whole set of formulaic sequences in learner language: (a) They contain continuous stretches of oral or written discourse rather than decontextualized words, phrases or sentences, and (b) they typically include data resulting from pedagogical tasks that allow learners to choose their own wording (e.g., essay writing, learner interviews; cf. Granger, Reference Granger, Deusen-Scholl and Hornberger2008; Granger, in press) rather than being requested to produce a particular word or structure. In addition, as the data are in electronic format, analysts can use a whole range of automated methods and tools to identify and analyze formulaic sequences in learner language.
This chapter aims to provide a synthesis of current research on formulaic language in learner corpora of L2 English (i.e., English as a foreign or second language). We first give an overview of the learner corpora used to study the learner phrasicon and describe the most popular method used to analyze the data. We then review the types of formulaic sequences that have been the focus of learner corpus research (LCR) so far and describe how they are identified and analyzed in learner corpora. Next, we present some of the major findings of LCR studies of the learner phrasicon, distinguishing between studies of co-occurrence and recurrence. This section puts particular emphasis on one of the main findings that LCR has helped establish, namely, the huge impact of the first language on the learner phrasicon. The chapter concludes with some proposals for future research in the field.
LEARNER CORPUS DATA USED
Studies of the learner phrasicon have utilized a range of data types that differ widely in terms of both learner and task variables. This heterogeneity makes it very difficult to compare the results. However, in spite of this diversity, it is possible to identify some general trends. A large majority of the studies have investigated the use of formulaic language by English as a foreign language (EFL) rather than English as a second language (ESL) learners, and the level of proficiency targeted has more often been upper-intermediate to advanced learners rather than beginners. Most studies have focused on learner writing, and more particularly argumentative essay writing (e.g., Nesselhauf, Reference Nesselhauf2005; Waibel Reference Waibel2008). However, the last few years have witnessed a marked increase in studies of formulaic language in learner speech (e.g., Aijmer, Reference Aijmer, Jucker, Schreier and Hundt2009; Crossley & Salsbury, Reference Crossley and Salsbury2011; De Cock, Reference De Cock, Frankenberg-Garcia, Flowerdew and Aston2011; Götz & Schilk, Reference Götz, Schilk, Hundt and Mukherjee2011). Although a number of longitudinal studies have recently been published (e.g., Crossley & Salsbury, Reference Crossley and Salsbury2011; Li & Schmitt, Reference Li, Schmitt and Wood2010), most studies are based on data collected at a single point in time (e.g., Chen & Baker, Reference Chen and Baker2010; Durrant & Schmitt, Reference Durrant and Schmitt2009; Gilquin, Reference Gilquin2007).
Many studies of the learner phrasicon make use of subcomponents of the International Corpus of Learner English (ICLE), one of the first learner corpora to be made available for research purposes (Granger, Dagneaux, & Meunier, Reference Granger, Dagneaux and Meunier2002). The ICLE is an ongoing project and now consists of about 3.7 million words of (mostly) argumentative essay writing by learners from 16 mother-tongue backgrounds representing several language families (including Romance, Germanic, and Asian languages; Granger, Dagneaux, Meunier, & Paquot, Reference Granger, Dagneaux, Meunier and Paquot2009). Other corpora of EFL learner writing used include the Uppsala Student English Corpus (about 1.2 million words), the Active Learning of English for Science Students (ALESS) corpus (about 700,000 words) and the Cologne-Hanover Advanced Learner Corpus (about 200,000 words). The recent surge of interest in the learner spoken phrasicon has occurred in parallel with the development of ICLE's spoken counterpart, the Louvain International Database of Spoken English Interlanguage (LINDSEI), a corpus of about 800,000 words of oral data produced by learners from 11 mother-tongue backgrounds (Gilquin, De Cock, & Granger, Reference Gilquin, De Cock and Granger2010).
One of the major advantages of corpora such as ICLE or LINDSEI is their size. Not only does a large corpus increase the degree of representativeness of the data and the generalizability of the results, but it is also a sine qua non for phraseological studies involving low-frequency words. However, small corpora may also be of considerable value. Wang and Shaw (Reference Wang and Shaw2008), for example, examined verb + noun collocations with have, do, take, and make in two self-compiled learner corpora produced by Chinese-speaking and Swedish-speaking learners of English. The learner corpora are rather small (about 20,000 words each), but all the learner texts have the same topic as prompt (Is it true that only rich countries can afford to worry about the environment?), thus making it possible to control for the influence of the topic on lexical choice (cf. Jarvis, Grant, Bikowski, & Ferris, Reference Jarvis, Grant, Bikowski and Ferris2003). Likewise, although longitudinal studies are generally characterized by a very limited set of informants, the number of words collected for each learner in recent studies has increased considerably. Li and Schmitt (Reference Li, Schmitt and Wood2010), for example, adopted a multiple case-study approach to identify patterns of development in the use of adjective-noun collocations by four learners. They collected eight essays and one dissertation for each participant, all written over a period of one academic year, and categorized them according to the academic term during which they were written (Term one, two, or three). The size of the longitudinal corpus they compiled amounts to about 150,000 words, with about 50,000 words for each term.
Up to now, the most frequently used method of investigating formulaic language in a learner corpus has been to compare the results of a learner corpus analysis with those obtained from the analysis of a comparable native corpus and identify errors and patterns of learner over- and underuseFootnote 1 of formulaic sequences. In most cases, the learner corpora used consist of texts produced by learners who have the same mother-tongue background: Chinese (Chen & Baker, Reference Chen and Baker2010), French (Henderson & Barr, Reference Henderson and Barr2010), German (Nesselhauf, Reference Nesselhauf2005), Hebrew (Laufer & Waldman, Reference Laufer and Waldman2011), Japanese (Allen, Reference Allen2011), Lithuanian (Juknevičienė, Reference Juknevičienė2008), Russian (Siyanova & Schmitt, Reference Siyanova and Schmitt2008), Spanish (González, Reference González, Campoy-Cubillo, Bellés-Fortuño and Gea-Valor2010), or Swedish (Groom, Reference Groom, Barfield and Gyllstad2009). This comparative method, better known as contrastive interlanguage analysis (CIA) (cf. Granger, Reference Granger, Aijmer, Altenberg and Johansson1996), may also involve the comparison of different learner varieties. Typically, studies compare learner populations that differ in terms of mother-tongue background. Thus, Waibel (Reference Waibel2008) compared the use of phrasal verbs in two corpora of argumentative texts produced by German and Italian learners, respectively, to assess the impact of the first language on learners’ use of both literal and idiomatic verb-particle combinations.
TYPES OF FORMULAIC SEQUENCES INVESTIGATED AND EXTRACTION METHODS
Following a well-established tradition in applied linguistics (cf. Biskup, Reference Biskup, Arnaud and Béjoint1992; Lennon, Reference Lennon1996), the most frequently studied type of formulaic sequence in LCR is verb + noun collocations, and more particularly verb + noun collocations with high-frequency delexical verbs such as take and make. Collocations can be roughly defined as usage-determined lexical combinations that are characterized by restricted co-occurrence of elements, for example, face a problem, take a step, reach a conclusion, and adopt an approach (cf. Granger & Paquot, Reference Paquot, Meunier and Granger2008). Unlike other L2 studies, learner-corpus-based studies usually consider the whole range of verb + noun combinations produced by learners, including free combinations such as want a car and read the paper (e.g., Altenberg & Granger, Reference Altenberg and Granger2001; Cross & Papp, Reference Cross, Papp, Gilquin, Papp and Díez-Bedmar2008; Howarth, Reference Howarth1996; Kaszubski, Reference Kaszubski2000; Nesselhauf, Reference Nesselhauf2005). The line between free and restricted combinations is a fine one, and learners have been shown to have difficulties with all types of verb-noun combinations, including those that are referred to as free. As several studies (e.g., Juknevičienė, Reference Juknevičienė2008; Laufer & Waldman, Reference Laufer and Waldman2011; Nesselhauf, Reference Nesselhauf and Tschichold2003; Zinkgräf, Reference Zinkgräf2008) have demonstrated, transfer from the learners’ first language (L1) is a major determinant of deviant verb + noun collocations (see later in this article).
Another category of formulaic sequences that has attracted the attention of learner corpus researchers and proved to be particularly error-prone is that of phrasal verbs. Learners use either the right verb with the wrong particle (a task which must be carried on [carried out] using the brain) or the right particle with the wrong verb (We tried to come back to [go back to] Los Angeles; De Cock, Reference De Cock and Rundell2005). Other deviations include collocational mismatches such as carry out a race or make up a proposal and overextension of the literal meaning of phrasal verbs to figurative contexts, for example, cut down a life or weed out a crime (Waibel, Reference Waibel2008).
Many researchers have highlighted the links between formulaic language and pragmatic competence (see Bardovi-Harlig, this volume). In LCR this link is explored through investigations of speech formulae, that is, word sequences that are used to organize the speakers’ messages and “employed in organizing turn-taking, indicating a speaker's attitude to other participants, and generally ensuring the smooth conduct of interaction” (Cowie, Reference Cowie, Cater and McCarthy1988, p. 133). For example, Aijmer (Reference Aijmer, Jucker, Schreier and Hundt2009) compared the use of I don't know (or ‘[I] dunno’) in the Swedish component of the LINDSEI corpus and a comparable native speaker corpus, and shows that this formula does not have the same function in native and learner speech. Swedish EFL learners overwhelmingly use I don't know as a speech management signal, whereas native speakers use it mainly to avoid asking questions directly.
While the pragmatic properties of formulaic sequences have long been recognized, it is mainly their role in speech that has been investigated. The part they play in writing has only recently been acknowledged. Corpus linguistic tools and methods have been instrumental in establishing the links between formulaic language and text structuring (e.g., Biber & Barbieri, Reference Biber and Barbieri2007; Conrad & Biber, Reference Conrad and Biber2004; Simpson-Vlach & Ellis, Reference Simpson-Vlach and Ellis2010). They helped identify a whole range of “regular form-meaning pairings” (cf. Pawley & Syder, Reference Pawley, Syder, Richards and Schmidt1983, p. 192) that are commonly used to organize the content of a text, fulfill rhetorical functions as diverse as introducing a topic, comparing, expressing a cause, summarizing, and concluding (e.g., as discussed in, it has been suggested, the main objective of, as a result of), or engage the reader in the argumentation process (e.g., it is possible, it is clear that). This range of textual or discourse-oriented sequences has recently been investigated in corpora of learner writing with a view to identifying learners’ difficulties and informing the teaching of and material development for English for Academic Purposes (e.g., Gilquin, Granger, & Paquot, Reference Gilquin, Granger and Paquot2007; Luzón Marco, Reference Luzón Marco, Campoy-Cubillo, Bellés-Fortuño and Gea-Valor2010; Neff van Aertselaer, Reference Neff van Aertselaer, Meunier and Granger2008; Paquot, Reference Paquot, Meunier and Granger2008, Reference Paquot2010).
While many different techniques have been used to extract formulaic language from learner corpora, they all have something in common; namely, they all involve some degree of automation. The most basic way of retrieving formulaic sequences is to make use of the concordancing option of a corpus-handling tool, that is, a program that presents all the instances of a linguistic item in their immediate context. Each occurrence is displayed on a single line, with the search item in the middle and the context on each side. This method has been used to extract all occurrences of continuous word combinations (e.g., I don't know; Aijmer, Reference Aijmer, Jucker, Schreier and Hundt2009) or all forms of a particular word (e.g., the verb make; Nesselhauf, Reference Nesselhauf2005; the preposition out; González, Reference González, Campoy-Cubillo, Bellés-Fortuño and Gea-Valor2010). Sorting the context to the left and/or to the right of the search item allows regular patterns to emerge.
Many corpus-handling tools contain a function that allows for the automatic retrieval of collocates, that is, the words that occur most frequently to the right and/or left of the search items. Lorenz (Reference Lorenz1999) probably carried out the first learner-corpus-based study that focused on statistically significant co-occurrences. He made use of the mutual information (MI) statistic to identify strongly associated intensifier-adjective combinations (e.g., heavily loaded, deeply rooted, highly estimated) in a corpus of essays produced by German learners of English. The MI score highlights co-occurrences that are not especially frequent, but are closely associated, and are thus likely to be very salient for native speakers (cf. Simpson-Vlach & Ellis, Reference Simpson-Vlach and Ellis2010). This extraction method, based on statistically significant co-occurrences, has seldom been used in LCR, probably because association measures tend to become unreliable with low-frequency data (e.g., Evert, Reference Evert, Lüdeling and Kytö2008), and learner corpora have until very recently been of limited size. However, two recent studies should be mentioned here. Groom (Reference Groom, Barfield and Gyllstad2009) made use of two association measures, the t score and MI score, to extract significant co-occurrences of the 10 most frequent prepositions from two subcorpora of the Uppsala Student English Corpus. Although the corpora investigated were not very large (about 250,000 words each), Groom's focus on very frequent prepositions made the use of association measures possible. Fan (Reference Fan2010), on the other hand, used a larger corpus (1 million words) and relied on the MI to measure the collocational strength between two much less frequent lexical items—the verbs “cause” and “lead to”—and their co-occurrences in the writing of Chinese learners.Footnote 2
A more common approach to measuring collocational strength in learner language is to make use of association measures computed from reference native or expert corpora, rank co-occurrences in learner corpora on that basis, and judge their acceptability. Durrant and Schmitt (Reference Durrant and Schmitt2009), for example, made use of the British National Corpus to calculate the t score and MI score for each adjacent premodifier-noun combination extracted from two learner corpora and two comparable corpora of native student writing (see later in this article for a summary of the results). Two other studies that use a similar methodology are Siyanova and Schmitt (Reference Siyanova and Schmitt2008) and Paquot (Reference Paquot2010).
Finally, corpus-handling tools also often provide an option that retrieves all repeated sequences of words of a given length (e.g., two-word sequences, three-word sequences) fully automatically. This method has been used extensively in corpus studies in the last few years to analyze lexical bundles, which are “recurrent expressions, regardless of their idiomaticity, and regardless of their structural status” (Biber, Johansson, Leech, Conrad, & Finegan, Reference Biber, Johansson, Leech, Conrad and Finegan1999, p. 990; e.g., as for, out of, as far as, as discussed in, take the example of, the extent to which) in various text types and genres (e.g., Biber, Conrad, & Cortes, Reference Biber, Conrad and Cortes2004; Hyland Reference Hyland2008; Stubbs & Barth, Reference Stubbs and Barth2003). This approach represents “corpus linguistic methodology at its most heuristic, i.e., as a raw discovery procedure” (De Cock, Reference De Cock2004, p. 227). It has highlighted the pervasiveness of a wide range of phraseological units that do not fit the categories that have traditionally been focused on (e.g., similes, phrasal verbs, or idioms; cf. Granger & Paquot, Reference Granger, Paquot, Granger and Meunier2008). Examples of studies that analyze learners’ use of lexical bundles include Allen (Reference Allen2011), Chen and Baker (Reference Chen and Baker2010), Crossley and Salsbury (Reference Crossley and Salsbury2011), De Cock (Reference De Cock, Campoy and Luzón2007), and Götz and Schilk (Reference Götz, Schilk, Hundt and Mukherjee2011).
MAJOR FINDINGS
The learner-corpus-based studies that focus on the phraseology of L2 speech or writing have yielded a wealth of interesting findings. However, two factors make it difficult to form a coherent picture. First, the studies focus on different categories of learners (e.g., EFL or ESL learners, with varying degrees of proficiency, different mother-tongue backgrounds) and the results are therefore not directly comparable, especially when only a minimum of ethnographic data is provided. Second, the types of word combinations investigated differ, as do the methodologies used to extract and analyze them. Any survey of the literature highlights what may, at first sight, look like contradictions, but is in reality nothing but a consequence of the considerable heterogeneity in both data and methods.
In this section we focus on the general trends which emerge from studies that are explicit about their methodologies and therefore allow meaningful comparisons of their results. We make a sharp distinction between two aspects of phraseology: co-occurrence and recurrence. These need to be considered separately as they tap different competencies and therefore pose different challenges for language learners. Special attention is devoted to transfer from the learners’ mother tongues, which has been shown to play a key role in learners’ phrasicons.
Co-occurrence
Co-occurrence is one of the two main types of patterns that lie at the heart of phraseology. It consists in the co-selection of (usually) two lexical items, which may be, but are not necessarily, contiguous. It refers to the preference that words have for certain co-occurring words, a phenomenon that Firth (Reference Firth and Firth1957) referred to as “mutual expectancy” (p. 179). In traditional phraseology, linguistically defined subsets of co-occurring items are referred to as collocations or restricted collocations, that is, lexically constrained combinations that allow for limited substitution within a particular grammatical construction (e.g., verb-object, adverb-adjective, or adjective-noun). For example, in English you can perform a task or do a task but you cannot make a task. These lexical restrictions are arbitrary and differ markedly from one language to another. They are one of the main stumbling blocks for learners, up to and including the most advanced ones. Restricted collocations are opposed to free combinations that are only constrained by semantic and grammatical constraints (e.g., you can drink coffee and any other liquid, but you cannot drink solids like bread).
In the new frequency-based approach to phraseology initiated by Sinclair (Reference Sinclair1987), the term collocation is used differently: It refers to statistically significant word co-occurrences, that is, lexical items occurring within a certain distance of the search item “with a greater frequency than the law of averages would lead you to expect” (p. 70). Collocations are extracted with the help of statistical methods (such as MI) rather than linguistic criteria, and they may therefore include not only collocations in the traditional sense but also combinations such as drink coffee that would be categorized as free combinations in traditional phraseology. To avoid any misunderstanding, we will only use the term collocation in its traditional meaning of a usage-based lexically restricted combination and will use co-occurrence to refer to the more general phenomenon of co-selection of lexical items or to statistically defined combinations.
The overall picture that emerges from learner-corpus-based studies is that learners’ use of co-occurring combinations is characterized by a mixture of underuse, overuse, and misuse. Laufer and Waldman (Reference Laufer and Waldman2011) recently showed that Hebrew-speaking EFL learners produced almost half as many verb-noun collocations as young adult native speakers. Other studies (e.g., Kaszubski, Reference Kaszubski2000) have shown that this general underuse goes hand in hand with an overuse of some collocations, especially very frequent collocations such as have a look, do work, or take care of. A possible explanation for learners’ heavy reliance on these collocations, which Nesselhauf (Reference Nesselhauf2005) referred to as “collocational teddy bears,” is that the learners feel confident using them (p. 69).
This combination of overuse of high-frequency collocations and underuse of less frequent collocations tallies with results from statistical studies of co-occurrence in learner corpora. Durrant and Schmitt (Reference Durrant and Schmitt2009), for example, showed that nonnative writers underuse less frequent, strongly associated premodifier-noun word pairs (i.e., relatively low-frequency co-occurrences identified by MI scores), which are probably highly salient for native speakers. However, they make at least as much use of high-frequency co-occurrences (as identified by high t scores) such as good example, long way, or hard work. As noted by Durrant & Schmitt (Reference Durrant and Schmitt2009), “This is an intuitively satisfying result: learners are quick to pick up highly frequent collocations, but less common, strongly associated items (e.g., densely populated, bated breath, preconceived notions) take longer to acquire” (p. 175).
Collocations are also particularly error-prone. Nesselhauf (Reference Nesselhauf2005) made use of dictionaries, corpora, and native speaker informants to investigate the degree of acceptability of around 2,000 verb-noun collocations produced by German EFL learners and found that approximately one third could be considered unacceptable or questionable. Similar figures were reported by Laufer and Waldman (Reference Laufer and Waldman2011) in their investigation of collocations in the writing of native speakers of Hebrew: Deviant collocations accounted for about a third of all the verb-noun collocations learners attempted to produce. Nesselhauf (Reference Nesselhauf2005) showed that the most frequent type of errors in verb-noun combinations involved the wrong choice of verb (e.g., *carry out races). Among the nouns that were most often used with deviant verbs were the signaling nouns action, aim, attitude, problem, question, statement, step, and conclusion. Other types of errors included prepositional errors (e.g., *lose control over), determiner errors (e.g., *get the permission), and completely wrong combinations (e.g., *hold children within bounds). Interestingly, Nesselhauf's study shows that verbs are not only frequently misused in collocations but also appear to be a major source of error in free combinations (e.g., *plead for emancipation [fight for emancipation]; cf. Nesselhauf, Reference Nesselhauf2005).
Studies such as Thewissen (in press) have shown that collocations remain an error-prone area at advanced proficiency levels. However, it is difficult to establish a clear link between language proficiency and phraseological competence, as proficiency is variously defined in terms of the number of years of study, the learner's age, or the Common European Framework of Reference (CEFR) levels. As a result, the studies may not be strictly comparable. However, the following two trends emerge: Higher proficiency in the language is usually characterized by a higher rate of use of collocations and phrasal verbs, but this quantitative gain is not paralleled by a qualitative gain, as many of the collocations continue to be used inaccurately.
Several studies (Hsu Reference Hsu2007; Lewis, Reference Lewis2009; Ohlrogge, Reference Ohlrogge, Corrigan, Moravcsik, Ouali and Wheatley2009) have provided evidence of a greater use of collocations, phrasal verbs, and formulaic sequences in better-rated pieces of writing. Laufer and Waldman (Reference Laufer and Waldman2011) found a statistically significant difference between the number of verb-noun collocations used by advanced and elementary students. However, there were no significant differences between intermediate learners and either of the other two groups, which suggests that progress in this area of language is relatively slow. This may explain why Li and Schmitt's (Reference Li, Schmitt and Wood2010) investigation of Chinese learners showed relatively little change in the production of adjective-noun co-occurrences over the course of a year, a period that may not be sufficient for significant improvement to take place. Interestingly, however, the study indicated a high degree of variability between learners, with some students producing more collocations after one year while others produced fewer (see also Ohlrogge, Reference Ohlrogge, Corrigan, Moravcsik, Ouali and Wheatley2009).
Taking a slightly different perspective, Groom (Reference Groom, Barfield and Gyllstad2009) and Waibel (Reference Waibel2008) both highlighted the impact of immersion-based L2 exposure on phraseological competence. Groom (Reference Groom, Barfield and Gyllstad2009) found a high correlation between the number of statistical co-occurrences and the amount of time spent in an English-speaking country, while for Waibel (Reference Waibel2008), the length of time abroad is a significant predictor of the number of phrasal verbs used.
More does not necessarily mean better. Nesselhauf (Reference Nesselhauf2005) found that students who had studied English for 10 to 17 years displayed a similar proportion of deviant collocations as students who had only studied English for five to 10 years. Similarly, Laufer & Waldman (Reference Laufer and Waldman2011) computed the ratio of the number of errors to the number of well-formed collocations and found no relationship between this ratio and the proficiency levels of the students. About a third of all collocations were erroneous, regardless of the learners’ level. However, when the number of collocation errors was computed as a proportion of the total number of words, advanced and intermediate learners were found to produce significantly more deviant collocations than basic learners. All in all, then, there appears to be an inverse relationship between proficiency and the correctness of collocations. As suggested by Thewissen (Reference Thewissen2008), this may paradoxically be a sign of increased phraseological richness: Higher-level learners attempt a much wider range of lexical phrases but the result is not fully error-free. While there may not be much difference in the number of errors produced by more and less advanced learners, the types of errors differ. The most advanced learners produce a large number of near hits (e.g., to turn over a new leave instead of to turn over a new leaf) compared to their lower intermediate counterparts who produced a higher number of wide-of-the-mark errors, that is, errors that resulted in the use of nonexistent phrases such as the time work's duration instead of working hours.
Recurrence
Recurrence, the second main type of phraseological patterning, can be defined as the repetition of contiguous strings of words of a given length (e.g., bigrams, trigrams). These recurrent strings, which following Biber et al. (Reference Biber, Johansson, Leech, Conrad and Finegan1999) we will refer to as lexical bundles, may include grammatically complete (that's a good idea, on the other hand) and incomplete (in the case of the, it should be noted that) strings. They are “conventionalized building blocks that are used as convenient routines in language production” (Altenberg, Reference Altenberg and Cowie1998, p. 122).
The results of automatic studies of recurrence in learner language are particularly difficult to compare. The lexical bundles investigated are of different sizes (from two- to six-word bundles), and the settings used to extract them may vary considerably. For example, Reppen (Reference Reppen, Barfield and Gyllstad2009) analyzed the 20 most frequent three-word sequences in three learner corpora; Groom (Reference Groom, Barfield and Gyllstad2009) investigated all two- to five-word lexical bundles with a minimum frequency of 10 occurrences per 250,000 words; Chen and Baker (Reference Chen and Baker2010) set the frequency threshold for determining four-word lexical bundles at four occurrences or more (approximately 25 occurrences per million words). However, a number of general trends can be identified.
While they have been shown to use fewer collocations than native speakers, learners tend to use more lexical bundles in writing when compared to native speakers. Ping (Reference Ping2009), for example, reported that Chinese learners used four times as many recurrent four-word sequences as native speakers (see also Juknevičienė, Reference Juknevičienė2009; Wei & Lei, Reference Wei and Lei2011). Bo and Shutang (Reference Bo and Shutang2005) noted that while there were more three- to six-word bundles in the Chinese Learner English Corpus, this overuse was especially true of longer sequences. Overall, less proficient learners seem to be more reliant on lexical bundles. This high degree of repetition may be due to their more limited lexical repertoire, but as shown by Ohlrogge (Reference Ohlrogge, Corrigan, Moravcsik, Ouali and Wheatley2009), it may also be due to a tendency in less proficient students to copy the writing prompt. Unlike learners’ use of collocations, the overall number of lexical bundles tends to decrease as proficiency in the language (Reppen, Reference Reppen, Barfield and Gyllstad2009) or the time spent in an English-speaking country (Groom, Reference Groom, Barfield and Gyllstad2009) increases.
More fine-grained qualitative analyses show the value of grouping lexical bundles according to their grammatical structure and/or functional use. Typically, studies of lexical bundles compare the structural characteristics and discourse functions of recurrent word sequences in learner and native corpora. For example, Chen and Baker (Reference Chen and Baker2010) compared four-word sequences in a corpus of learner academic writing, a corpus of British student academic writing and a corpus of expert writing and made use of the structural classification of lexical bundles provided in the Longman Grammar of Spoken and Written English (Biber et al., Reference Biber, Johansson, Leech, Conrad and Finegan1999) to categorize them (e.g., noun-phrase with postmodifier fragment, e.g., the nature of the; preposition + noun-phrase fragment, e.g., as a result of; passive verb + PP fragment, e.g., is based on the). They then adopted the functional taxonomy of lexical bundles devised by Biber and colleagues (e.g., Biber & Barbieri, Reference Biber and Barbieri2007) and grouped word sequences into referential bundles (e.g., in the context of, in a number of, at the same time), stance bundles (e.g., are more likely to, it can be argued), and discourse organizers (e.g., in this essay I, on the other hand). This classification brought out a mixed pattern of under- and overuse. Chen and Baker found that Chinese learners made relatively little use of bundles with passive verb forms followed by a preposition (e.g., can be seen in, can be used for; see also Allen, Reference Allen2011) and quantifying bundles (e.g., in so far as, the extent to which, the degree to which, to a certain extent). They also used very few hedging expressions (e.g., are more likely to, is considered to be). Several studies (cf. Juknevičienė, Reference Juknevičienė2009; Reppen, Reference Reppen, Barfield and Gyllstad2009) have revealed an underuse of the most academic-like recurrent word combinations, such as noun phrases with postmodifier fragments (e.g., the issue of, the risk of, part of the, the idea that), coupled with an overuse of speech-like bundles, such as personal pronoun + verb (e.g., I think).
There have been relatively few studies of recurrence in learner speech. One exception is De Cock (Reference De Cock2004), who extracted recurrent strings of two to six words from the French subcorpus of LINDSEI and compared their frequency and use with that in a comparable native speaker corpus. The study showed that learners’ preferred sequences were less interactional than native speakers’ and included very few markers of vagueness (e.g., or something, and things, or anything, sort of, kind of). Learners were also shown to have a tendency to favor some rather forceful recurrent sequences such as yes/yeah of course, which may lead them to sound overemphatic or even impolite. Overall, the French learners proved to be lacking in routinized ways of interacting and building rapport with their interlocutors and of toning down and weaving the right amount of imprecision and vagueness into their speech.
Studies focusing on the use of speech formulae point in the same direction: Learners do not use all the functions of pragmatic markers or use them with different frequencies from native speakers (e.g., Aijmer, Reference Aijmer, Jucker, Schreier and Hundt2009; Müller, Reference Müller2005). A recent longitudinal study by Crossley and Salsbury (Reference Crossley and Salsbury2011) showed that the amount of time spent in an English-speaking environment can have a significant influence on learners’ phraseological competence. They showed that, as time goes by, learners begin to use two-word lexical bundles at frequencies that approximate those of L1 speakers. Taking a different perspective, Adolphs and Durrow (Reference Adolphs, Durrow and Schmitt2004) emphasized another factor that can foster the adoption of nativelike lexical bundles, namely, the quality of the students’ cultural and social integration.
The Impact of the Learner's Mother Tongue
Although the learner's mother tongue is clearly not the only variable that has an impact on the phraseological production of EFL learners, it is the factor that has been most discussed in LCR. Corpus-based studies of collocational errors tend to support the claim that the influence of their first language (L1) accounts for a large proportion of EFL learners’ inappropriate collocations. Nesselhauf (Reference Nesselhauf2005) examined all the inappropriate verb-noun collocations produced by German learners and observed that about 50 percent of them were probably influenced by German phraseology. Similarly, Laufer and Waldman (Reference Laufer and Waldman2011) reported that L1 influence was evident in about half of the erroneous verb-noun collocations produced by Hebrew learners at three levels of proficiency and did not decrease with time. Moreover, 89 percent (42 out of 47) of the recurrent errors (i.e., errors made by several learners) were word-for-word translations of Hebrew verb-noun collocations that convey the same meaning as but use different verbs from the English collocations (see also Juknevičienė, Reference Juknevičienė2008).
Words with figurative meanings have sometimes been considered less transferable than those with core meanings (cf. Kellerman, Reference Kellerman1978), a hypothesis that does not seem to be supported by computer-aided error analyses. Nesselhauf (Reference Nesselhauf2005), for example, observed that overall the majority of verbs that were likely to have been influenced by the L1 were used in a figurative sense in verb-noun combinations (e.g., take in take a step). More generally, LCR provides evidence to counter the claim that transferability decreases with the degree of idiomaticity of a verb-noun combination (cf. Kellerman, Reference Kellerman1978, Reference Kellerman1979). Rather, collocations seem to be the type of combination that is most susceptible to transfer (e.g., *make part of [Fr. faire partie de] for be part of, *make dreams [Fr. faire des rêves] for have dreams; Nesselhauf, Reference Nesselhauf and Tschichold2003, p. 278), and the degree of L1 influence is somewhat less strong for free verb-noun combinations.
More interestingly perhaps, learner-corpus-based studies have also provided more subtle evidence of L1 influence on EFL learners’ patterns of overuse and underuse. Studies of EFL learners’ use of phrasal verbs find a general underuse of phrasal verbs by learners with Romance mother tongues (Gilquin & Granger, Reference Gilquin, Granger, Hundt and Mukherjee2011; González, Reference González, Campoy-Cubillo, Bellés-Fortuño and Gea-Valor2010; Waibel, Reference Waibel2008). This is backed up by a number of experimental studies which have reported underuse of English phrasal verbs by learners whose first languages do not have this type of construction (e.g., Dagut & Laufer, Reference Dagut and Laufer1985; Laufer & Eliasson, Reference Laufer and Eliasson1993; Liao & Fukuya, Reference Liao and Fukuya2004; Sjöholm, Reference Sjöholm, Haastrup and Viberg1998) and confirm the hypothesis that “the absence of a structural feature in the L1 may have as much impact on the L2 as the presence of a different feature” (Ellis, Reference Ellis1994, pp. 311–312). Learners have also been shown to rely heavily on English collocations that are congruent with word combinations in their L1. Thus, the few adverb-adjective collocations used by French EFL learners in Granger's (Reference Granger and Cowie1998) study typically have a direct translation equivalent in French. Examples include closely linked (Fr. étroitement lié), deeply rooted (Fr. profondément enraciné), and severely punished (Fr. sévèrement puni). Similarly, Gilquin (Reference Gilquin2007) reported that French learners tend to underuse collocations with the verb make and limit themselves to those which have a direct equivalent in French (make progress = [Fr. faire des progress], make an effort = [Fr. faire un effort]) and are therefore safer (see also Altenberg & Granger, Reference Altenberg and Granger2001).
This also seems to hold true for recurrent sequences and, more particularly, discourse-oriented formulae. Neff et al. (Reference Neff, Dafouz, Herrera, Martínez, Rica, Diez, Prieto, Sancho, Granger and Petch-Tyson2003) showed that writers with an L1 Romance mother tongue relied heavily on we + modal verb + verb of mental/verbal process clusters, which correspond closely to Romance languages’ use of the first person plural to address readers. In a follow-up study, Neff et al. (Reference Neff, Ballesteros, Dafouz, Martínez, Rica, Facchinetti and Palmer2004) partly explained Spanish speakers’ use of we must + a reporting verb with an illocutionary force as a transfer from Spanish, in which the deontic verb deber can mean either must or should. More recently, Neff van Aertselaer (Reference Neff van Aertselaer, Meunier and Granger2008) hypothesized that Spanish expert and novice writers’ preference for passive structures in the present tense when they are writing in English probably reflects a transfer of Spanish discourse strategies, that is, a translation of Spanish se impersonal passive phrases, for example, it is said (Es. se dice).
Paquot (Reference Paquot2010) made use of Jarvis's (Reference Jarvis2000) unified framework for the study of L1 influence to show that transfer was at work in French-speaking learners’ overuse of a number of discourse-oriented formulae (e.g., imperative constructions with let us such as Let us take the example of or Let us now turn to, and sentence-initial to-infinitive constructions such as To illustrate this or To conclude). As these word combinations are found in English, they can be regarded as cases of positive transfer. However, these structures in learner writing often mirror the stylistic profile (i.e., register preferences, function and frequency) of their French congruent sequences and do not conform to English academic writing conventions (see also Paquot, Reference Paquot, Meunier and Granger2008).
All in all, recent learner-corpus-based studies have illustrated the importance of L1 influence and identified a number of transfer effects that have not previously been much documented, but are well worth exploring in more detail. In fact, they have shown that the influence of L1 on learners’ use of formulaic language may be much more substantial than was previously thought, and may play a role in L2 learning that we do not yet fully comprehend.
CONCLUSION
Our survey has shown that learner corpora are a powerful resource for exploring learners’ use of formulaic language. Although LCR is still a very young field, phraseological studies have been particularly numerous and have generated a wealth of interesting results. However, the high degree of heterogeneity of both data and methods that characterizes the field should lead us to exercise caution in the formulation of generalizations. The trends we have identified should be interpreted as mere hypotheses that need to be tested against other learner corpus data, and our knowledge of the learner phrasicon would undoubtedly benefit from replication studies such as those conducted in Cobb (Reference Cobb2003). One lesson that can clearly be derived from our survey, however, is that any investigation of formulaic language needs to be very explicit as to the characteristics of the learner populations under investigation, the types of word combinations, and the methods used to extract them as well as the settings adopted (e.g., size of lexical bundles, minimum frequency threshold) and the criteria used to classify them.
The fact that learner corpus data are in electronic format has made it possible to use quantitative methods to investigate the learner phrasicon and distinguish two major aspects of phraseology, namely co-occurrence and recurrence. Although these two aspects are closely related and partly overlap, they have been shown to pose quite different problems for learners, and it is therefore advisable to treat them separately. The research results suggest that the development of L2 phraseological competence is characterized by an increase in the appearance of target-like low/medium frequency collocations and a decrease in the use of lexical bundles. They also show that accuracy is relatively slow to develop and that phraseological errors change over time, progressively morphing from totally non-native-sounding combinations, often based on the learners’ L1, to units that closely approximate native speaker use. In this connection, Thewissen (in press) spoke of a “possible accuracy-complexity trade-off effect”: At an advanced level, learners take more risks, try out more complex lexical phrases and as a result, produce errors, but these errors are of a different, more advanced nature than the basic phraseological errors typical of earlier stages.
The ease with which formulaic language can be identified automatically should not lead us to overlook the fact that the automatically extracted units are but raw materials, which still require a great deal of manual work (Granger, Reference Granger, Barfield and Gyllstad2009). Both recurring and co-occurring items need to be classified structurally and/or functionally and the vast literature emanating from the traditional approach to phraseology provides an invaluable battery of theoretical constructs for this purpose (cf. Cowie, Reference Cowie, Asher and Simpson1994).
In spite of the flurry of learner-corpus-based studies on the L2 phrasicon, some important areas are still largely underexplored, among which the following three are particularly worthy of attention. First, there is a need for more investigations of the L2 spoken phrasicon which is interesting in its own right but will also allow for useful comparisons with the L2 written phrasicon. Second, more truly longitudinal studies are needed to identify patterns of phraseological development in L2 learners and compare them with those based on quasi-longitudinal data. Third, there is a need for more studies along the lines of Nesselhauf (Reference Nesselhauf2009), Gilquin and Granger (Reference Gilquin, Granger, Hundt and Mukherjee2011), and Götz and Schilk (Reference Götz, Schilk, Hundt and Mukherjee2011), which compare the phrasicon in foreign and second language varieties of English.
The field of L2 phraseological studies is vast and its potential applications numerous. A better knowledge of the L2 phrasicon may lead to a profound reshaping of foreign language learning and teaching (see Meunier, this volume), improved scoring rubrics in language testing (Ohlrogge, Reference Ohlrogge, Corrigan, Moravcsik, Ouali and Wheatley2009), and more efficient algorithms in a range of natural language processing applications, in particular automatic error detection, automatic essay scoring, and automatic L1 identification (Jarvis & Paquot, Reference Jarvis, Paquot, Jarvis and Crossley2012; Leacock, Chodorow, Gamon, & Tetreault et al., Reference Leacock, Chodorow, Gamon and Tetreault2010). All these applications, however, require solid empirical data, and as we hope to have shown in this survey, learner corpora provide a particularly rich and versatile resource to use to this end.
