Frequency Effects and Second Language Lexical Acquisition Word Types, Word Tokens, and Word Production

Frequency effects and second language lexical acquisition Word types, word tokens, and word production

Scott Crossleyi, Tom Salsburyii, Ashley Titaki and Danielle McNamaraiii iGeorgia State University / iiWashington State University / iiiArizona State University

Frequency effects in an L1 and L2 longitudinal corpus were investigated using Zipfian distribution analyses and linear curve estimations. The results demonstrated that the NS lexical input exhibited Zipfian distributions, but that the L2 lexical output did not match the NS Zipfian patterns. Word frequency analyses indicated that NS interlocutors modify their lexicon such that frequency scores decrease as a function of time that L2 learners have studied English. In contrast, the word frequency scores for the L2 output increased as a function of time. Post-hoc analyses indicated that differences in frequency scores between NS input and L2 output were best explained by the repetition of infrequent words, but not frequent words by L2 learners in the early stages of language acquisition. The results question absolute frequency interpretations of lexical acquisition for L2 learners and provide evidence for usage-based approaches for language learning.

Keywords: lexicon, corpus, computational, frequency, usage-based

1. Introduction

The frequencies of linguistic items to which second language (L2) learners are exposed have important facilitative effects in L2 acquisition (Ellis 2002, MacWhinney 1997). Such effects are referred to as ‘frequency effects’, and theoretically relate to the common observation that the more frequently a linguistic item occurs in the language data to which an L2 learner is exposed (e.g. input), the more likely that item will be acquired by the L2 learner (Ellis 2002, Gries 2008). These frequency effects are based on structural regularities within language and, because frequency effects are distributional, language learning is argued to be an implicit

International Journal of Corpus Linguistics 19:3 (2014), 301–332. doi 10.1075/ijcl.19.3.01cro issn 1384–6655 / e-issn 1569–9811 © John Benjamins Publishing Company 302 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

phenomenon that is based on statistical learning premised on the strength of associations between representations (Ellis 2006a, MacWhinney 1997). However, frequency effects are generally not modeled simply on absolute occurrences of linguistic items within language exposure. Although absolute occurrence of an item is an important element of frequency and language acquisition, contemporary theories of language acquisition (i.e. associative learning through usage) consider frequency in terms of construction frequency, type frequency, and recency. Usage-based approaches also consider the importance of form and function in language acquisition. Thus, frequency effects in language acquisition are based on usage patterns (i.e. the more times a linguistic item is meaningfully experienced and used, the more likely it is incorporated into a language system) and not purely based on the raw frequency of a linguistic item’s occurrence (i.e. absolute frequency of occurrence), although absolute frequency is an important component of language acquisition (i.e. Ziphian distributions, Ellis 2012). Frequency effects are argued to be common in phonology, lexis, grammar, and syntax (Ellis 2002). The focus of this study is on frequency effects in the lexis and specifically on the absolute frequency of individual words in L2 input. We define absolute frequency as the raw occurrence of linguistic items in the absence of context. In the case of lexical items, absolute frequency refers to the raw token counts for individual words. Our hypothesis is that absolute frequency is not a strong determiner of L2 word acquisition and lexical output for beginning L2 learners. To address this hypothesis, we examine if the production of words by L2 learners is based simply on the raw frequency of exposure to words in first language (L1) input. To our knowledge, no studies have directly explored the relationship between spoken word frequency in native speaker (NS) input and spoken word frequency in L2 output. Our purpose is to reject the null hypothesis that the absolute frequency in L2 input can explain lexical production in L2 output and, in rejecting the null hypothesis, examine the possibility that other elements of a usage-based approach to language learning (e.g. forms and functions) can explain word production in L2 learners. To accomplish this, we examine data from a year- long longitudinal study of naturalistic speech between six L2 learners and 13 NS interlocutors. The longitudinal data in this study allows us to examine absolute frequency effects in L2 acquisition over time, affording insights into the relationship between NS input and L2 acquisition (Ellis & Collins 2009). Our primary purpose is to investigate absolute word frequency in NS input and its effects on L2 output. Our key research questions are: i. Do distributions of frequently shared words in NS input and L2 output follow Zipfian distributions such that the frequency of the words produced are inversely proportional to their rank on the frequency chart? Frequency effects and second language lexical acquisition 303 ii. Do NS modify the linguistic input they provide to L2 learners to make that input more comprehensible (i.e. use more frequent words when speaking to L2 learners at an early stage of learning)? iii. Do L2 learners produce more frequent words at an early stage as compared to a late stage of language learning as a result of exposure to more frequent words in NS input? iv. Do the words produced by L2 learners and their NS interlocutors exhibit similar frequency patterns throughout the stages of language learning? We predict, based solely on absolute frequency effects, that frequent words shared in both NS input and L2 output will not follow the same Zipfian distributions, demonstrating no absolute frequency effects for lexical acquisition. In addition, we predict that native speakers will use more frequent words at the beginning of the study when the L2 learners have lower language proficiency. Furthermore, as L2 language proficiency increases, we predict that native speakers will use more infrequent words. Concomitantly, we predict that L2 learners’ and native speakers’ word frequency use will, over time, begin to align with one another.

1.1 Input and output

Input and output are important attributes of second language acquisition. Input refers to the language to which the L2 learner is exposed and often serves as linguistic evidence from which language hypotheses are formed. NS input is often modified in the form of “foreigner talk” or “teacher talk.” Both of these forms are simplified at the lexical, phonological, and syntactic levels, assumedly to facilitate comprehension and processing of linguistic features (Gaies 1983, Hatch 1983). Strong arguments for the role of input in L2 acquisition state that L2 acquisition relies on the quality of available input (Ellis & Collins 2009) and that L2 acquisition is input driven (Wulff et al. 2009). However, by themselves, the linguistic features contained in input cannot sufficiently explain L2 acquisition. Acquisition also depends on numerous learner-based variables such as noticing, processing, storing, and production (i.e. output). Output is argued to allow L2 learners to move from lexical to syntactic processing and afford L2 learners the opportunity to experiment with new syntactic forms by testing hypotheses about language structure (Swain 1985). Output may also facilitate learning in that it directs learners’ attention to notice input, thus introducing the learner to new language data from which to develop linguistic knowledge. In this sense, output sensitizes learners to patterns and associations in future input (Swain 1995). Studies investigating output in L2 learning have demonstrated that it is an important component of noticing (Izumi et al. 1999) and 304 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

increasing the frequency (Toth 2006) and accuracy of use of grammatical struc- tures (Izumi & Bigelow 2000).

1.2 Zipfian distribution

Every level of language (i.e. phonology, morphology, lexicon, syntax, and discourse) is characterized by the disproportionate frequencies of certain individual items. Lexically, these frequencies generally follow a Zipfian (1935) distribution, with the most frequently occurring types accounting for a majority of tokens. Zipf’s Law states that the frequency of any item will be inversely proportional to its rank in a frequency table. The law suggests that while learners will be exposed to thousands of different items, the items to which they are regularly exposed to will be a much smaller set of highly frequent items, which decrease in input variability. Theoretically, these distributions aid in the process of language learning due to the increased probability of form-function mapping for the more frequent items (Ellis 2006a, 2006b; MacWhinney 1997) and the simple notion that the more times one is exposed to an item the faster and more accurately it will likely be processed (Ellis 2006b). Thus, with experience, a learner’s cognitive system becomes tuned to expect linguistic items according to their raw probability of occurrence in the input (Ellis 2006a). Practically speaking, as learners hear and read the same set of words consistently, these words are learned and produced more quickly (Mintz et al. 2002). There has been some evidence that L2 output follows distributional frequency effects in terms of the absolute frequency of words produced. For instance, Laufer & Nation (1995) and Bell (2003) report that lower level L2 writers produce more high frequency words than higher-level L2 learners. Similar results have been reported by Crossley et al. (2011) and by Crossley et al. (2012), who find that more advanced L2 speakers produced less frequent words than beginning level L2 learners. These studies indicate that beginning level L2 learners are more likely to produce words to which they are infrequently exposed in terms of absolute frequency. Conversely, Crossley et al. (2010), focusing on spoken language in a longitudinal study, demonstrate that beginning L2 learners produced words of lower frequency early and words of a higher frequency later. Similar findings have been reported in L1 studies. For instance, high frequency words found in caretaker input are in some cases associated with a later age of acquisition among L1 children. For example, closed-class words are the most frequent category of words found in caretaker input, but the slowest to be acquired by children (Goodman et al. 2008). Such findings would seemingly contradict absolute frequency effects in word learning, because learners would be expected to produce more frequent words early as a function of the distributional properties of the input. These findings Frequency effects and second language lexical acquisition 305 likely, however, indicate that it may not be the raw frequency of words that is important, but rather the usage of frequent words that lead to acquisition.

1.3 Usage-based approaches to acquisition

Frequency and usage-based approaches converge under the notion that it is not the raw input that learners are exposed to that necessarily leads to acquisition, but rather how the input is patterned in conversation (Ellis 2012). The meaningful patterns of input that occur frequently are often referred to as linguistic constructions, which are basically conventionalized form-meaning mappings. These mappings can encase units of language at the phonological, morphological, syntactic, and lexical level (Robinson & Ellis 2008, Goldberg 2003). Lexically, form-meaning mappings link words, idioms, and lexical phrases to their discourse or semantic functions (Ellis 2012). The form-meaning and/or form-function mappings that occur most frequently are argued to be learned first. However, it is not only the frequency of a construction that leads to learning, but also type frequencies, and recency effects. Outside of frequency, usage-based learning also relies on the salience and perception of the item, the prototypicality of item, and the redundancy of items (Boyd & Goldberg 2009; Ellis 2006a, 2012). In reference to type frequency, the potential for a greater number of items to substitute into a specific slot will afford quicker acquisition of a form (e.g. the number of nouns that can be slotted before the regular plural suffix -s in English). Recency effects refer to the tendency for language to prime similar language such that exposure to a linguistic item will increase the likelihood of that item being used. The salience or perceptibility of linguistic items refers to their perceived linguistic strength with low salience cues more difficult to acquire (Ellis 2006a). For instance, the low phonological salience of the past tense might explain why it is acquired later than the progressive aspect, which has greater salience (Boyd & Goldberg 2009). Prototypicality refers to the notion that some members of linguistic categories are more typical than other members. For instance, put is a prototypical verb for use in a verb-object-location construction (e.g. He put the glass on the table). Such prototypical verbs occur more frequently in specific constructions and promote the learning of the construction as well as the specific verb associated with the construction. Redundancy refers to features that are often not acquired because they are unnecessary for the interpretation of a message. For instance, verbs inflected for tense are often unnecessary for interpreting a message when temporal adverbs precede them (Ellis & Collins 2009). All of these usage-based elements can work individually to assist in language acquisition or, more likely, can work together (Wulff et al. 2009). However, in some cases, usage-based elements may supersede exposure to language based on 306 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

absolute frequency in terms of language acquisition. For instance, in the case of closed-class words as discussed above, children acquire them slowly even though they are frequent in caretaker input likely because they have low salience and may be redundant with other linguistic cues. Thus, absolute word frequency alone is not likely to explain language acquisition.

2. Methods

Our goal in this study is to investigate the effects of absolute word frequency in NS input to L2 learners in relation to L2 lexical production. Thus, we are interested in investigating whether absolute word frequency in L2 input has a relationship with absolute word frequency in L2 output (i.e. does an input-acquisition relationship exist? Ellis & Collins 2009). In this manner we can test the hypothesis that L2 lexical acquisition is primarily driven by the distributional properties of words to which L2 learners are exposed (i.e. can raw word frequency alone explain lexical acquisition?). To examine this hypothesis, we collected a longitudinal corpus of naturalistic spoken data between L2 learners and NS interlocutors. We computed the observed frequency values (i.e. frequency counts taken from the corpus itself) for the words in the input and the output at the beginning, middle, and end of the study. We then used these observed frequency values to examine if the 50 most common words shared between both NS interlocutors and L2 learners exhibited Zipfian distributions (cf. Ellis & Ferreira-Junior 2009). We then computed word frequency scores for the input and output of the collected speech samples using external frequency values calculated from the CELEX database (Baayen et al. 1995) and the British National Corpus (BNC 2007). We used the external frequency data to examine developmental trends in the L2 learners and potential modifications on the part of the NS interlocutors. We also compared correlational trends between L2 word frequency input and output. Thus, we looked at word frequency as a product of the longitudinal learner corpus (i.e. the observed frequencies) and as a product of larger, more representative corpus of NS language use (i.e. the CELEX database and the BNC).

2.1 Corpus

Our corpus came from six L2 learners enrolled in a year-long intensive English program at a large American university. Thirteen native speakers of English con- versed with the six L2 learners about every 2 weeks (not including program and university breaks) over the one-year period. The 30 to 45 minute sessions were Frequency effects and second language lexical acquisition 307 tape recorded and later transcribed. These transcriptions were then divided based on the participant to form the input and output corpora used in this study. The native speaker data formed our NS input corpus, while the L2 learner data formed our L2 output corpus. The L2 learners’ proficiency levels were tested upon arrival to the program using internal assessments. All participants in the study tested into the lowest proficiency level, Level 1, of a 6-level program. The L2 learners’ language growth was also assessed every other month through the institutional TOEFL. Over the course of the year, the L2 learners progressed through the six levels of the program. The learners were not members of the same class, but rather dispersed among sections within a level. Each learner in the study was given a pseudonym; this paper reports on data from Eun Hui (L1 Korean), Faisal (L1 Arabic), Jalil (L1 Arabic), Kamal (L1 Arabic), Marta (L1 Spanish), and Takako (L1 Japanese). The participants ranged in age from 18 to 29 years old. None of the participants had lived in the US for longer than 3 weeks prior to the start of the study. All participants reported that they had studied English at the secondary level in their countries of origin. Because of their prior English language instruction, their low TOEFL scores (M = 388.333, SD = 49.794), and because the L2 learners tested into the beginning level of the English program, we considered them false begin- ners. The 13 native speaker interlocutors were recruited from a graduate level course in second language acquisition taught at the same university. Interlocutors rotated as the year progressed and, as a result, each learner interacted with at least 3 or more interlocutors over the year of observation (M = 3.667, SD = 1.211). This rotation was necessary to control for familiarity effects between the L2 learner and the NS speaker. Not all interlocutors conducted an equal number of meetings with the L2 learners. Three interlocutors conducted only one meeting each. One interlocutor conducted 31 meetings. The average number of meetings per interlocutor was 7.615 (SD = 7.621). Participants (interlocutors and L2 learners) were given a variety of elicitation materials, but analysis of the transcripts demonstrate that the sessions were generally characterized by free conversation. However, because the purpose of the meetings was to elicit free conversation from the L2 learners, the NS input included a disproportionate number of question forms, especially at the beginning of the study when the L2 learners were less proficient in their English skills and unable to maintain conversations without assistance. The NS input to which the L2 learners were exposed in this study is taken to be illustrative of the type of naturalistic input (i.e. language found in natural and instructional settings) to which the L2 learners were typically exposed; however, we acknowledge the limitations of such an assumption, but note that controlling all NS input to which learners are exposed is impossible. 308 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Table 1. Descriptive statistics for longitudinal output data Learner Number of Average num- Average Average number Number of meetings/ ber of words number of of words per words in transcripts per utterance utterances transcript output Eun Hui 16 22.609 54.688 1,236.438 19,783 Faisal 12 27.126 72.667 1,971.167 23,654 Takako 16 23.588 51.875 1,223.625 19,578 Kamal 14 24.478 51.000 1,248.357 17,477 Jalil 15 39.073 64.933 2,537.133 38,057 Marta 16 30.502 63.813 1,946.438 31,143

The output corpus consisted of the spoken data collected from the six L2 learners only. A total of 99 transcripts were collected. However, we only focused on 89 of these 99 transcripts. Ten transcripts were excised from the corpus because the data contained within the transcripts largely focused on picture description tasks or other elicitation tasks that were not characteristic of free conversation and, thus, not representative of naturalistic language exposure. In preparation for the analysis of the learner corpus, transcriptions of each elicitation session were cleaned to eliminate interjections and non-English words. Each elicitation session was saved as a single text file containing the words of only the L2 learner in focus and not the NS interlocutor or other learners participating in the session. The text file was manually and electronically checked for spelling errors. Descriptive data for the output corpus is presented in Table 1. The input corpus consisted of the spoken data collected from the native speakers when interacting with the L2 learners. We selected the 89 native speaker transcripts that matched the transcripts found in the output corpus. Like the output corpus, we cleaned the input corpus to eliminate interjections. Each session was

Table 2. Descriptive statistics for longitudinal input data Learner Number of Average num- Average Average number Number of meetings/ ber of words number of of words per words in transcripts per utterance utterances transcript input Eun Hui 16 18.807 73.500 1,382.313 22,117 Faisal 12 13.820 87.917 1,215.000 14,580 Takako 16 15.325 61.063 935.813 14,973 Kamal 14 11.106 64.071 711.571 9,962 Jalil 15 9.993 73.733 736.800 11,052 Marta 16 16.013 65.375 1,046.875 16,750 Frequency effects and second language lexical acquisition 309 saved as a single text file containing only the words spoken by the interlocutor in focus. Descriptive data for the input corpus is presented in Table 2.

2.2 Word frequency indices

In addition to the observed frequency counts taken from the corpus, we also collected external frequency counts from two representative corpora of English: CELEX and the BNC.

2.2.1 CELEX We selected two CELEX word frequency indices from Coh-Metrix (Graesser et al. 2004). CELEX word frequency measurements consist of frequencies derived from the 1995 version of the COBUILD corpus, a 17.9 million word corpus. Coh- Metrix reports frequency values taken from the entire CELEX corpus (including both written and spoken texts) and for the spoken subset corpus contained in CELEX, which consists of 1.3 million spoken tokens. The Coh-Metrix indices cal- culate a mean logarithm (to the base of 10) for all the word tokens in the text except those not contained in the CELEX database. If a word in a text is not included in the CELEX corpus, it is not computed in the Coh-Metrix indices. The two indices we selected to use in this analysis were “word frequency all words” and “word frequency all spoken words”. We predict that both L2 output and NS input will demonstrate decreasing word frequency trends over time.

2.2.2 BNC We computed two frequency indices from Vocabprofiler (Cobb 2002): “Percentage of level 1 words” and “Percentage of words beyond the second band”. Vocabprofiler uses frequency counts from the BNC (version 3.2) to report the percentage of words in a text that occur within 1,000 word bands (i.e. the first band contains the 1,000 most frequent words in English and is considered the first level; the second band contains the next 1,000 most frequent words and could be considered the second band). The basic hypothesis behind such an approach is that low proficiency speakers will produce more words in low-level bands while high proficiency speakers will produce more words in high-level bands (Laufer & Nation 1995). Our first index, “Percentage of level 1 words”, was selected under the hypothesis that early NS input and L2 output would contain greater percentages of more frequent words. Our second index, “Percentage of words beyond the second band”, was selected because the production of words beyond the first 2,000 most frequent words of English characterizes intermediate and advanced L2 speakers (Laufer 2000). 310 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

3. Results

In the following sections, we present the results for the language proficiency development in our selected L2 participants (Section 3.1), the Zipfian distributional analyses (Section 3.2), the correlations between NS input and L2 output (Section 3.3), a decontextualized analysis (Section 3.4), the linear curve estimations (Sections 3.5 and 3.6), input to output correlations (Section 3.7), and post- hoc analyses (Section 3.8).

3.1 Language proficiency

To investigate the development of L2 proficiency, we conducted a repeated-mea- sures Analysis of Variance (ANOVA) using the learners’ TOEFL scores to analyze whether temporal intervals affected scores. Our prediction was that as time spent learning English increased, TOEFL scores would increase. Such an analysis would provide evidence of general linguistic growth. Only four learners completed all six of the tests given over the course of the year. One learner missed the second TOEFL and another learner missed the fifth. Thus, the ANOVA for the TOEFL scores included only four testing sessions, the first, third, fourth, and sixth. The results indicated that the L2 learners’ TOEFL scores increased as a function of time, F(3, 15) = 22.782, p < .001 (see Table 3). Within-subjects contrasts indicated that the TOEFL scores from the last examination on the 52nd week (the final TOEFL administration) were significantly different from the first examination on the sixth week (the first TOEFL administration), F(5, 15) = 33.983, p < .01. Additionally, there were significant differences in TOEFL scores between the 6th week and the 22nd week (the second TOEFL administration), F(3, 15) = 17.234, p < .01, and the 6th week and the 42nd week (the third TOEFL administration), F(3, 15) = 35.798, p < .01. There was also a significant linear trend,F (1, 25) = 40.076, p < .001. These findings provide evidence that significant L2 development likely occurred during the year of study.

Table 3. Means and standard deviations for TOEFL scores Testing session Week Mean Standard deviation 1 6 358.330 49.790 3 22 418.830 33.040 4 42 450.660 30.120 6 52 458.830 29.250 Frequency effects and second language lexical acquisition 311

3.2 Zipfian distribution analysis

To assess whether L2 output and NS input shared similar Zipfian distribution patterns, we compared the 50 most common words shared by the NS interlocutors and L2 learners during the 1st, 27th, and 50th weeks of the study. Such an approach allowed us to investigate if L2 learners use frequent words found in the NS input more often than infrequent words. To ensure the majority of the words were content words and not function words (i.e. words that were not nouns, verbs, ad- jectives, or adverbs), we coded the shared words. Of the 50 words for the 1st week, 38% were function words. In the 17th week, 28% of the words were function words and in the 50th week, 36% of the words were function words. The words for each week along with their observed normed frequency of occurrence in the corpus are located in Table 4.

Table 4. 50 most frequent words shared in NS input and L2 output Week 1 Week 27 Week 50 Words Input Output Words Input Output Words Input Output freq. freq. freq. freq. freq. freq. you 8.945 1.517 you 8.236 2.258 you 6.559 2.434 do 3.605 0.289 that 2.520 0.645 to 2.930 1.888 to 2.384 2.576 to 2.520 2.046 that 2.646 1.110 in 2.327 2.937 so 2.263 0.662 the 2.524 2.379 and 2.212 2.792 the 2.170 2.793 I 2.453 7.060 the 2.212 2.455 I 2.147 7.997 a 2.271 0.800 what 2.212 0.506 what 2.053 0.798 it 2.058 1.408 I 1.640 6.042 it 1.890 2.029 so 1.673 1.281 a 1.564 0.361 do 1.843 0.535 and 1.612 2.429 your 1.526 0.361 in 1.727 1.520 of 1.500 0.806 have 1.450 0.313 and 1.633 2.674 in 1.338 1.259 is 1.392 1.926 like 1.610 1.910 your 1.227 0.177 like 1.335 0.939 a 1.447 0.696 know 1.075 1.441 it 1.221 0.770 about 1.097 0.756 what 1.024 0.469 are 1.183 0.698 city 1.097 0.509 is 0.912 1.043 how 1.087 0.530 have 0.980 0.976 be 0.892 0.265 number 1.011 3.370 know 0.910 1.902 have 0.892 1.115 here 1.011 0.530 of 0.840 0.475 do 0.811 0.408 English 0.973 1.276 is 0.817 0.942 was 0.811 0.811 312 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Table 4. (continued) Week 1 Week 27 Week 50 Words Input Output Words Input Output Words Input Output freq. freq. freq. freq. freq. freq. about 0.801 0.265 they 0.770 1.163 like 0.689 2.915 when 0.801 0.361 but 0.677 1.477 are 0.679 0.359 okay 0.725 0.626 okay 0.677 0.501 he 0.679 1.225 me 0.706 0.698 think 0.630 0.739 for 0.669 0.800 for 0.668 0.794 or 0.607 0.730 they 0.649 0.701 or 0.629 0.481 here 0.537 0.526 but 0.639 1.822 where 0.629 0.313 are 0.513 0.280 really 0.639 0.193 he 0.610 0.987 if 0.513 0.365 just 0.598 0.546 good 0.553 0.554 there 0.513 0.323 think 0.598 0.475 go 0.534 0.963 with 0.513 0.509 when 0.588 0.458 from 0.515 0.554 go 0.490 0.993 if 0.578 0.734 yes 0.515 1.204 can 0.467 0.518 about 0.497 0.624 speak 0.496 0.506 yes 0.467 0.976 or 0.497 0.646 old 0.477 0.337 for 0.443 0.475 people 0.487 0.342 this 0.458 0.457 just 0.443 0.645 yes 0.466 0.999 but 0.420 1.228 one 0.443 0.526 with 0.456 0.408 know 0.420 0.578 good 0.420 0.484 can 0.446 0.679 study 0.420 0.722 this 0.420 0.789 mean 0.446 0.282 very 0.420 1.011 want 0.420 0.340 number 0.436 0.591 can 0.401 0.433 when 0.420 0.798 how 0.426 0.215 name 0.401 0.361 not 0.397 0.509 go 0.416 0.375 think 0.401 0.481 we 0.397 0.705 not 0.405 0.734 class 0.381 0.433 he 0.373 1.553 well 0.385 0.182 they 0.381 0.385 see 0.373 0.195 had 0.375 0.171 with 0.381 0.361 something 0.350 0.263 this 0.365 1.021 one 0.362 0.963 time 0.350 0.475 no 0.355 0.845 there 0.362 0.289 maybe 0.327 0.654 there 0.355 0.392 not 0.305 0.650 # 0.303 0.772 she 0.345 0.878 brother 0.286 0.409 take 0.303 0.187 me 0.324 1.170 friends 0.286 0.409 at 0.280 0.170 at 0.314 0.171 then 0.286 0.289 came 0.280 0.272 very 0.314 0.580 Frequency effects and second language lexical acquisition 313

3.2.1 NS interlocutors The frequency distributions of the 50 most common words shared between NS interlocutors and L2 learners for the three weeks of interest are shown in Figure 1. The word frequency distributions for each week appear to be Zipfian in that the frequency of the words are inversely proportional to their rank on the frequency chart. To test the strength of these inverse relations, we plotted the frequency distributions as log word frequencies against log word ranks (Figure 2). The strength of the R2 value that resulted from Pearson correlations between log frequency and log rank demonstrates that the distribution is a straight-line function and thus Zipfian (R2 between .891 and .934).

10 9

9 8

8 7 7 6 6 5 5 4 4 3 3 Word frequency NS input Word frequency NS input 2 2

1 1

0 0 1 6 11 16 21 26 31 36 41 46 1 6 11 16 21 26 31 36 41 46 Commonly shared words in input and output Commonly shared words in input and output 7

2 Word frequency NS input 1

0 1 6 11 16 21 26 31 36 41 46 Commonly shared words in input and output Figure 1. Type-token frequency distribution for 50 most commonly shared words in interlocutors’ input compared to learners’ output, organized by rank order (weeks 1, 27, and 50) 314 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

r2 = .904 r2 = .934 2 2

1.8 1.8 1.6 1.6 1.4 1.4 1.2 1.2

1 1

0.8 0.8

0.6 0.6 Log word frequency NS input

Log rank NS input word frequency 0.4 0.4

0.2 0.2 0 1

1.69019608 0.698970004 1.146128036 1.361727836 1.505149978 1.612783857 1.278753601 1.447158031 1.568201724 1.662757832 Log rank NS input word frequency Log rank NS input word frequency r2 = .891 2

1.8

1.6

1.4

1.2

0.8

0.6 Log word frequency NS input 0.4

0.2 0

1.51851394 0.9542425091.2304489211.397940009 1.612783857 Log rank NS input word frequency Figure 2. Zipfian type-token frequency distribution for 50 most commonly shared words in interlocutors’ input as compared learners’ output, organized by log rank (weeks 1, 27, and 50)

3.2.2 L2 learners The frequency distributions of the 50 most common words shared between NS interlocutors and L2 learners for the three weeks of interest are shown in Figure 3. The distributions are organized based on the rank order of the words as found in the NS interlocutors’ data in Figure 1 in order to test relationships between the NS input and the L2 output. Accordingly, the first words in the charts in Figure 3 correspond to the first words in the charts in Figure 1 and so forth for each of the 50 words. When using the NS input rank order, the word frequency distributions for the L2 output do not correspond to the word frequency distributions in the Frequency effects and second language lexical acquisition 315

7 9

6 8

5 7 6 4 5

3 4

3 2 Word frequency NNS output

Word frequency NNS output 2 1 1

0 0 1 6 11 16 21 26 31 36 41 46 1 6 11 16 21 26 31 36 41 46 Commonly shared words in input and output Commonly shared words in input and output 8

2 Word frequency NNS output

0 1 6 11 16 21 26 31 36 41 46 Commonly shared words in input and output Figure 3. Type-token frequency distribution for 50 most commonly shared words in learners’ output as compared to interlocutors’ input, organized by rank order in interlocutors’ input (weeks 1, 27, and 50)

NS input and do not appear to demonstrate Zipfian distributions. To test if there was an inverse relationship between the word frequency and rank, we plotted the frequency distributions as log word frequency against log word rank (Figure 4). The resulting R2 value from Pearson correlations between frequency and rank demonstrates that the distribution is not a straight-line function and thus does not support the notion that the distribution is Zipfian (R2 between .212 and .369).

3.3 Correlations between NS input and L2 output

Figures 1 and 3 indicate that the rank order of the 50 most commonly shared words between the NS input and L2 output are dissimilar. Weak to moderate 316 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6 Log word frequency NNS output Log word frequency NNS output 0.4 0.4

0.2 0.2 0 1

1.69019608 0.698970004 1.146128036 1.361727836 1.505149978 1.612783857 1.278753601 1.447158031 1.568201724 1.662757832 Log rank NNS output word frequency Log rank NNS output word frequency 2

1.8

1.6

1.4

1.2

0.8

0.6

Log word frequency NNS output 0.4

0.2

1.51851394 0.954242509 1.230448921 1.397940009 1.612783857 Log rank NNS output word frequency Figure 4. Zipfian type-token frequency distribution for 50 most commonly shared words in learners’ output as compared to interlocutors’ input, organized by log rank of interlocutors’ input (weeks 1, 27, and 50)

correlations between the frequency of the shared words in the input and output confirm this for each of the three temporal intervals of interest: Week 1, r = 0.285, p < .050, N = 50; Week 27, r = 0.403, p < .050, N = 50; Week 50, r = 0.495, p < .001, N = 50.

3.4 Decontextualized analysis

The initial word analyses demonstrated differences in the production of personal pronouns and question forms between the L2 output and the NS input that likely resulted from the nature of the corpus (i.e. elicited conversation). To control for Frequency effects and second language lexical acquisition 317

Table 5. Pearson correlations between log frequency and rank order for the 50 most frequent shared words (without first person personal pronouns and question forms) Week Data type r R2 1 Input −0.962 0.925 Ouput −0.550 0.303 27 Input −0.960 0.922 Ouput −0.546 0.298 50 Input −0.956 0.915 Ouput −0.558 0.311 these lexical forms, a second Zipfian analysis was conducted on the 50 most commonly shared words in the transcripts for the 1st, 27th, and 50th week excluding first person personal pronouns (e.g. you and I) and question forms (e.g. what, do, when). Here we only report the straight-line functions for each week that resulted from plotting the observed frequency distributions as log word frequency against log word rank (see Table 5). The resulting R2 values from Pearson correlations demonstrate that the distributions for the input pattern are straight-line functions while those for the output are not. Thus, the distribution of the NS input is Zipfian, but the L2 output for the same words is not Zipfian. We also report the correlations between the frequency of the shared words in the input and output. These reported moderate correlations and suggest that the frequency of occurrence for the 50 most commonly shared words (absent first person pronouns and question forms) between the NS input and L2 output are not strongly similar: Week 1, r = 0.685, p < .001, N = 50; Week 27, r = 0.632, p < .001, N = 50; Week 50, r = 0.563, p < .001, N = 50.

3.5 Linear curve estimations NS input

To investigate if the frequency values for words in NS input declined as a function of L2 learners’ developing linguistic proficiency, we conducted linear curve estimates to model changes in the frequency of the words present in the NS input for the two CELEX frequency indices reported by Coh-Metrix and the two indices we computed from Vocabprofiler. A linear curve estimation affords the examination of word frequency changes as a function of time spent studying English by calculating a goodness of fit model between the frequency variable and the time variable. If native speakers are modifying their input to make it more comprehensible, the frequency of words produced by NS interlocutors (i.e. input frequency) should demonstrate significant, negative linear trends with time spent studying English as a result of fewer linguistic modifications (i.e. word simplification) made 318 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Table 6. Input linear curve estimates for selected frequency indices (N = 89) Index r r2 F p Word frequency all words CELEX −0.307 0.094 9.052 < .010 Word frequency all spoken words CELEX −0.230 0.053 4.851 < .050 Percentage of level 1 words 0.088 0.007 0.680 > .050 Percentage of words beyond second band 0.136 0.019 1.635 > .050

by the NS interlocutors throughout the year. Accordingly, NS would be expected to produce more high frequency words at the beginning of the year and more low frequency words at the end. The input linear trend estimates for the NS input showed that as time studying English increased, significant, negative linear trends were demonstrated for each CELEX index. However, these trends were relatively weak. No significant trends were reported for the Vocabprofiler indices (see Table 6 for more details). The CELEX indices demonstrated that the NS interlocutors tended to produce more low frequency words over time. The CELEX findings help support the notion that NS interlocutors likely modify the frequency of their words in relation to the amount of time the L2 learner has spent studying English.

3.6 Linear curve estimations L2 output

We also conducted linear curve estimates to model changes in word frequency production in L2 output using all four word frequency indices. A distributional frequency account of word learning would predict that the frequency of words produced by L2 learners would match the NS input and thus correlate negatively with the time variable. Such a finding would demonstrate that L2 learners produce more low frequency words as a function of time spent studying English (i.e. producing more high frequency words at the beginning of the year and more low frequency words at the end). However, contrary to expectations, our analysis showed that L2 learners produced more high frequency words as a function of time (per the CELEX indices). Additionally, L2 learners produced a greater percentage of level 1 words and a lower percentage of words beyond the second level as a function of time (see Table 7 for more details). The “percentage of level 1 words” explained the greatest amount of the variance for time spent studying English. These findings demonstrate that L2 learners initially produce more low frequency (rare) words and, as time spent studying English increases, L2 learners begin to use more frequent words. Frequency effects and second language lexical acquisition 319

Table 7. Output linear curve estimates for selected frequency indices (N = 89) Index r r2 F p Word frequency all words CELEX 0.395 0.156 16.075 < .001 Word frequency all spoken words CELEX 0.460 0.211 23.331 < .001 Percentage of level 1 words 0.467 0.218 24.254 < .001 Percentage of words beyond second band −0.313 0.098 9.468 < .010

3.7 Input to output correlations

We conducted Pearson correlations between the word frequency scores in the input and output for all four selected frequency indices to examine a potential relationship between the NS input and the L2 output (see Table 8). Contrary to predictions, the analyses yielded only one significant correlation between word frequency scores in the NS input and the L2 output for the selected frequency indices (“percentage of words beyond second band”). However, the r value for this correlation was weak showing that L2 output and NS input have marginally similar patterns in their use of infrequent, but not frequent words. In general, the correlational findings from this analysis do not support the notion that there are strong relationships between the frequency scores of the words present in the NS interlocutor data (the NS input) and the frequency scores of the words produced by the L2 learners (the L2 output). Illustrations of the word frequency input and output correlational trends for each word frequency index are presented in Figure 5.

Table 8. Correlations between input and output frequency values Index r p Word frequency all words CELEX −0.099 0.355 Word frequency all spoken words CELEX −0.061 0.571 Percentage of level 1 words 0.196 0.066 Percentage of words beyond second band 0.245 0.020 320 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

3.35 3.5

3.3 3.45

3.4 3.25 Output 3.35 Input Output Input 3.2 3.3 Frequency value Frequency value 3.25 3.15 3.2

3.1 3.15 1 2 8 10 13 20 22 27 30 32 38 40 42 44 46 48 1 2 8 10 13 20 22 27 30 32 38 40 42 44 46 48 Week of study Week of study

5 96

4.5 95 94 4 93 3.5 Input 92 Input Output Output 3 91 90 Frequency value 2.5 Frequency value 89 2 88 1.5 87 1 2 8 10 13 20 22 27 30 32 38 40 42 44 46 48 1 2 8 10 13 20 22 27 30 32 38 40 42 44 46 48 Week of study Week of study Figure 5. NS input and L2 output (CELEX and Vocabprofiles word frequency indices) as a function of time

3.8 Post-hoc analyses

We conducted a series of post-hoc analyses to illuminate the findings in our initial analyses. These post-hoc analyses were precipitated by the lack of Zipfian trends in the L2 output when compared to the NS input and the unexpected finding that the L2 learners’ produced more, not less, frequent words as a function of time. Such patterns did not strongly correlate with the word frequency scores in the NS input to which the learners were exposed and, thus, the findings from this data set do not lend support to a theory of word learning that is based solely on absolute frequency effects. We hypothesize that the frequency trends reported by our indices may be the result of the indices measuring the frequency of all the words in the text (the tokens), but not the individual words (the types). Thus, in our post-hoc analyses, we investigated whether the frequency of the individual word types produced by the L2 learners decreases over time and whether this pattern matches that found in the NS interlocutor data. Additionally, we examined if the frequency trends noted in our initial analyses are the result of infrequent word repetition on the part of L2 Frequency effects and second language lexical acquisition 321 learners at an early stage of learning. Such word repetition would not be measured by the token-based indices used in the initial analysis and may explain why word frequency values for L2 samples increase over time.

3.8.1 Type frequency analysis We conducted correlation analyses between time spent studying English and word type frequency values for both the NS input and L2 output. We also conducted t- tests between the frequency values of the types found in the NS input and in the L2 output. For both of these analyses, we examined frequency values for all the word types in the samples, those word types produced by three or more participants (i.e. frequent word types), and those word types produced by two or fewer participants (i.e. infrequent word types). To gather the data for this analysis, we combined the transcripts from the weeks wherein at least five L2 speakers participated. Such an approach allowed us to examine 11 weeks of data (weeks 1, 8, 10, 13, 20, 27, 30, 32, 40, 44, and 50) for both the NS input and the L2 output. To ensure that word types produced by a plurality of the participants (i.e. three or more participants) were, in fact, more frequent, we compared the frequency values of these word types to those word types produced by 2 or fewer of the participants using the CELEX index “word frequency all words”. This analysis demonstrated that words shared by three or more participants (M = 2.891, SD = 0.088) were significantly more frequentt (10) = 19.900, p < .001, than words shared by two or fewer participants (M = 1.990, SD = 0.068). We next conducted correlational analyses between the frequency of the word types in the L2 output and NS input data. The correlations between time spent studying English and type frequency values for both NS input and L2 output demonstrated negative trends such that word types with higher frequency values were produced at the beginning of the study and word types with lower frequency values were produced toward the end of the study (supporting our original predictions). Significant correlations for the input were reported for CELEX indices in all word types and infrequent word types (those words used by two or fewer participants). No significant correlations were reported for the frequent word types condition (those words used by three or more participants). For the output, significant correlations were reported for all the CELEX indices in all the conditions (see Table 9 for details). This analysis yielded significant trends toward the production of more infrequent word types in both the NS input and L2 output as a function of time studying English. We also tested the relationships between the frequency values of word types in the NS input to the frequency values of word types in the L2 output. For both CELEX indices, a strong relationship between word type frequency values was reported, (CELEX all words: r = 0.844, p < .001, N = 22; CELEX all spoken words: 322 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Table 9. Correlations between time studying English and word type frequency Condition CELEX all words CELEX all spoken words r p r p Input All −0.860 <.001 −0.862 <.001 2 or fewer −0.768 <.010 −0.799 <.010 3 or more −0.536 >.050 −0.439 >.050 Output All −0.823 <.010 −0.843 <.001 2 or fewer −0.786 <.010 −0.801 <.010 3 or more −0.740 <.010 −0.758 <.010

r = 0.792, p < .001, N = 22) demonstrating similar frequency trends in the word types in the NS input and the L2 output. Thet -test analyses yielded significant differences in the word frequency scores between the NS input and L2 output data for all word types, frequent word types (those produced by three or more participants), and infrequent word types (those produced by two or fewer participants). Overall, word types in the NS input had significantly higher frequency scores than word types in the L2 output in all conditions. This analysis indicates that L2 speakers produce word types with significantly lower frequency values than their NS interlocutors. Descriptive and t-test statistics for these analyses are provided in Table 10.

Table 10. Descriptive and t-test statistics for type frequency (CELEX word frequency values): Input and output Index Mean Standard t p deviation All words All words Input 2.330 0.095 2.996 < .010 Output 2.227 0.063 Word produced by three or Input 3.022 0.105 3.280 < .010 more participants Output 2.849 0.141 Word produced by two or fewer Input 2.076 0.088 2.628 < .050 participants Output 1.984 0.074 All spoken words All words Input 2.334 0.116 3.478 < .010 Output 2.192 0.071 Word produced by three or Input 3.111 0.117 3.594 < .010 more participants Output 2.899 0.157 Word produced by two or fewer Input 2.038 0.109 3.014 < .010 participants Output 1.913 0.083 Frequency effects and second language lexical acquisition 323

3.8.2 Frequent and infrequent word repetition analysis We next examined the repetition of frequent word types (those produced by three or more participants) and infrequent word types (those produced by two or fewer participants) in both the NS input and L2 output. Our objective was to examine if infrequent words were repeated more often in the early L2 output as compared to frequent words. If such a trend were found, it would provide an explanation for why L2 learners’ transcripts contain lower token word frequency scores at the beginning of the study as compared to the end. For this analysis, we created word lists for the frequent and infrequent words present in the NS input and L2 output from weeks 1, 8, 10, 13, 20, 27, 30, 32, 40, 44, and 50. These word lists included the number of words in the text, the word types, and the number of occurrences of those word types. To analyze word repetition, we computed the average word repetition score by calculating the number of occurrences for each word token in the transcript divided by the number of total words in the transcript. We then conducted correlations between the average word repetition scores and the cor- responding weeks in which they were produced. The correlation between the word repetitions and the weeks of study for the infrequent words in the L2 output was negative and significant (r = −.911, p < .001, N = 11). However, the correlation for the frequent words was positive and not significant (r = .208, p = .540, N = 11). This finding indicates that L2 learners repeat infrequent words at the beginning of the longitudinal study and the repetition of infrequent words decreases as time spent studying English increases. No such trend is noted for frequent words. The correlation between word repetition and the weeks of study for the infrequent words in the NS input was negative and significant (r = −.866, p < .001, N = 11) as was the correlation for the frequent words (r = −.711, p < .050, N = 11). This finding indicates that NS input contains repetitions of frequent and infrequent words at the beginning of the longitudinal study, but the repetition of words, regardless of frequency, decreases as time spent studying English increases. These findings support the notion that L2 learners repeat infrequent words more often at the beginning of the study with trends toward less repetition of infrequent words as time spent studying English increases. The same is not true for frequent words, which appear to be repeated at the same rate regardless of time spent studying English. The repetition of infrequent words by L2 learners early in the study, but not frequent words, likely explains the word frequency scores reported in the initial analysis (i.e. the production of more infrequent words at the beginning of the study and more frequent words at the end of the study). The input L2 learners receive from their NS interlocutors contains repetitions of both frequent and infrequent words early, but these word repetitions decrease as the L2 learners’ time studying English increases, thus helping to explain why the NS 324 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

interlocutors produce more frequent words early in the study and more infrequent words as the study continues.

4. Discussion

These analyses present a rather complex relationship between absolute word frequency in NS input and its connections to absolute word frequency in L2 output. The analyses demonstrate that the lexical input L2 learners are exposed to is Zipfian, but that the matching L2 output is not. Word frequency analyses based on token counts demonstrate that L2 learners produce words with lower frequency values, fewer words in the first frequency band, and more words beyond the most frequent 2,000 words at the beginning of the longitudinal study. Over time, the data show significant linear trends toward the production of words with higher frequency values, more words in the first frequency band, and fewer words beyond the most frequent 2,000 words by the end of the observation period. In general, this pattern is the reverse of what we find in the NS input; although, over time, the frequency of words in the input and in the output appears to begin to converge (see Figure 5). If we examine only word types, we see trends in both L2 learners and their NS interlocutors toward producing words types with lower frequency values as a function of time studying English. These trends yield strong correlations between changes in the frequency values of the word types in the NS input and L2 output over time. Our explanation of these patterns rests on evidence that L2 learners, unlike their NS interlocutors, repeat more infrequent words early in the study with a trend towards less repetition of infrequent words by the end of the study. However, L2 learners’ repetition of frequent words does not correlate with time spent studying English, resulting in lower token frequency scores as compared to the NS interlocutors, who repeat both infrequent and frequent words at the beginning of the study with a significant trend toward lower word repetition by the end of the study. The analyses of word type frequency raise further complications because they indicate that L2 learners produce words that have significantly lower frequency values than their NS interlocutors overall, even when the word types are controlled for frequency (i.e. word types produced by three or more participants and those produced by two or fewer participants). Thus, we are left with the notion that whereas NS interlocutors move toward the production of words with lower frequency values, L2 learners move toward the production of words with higher frequency values as they become more proficient in English. While there is a weak correlation between the production of more infrequent words (described as words beyond the most frequent 2,000 words) Frequency effects and second language lexical acquisition 325 over time between L2 learners and their interlocutors, L2 learners still produce word types with significantly lower frequency values in all conditions (all words, frequent words and infrequent words) than their NS interlocutors. Additionally, there is not strong evidence that L2 word production matches the Zipfian distribution trends found in the NS input. This is not to say that L2 output is not Zipfian (it is), but that the NS input and the L2 output for the same words do not demonstrate Zipfian distribution patterns. Thus, for the L2 learners in this study, the frequency values of the words in the NS input are not correlated to the frequency values of the words in the L2 output, indicating that the L2 learners in this study did not produce frequent words at a similar rate as their interlocutors. To put these findings in context and to provide tangible examples, we look at the lexical output of one learner (Eun Hui) over the course of the study. We focus on the results produced by Vocabprofiler because these results better lend themselves to qualitative analysis unlike those reported by the CELEX indices. Table 11 shows the percentages of words at band 1, band 2, and band 3 and above produced by Eun Hui in the 2nd, 20th, 38th, and 50th week along with descriptive statistics for the speech samples in which the indices were computed. The descriptive statistics demonstrate an increase in the number of tokens per sample as would be expected with increasing linguistic proficiency. However, the percentage of words found in the first band demonstrates that, over time, Eun Hui begins to produce a higher percentage of words that are among the most 1,000 frequent words in English. Eun Hui also begins to produce a lower percentage of words beyond the top 2,000 most frequent words in English. Both of these findings run counter to expectations of developing linguistic competence and distributional approaches to language learning. In Table 12, we present all the word types and their frequencies produced by Eun Hui beyond the 2,000 most frequent words (we do not present the words from bands 1 and 2 in the interest of space) from the weeks analyzed above. For ease of processing, we combine the 18 bands (bands 3–20) into four, larger groups. The differences in the number of words produced in each transcript makes

Table 11. Descriptive statistics and percentage of words in each band and those bands above 2 for Eun Hui’s output Week Tokens Types TTR Band 1 Band 2 Band 3 and above week 2 267 129 0.48 79.780 4.870 9.730 week 20 1,198 334 0.28 83.810 5.930 7.100 week 38 1,916 353 0.18 89.970 4.810 4.190 week 50 2,063 347 0.17 92.290 4.310 2.560 326 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Table 12. Types and token counts from selected speech samples taken from Eun Hui’s output Week Band 3–7 stems Band 8–12 Band 13–16 Band 17–20 (tokens) stems stems stems (tokens) (tokens) (tokens) 2 accountant (2) angry (2) grammar (1) classmate (2) MBA (1) stepsister hobby (2) pet (1) silver (1) tennis (1) gem (1) (1) Korea (11) 20 adapt (2) Asian (1) Bible (4) calendar (3) undergraduate Canada (2) climate (1) cloth (1) dining (1) lunar (3) (1) drum (4) instrument (1) mask (2) canyon museum (1) soil (1) campus (2) disco (1) dormitory (2) leather (3) Korea (12) peasant (2) ping classmate (2) (9) pronunciation (1) teller (1) Buddhist gong (6) (2) pong (9) 38 apartment (1) core (2) Denmark (1) undergraduate MBA (2) divorce (2) eldest (2) essay (4) host (3) (1) semester ginseng (1) liquid (2) lonely (1) movie (2) personal- (4) airplane ity (2) tennis (1) deposit (1) ocean (2) (1) backache professor (1) studio (2) vocabulary (1) (4) elective (2) gym (2) journal (1) Korea (13) waist (4) pong (2) ping (2) pronunciation (1) aerobic (1) plantation (1) 50 essay (1) exhaust (2) grammar (4) classmate (1) motel (1) orientation (2) reputation (2) teenager (8) envy (5) president (3) vocabulary (1) Korea (5) orphan (3) prostitute (8) stag (1) vacation (4)

comparisons between the weeks difficult, but there is a clear pattern in the number of words produced at each band group from week 1 to week 50 (as seen in the “Band 3 and above” column 7 of Table 11) with fewer word types being produced in upper frequency bands as a function of time. If considering only word types, the normed frequency for number of word types above the 2,000 most frequent words (normed for 1,000 words) displays similar trends with 45 word types above the 2,000 most frequent words in the 2nd week; 24 words types above the 2,000 most frequent words in the 20th week; 17 word types above the 2,000 most frequent words in the 38th week and 8 word types above the 2,000 most frequent words in the 50th week. While findings such as these may provide counter evidence to a strict frequency account of word learning, they do not, in themselves, contradict frequency Frequency effects and second language lexical acquisition 327 effects in language learning. This study simply demonstrates that there is not strong evidence that absolute word frequency alone helps to explain individual word production. There are many potential reasons for this. The first potential reason is that many frequent words may associate with difficult grammatical and syntactic form-function relations and thus will not be produced early by L2 learners. In addition, these form-function relations may be less salient phonologically (i.e. less likely to be learned under a usage-based approach) and thus more difficult to acquire. Lastly, there is the potential for cross-linguistic influence in L2 word production and acquisition that needs consideration. We will discuss all three of these in detail. One problem with very frequent words is that they may represent complex form-function relations that are difficult to acquire (and may be less salient as a cue for learning). For instance, articles are difficult to learn because although their form is simple, their function is complicated (Pica 1983). Phonologically, articles are also generally of low salience (Shockey 2003). Perfect auxiliaries are also lower in phonological salience (Shockey 2003), simpler in form, but complicated in function, and acquired later by L2 learners (Bardovi-Harlig 2000). In contrast, question formations have a simple function, but a difficult form and may be acquired later (Gass & Mackey 2002, McDonough & Kim 2009). In cases such as these, a word’s raw frequency may not facilitate acquisition because the associations between the word and its grammatical or syntactic form and function may be complex or phonologically less salient and thus less likely to be noticed and acquired (Ellis 2006a). A quick, qualitative analysis of the NS input and output data from the first weeks of this study supports such assertions. Looking at only the top 50 most frequent words in both NS input and output and comparing which words are shared between the conditions demonstrates that NS interlocutors’ 50 most frequent words included questions words (did, when, where), indefinite articles (a), and perfect auxiliaries and verbs (have, been) that were not included in the top 50 words produced by the L2 learners. These words, along with other categories present in the most frequent words of the NS interlocutors, but not the L2 learners such as possessive pronouns (his, her, your) and complementizers/demonstratives (that), indicate that many words frequently produced by NS interlocutors may be infrequent in the L2 data because of their association to complex grammatical and syntactic form function relations as well as to their association with lower phonological salience. Another potential reason for the findings in this study may be that the initial state of L2 learning is not tabula rasa, as it is with NS learners, but rather tabula repleta (Ellis 2006a, 2006b). Thus, we have to consider the effects of the learner’s first language, which will influence word production and acquisition in a second language (Adjemian 1983, Singleton 1999). For example, L2 learners may produce 328 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

adverbs relatively early because they often already have experience with adverbs in their first language (Bardovi-Harlig 1992). Thus, it is highly possible that the production of infrequent words at an initial stage of L2 learning is the result of advanced lexical knowledge in the NS. L2 learners may also carry over word frequency patterns from their first language to their second. These word frequency patterns may be dissimilar and may result in L2 learners producing words that, while frequent in their first language, may not be frequent in their second language. Such tendencies may be greater in learners whose first language shares a greater number of cognates with their second language.

5. Conclusions

There are a few areas of L2 learning where absolute frequency interpretations seem debatable. These include implicational scales of morphosyntax (i.e. third-person singular -s; Goldschneider & DeKeyser 2001) and question formation (Gass & Mackey 2002, McDonough & Kim 2009). We would add to this list the acquisition of lexical items, especially for beginning level L2 learners. As noted before, we find little evidence that the words produced by NS interlocutors and L2 learners share Zipfian distribution patterns. We also find little correlation between the frequency of words in NS input and L2 output and instead find that the frequency of words in the NS input is significantly more frequent than the L2 output. These findings are likely the result of L2 learners producing and repeating infrequent words early in their language studies. This is not to say that learners are not intuitive statisticians (Ellis 2006a, 2006b), but rather that the statistics learners employ are likely far more complex than simple frequency counts. As Gries (2008, 2010) notes, frequency of occurrence is often a flawed predictor of processing, especially in the absence of context. A primary reason for this is that the frequency is highly intercorrelated with other linguistic aspects such as saliency, recency, and concreteness. Thus, it is likely that L2 learners use a variety of linguistic properties to facilitate word learning such as context, perceptual salience, word distinctiveness, and word prototypicality. In addition, it is likely that the conditional frequency of the word within a specific construction (i.e. verb-argument constructions or tense-aspect patterns) is a better predictor of lexical acquisition than absolute word frequency (Ellis & Ferreira-Junior 2009). Such properties should be explored in future studies, especially those lexical properties related to prototypicality that are easy to compute such as word concreteness, imagability, and familiarity. Future studies should also link word production to lexical accuracy and analyze larger longitudinal corpora, if available, to ensure the findings of this study are representative of English L2 Frequency effects and second language lexical acquisition 329 learners as a whole and not just the participants in this study. Lastly, the data reported here could be interpreted as indicating that learners begin to approximate the Zipfian distributions found in the NS data toward the end of the study. Thus, while beginning level learners may not have the language experience necessary to produce words at similar frequencies as found in native speaker speech, with time and experience, the frequency of words in L2 output may begin to more closely match that found in L1 input. Therefore, future studies should also consider the relationship between input and output frequency for more advanced learners.

References

Adjemian, C. 1983. “The transferability of lexical properties”. In S. Gass & L. Selinker (Eds.), Language Transfer in Language Learning. Rowley, MA: Newberry House, 250–268. Baayen, R. H., Piepenbrock, R. & Gulikers, L. 1995. The CELEX Lexical Database (Release 2) [CD-ROM]. Philadelphia, PA: University of Pennsylvania, Linguistic Data Consortium. Bardovi-Harlig, K. 1992. “The relationship of form and meaning: A cross sectional study of tense and aspect in the interlanguage of learners of English as a second language”. Applied Psycholinguistics, 13 (3), 253–278. DOI: 10.1017/S0142716400005634 Bardovi-Harlig, K. 2000. Tense and Aspect in Second Language Acquisition: Form, Meaning, and Use. Oxford: Blackwell. Bell, H. 2003. Using Frequency Lists to Assess L2 texts. Unpublished PhD dissertation, University of Wales Swansea, Swansea. Boyd, J. K. & Goldberg, A. E. 2009. “Input effects within a constructionist framework”. Modern Language Journal, 93 (3), 418–429. DOI: 10.1111/j.1540-4781.2009.00899.x The British National Corpus. 2007. Distributed by Oxford University Computing Services on be- half of the BNC Consortium. Available at: http://www.natcorp.ox.ac.uk/ (accessed October 2013). Cobb, T. 2002. Web VocabProfiler. Available at: http://www.er.uqam.ca/nobel/r21270/cgi-bin/ webfreqs/web_vp.cgi (accessed October 2013). Crossley, S. A., Salsbury, T. & McNamara, D. S. 2010. “The development of polysemy and frequency use in English second language speakers”. Language Learning, 60 (3), 573–605. DOI: 10.1111/j.1467-9922.2010.00568.x Crossley, S. A., Salsbury, T., McNamara, D. S. & Jarvis, S. 2011. “What is lexical proficiency? Some answers from computational models of speech data”. TESOL Quarterly, 45 (1), 182– 193. DOI: 10.5054/tq.2010.244019 Crossley, S. A., Salsbury, T. & McNamara, D. S. 2012. “Predicting the proficiency level of language learners using lexical indices”. Language Testing, 29 (2), 240–260. DOI: 10.1177/0265532211419331 Ellis, N. 2002. “Frequency effects in language processing: A review with implications for theories of implicit and explicit language acquisition”. Studies in Second Language Acquisition, 24 (2) 143–188. 330 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Ellis, N. 2006a. “Cognitive perspectives on SLA: The associative-cognitive CREED”. AILA Review, 19, 100–121. DOI: 10.1075/aila.19.08ell Ellis, N. 2006b. “Language acquisition as rational contingency learning”. Applied Linguistics, 27 (1), 1–24. DOI: 10.1093/applin/ami038 Ellis, N. 2012. “What can we count in language, and what counts in language acquisition, cognition, and use?” In S.T. Gries & D. S. Divjak (Eds.), Frequency Effects in Language Learning and Processing (Vol. 1). Berlin: Mouton de Gruyter, 7–34. Ellis, N. & Collins, L. 2009. “Input and second language acquisition: The roles of frequency, form, and function introduction to the Special Issue”. Modern Language Journal, 93 (3), 329–335. DOI: 10.1111/j.1540-4781.2009.00893.x Ellis, N. & Ferreira-Junior, F. 2009. “Construction learning as a function of frequency, frequency distribution and function”. Modern Language Journal, 93 (3), 370–385. DOI: 10.1111/j.1540- 4781.2009.00896.x Gaies, S. J. 1983. “The investigation of language classroom processes”. TESOL Quarterly, 17 (2), 205–217. DOI: 10.2307/3586650 Gass, S. & Mackey, A. 2002. “Frequency effects and second language acquisition: A complex picture?” Studies in Second Language Acquisition, 24, 249–260. DOI: 10.1017/ S0272263102002097 Goldberg, A. E. 2003. “Constructions: A new theoretical approach to language”. Trends in Cognitive Science, 7 (5), 219–224. DOI: 10.1016/S1364-6613(03)00080-9 Goldschneider, J. M. & DeKeyser, R. 2001. “Explaining the ‘natural order of L2 morpheme acquisition’ in English: A meta-analysis of multiple determinants”. Language Learning, 51 (1), 1–50. DOI: 10.1111/1467-9922.00147 Goodman, J., Dale, P. & Li, P. 2008. “Does frequency count? Parental input and the acquisition of vocabulary”. Journal of Child Language, 35, 515–531. DOI: 10.1017/S0305000907008641 Graesser, A. C., McNamara, D. S., Louwerse, M. M. & Cai, Z. 2004. “Coh-Metrix: Analysis of text on cohesion and language”. Behavioral Research Methods, Instruments, and Computers, 36 (2), 193–202. DOI: 10.3758/BF03195564 Gries, S. T. 2008. “Dispersions and adjusted frequencies in corpora”. International Journal of Corpus Linguistics, 13 (4), 403–437. DOI: 10.1075/ijcl.13.4.02gri Gries, S. T. 2010. “Dispersions and adjusted frequencies in corpora: Further explorations”. In S. T. Gries, S. Wulff & M. Davies (Eds.), Corpus Linguistic Applications: Current Studies, New Directions. Amsterdam: Rodopi, 197–212. Hatch, E. 1983. Psycholinguistics: A Second Language Perspective. Rowley, MA: Newbury House Publishers, Inc. Izumi, S., Bigelow, M., Fujiwara, M. & Fearnow, S. 1999. “Testing the output hypothesis”. Studies in Second Language Acquisition, 21 (3), 421–452. DOI: 10.1017/S0272263199003034 Izumi, S. & Bigelow, M. 2000. “Does output promote noticing and second language acquisition”. TESOL Quarterly, 34 (2), 239–278. DOI: 10.2307/3587952 Laufer, B. 2000. “Task effect on instructed vocabulary learning: The hypothesis of involvement”. Selected papers from AILA ’99 Tokyo. Tokyo: Waseda University Press, 47–62. Laufer, B. & Nation, P. 1995. “Vocabulary size and use: Lexical richness in L2 written production”. Applied Linguistics, 16 (3), 307–322. DOI: 10.1093/applin/16.3.307 Frequency effects and second language lexical acquisition 331

MacWhinney, B. 1997. “Second language acquisition and the Completion Model”. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in Bilingualism: Psycholinguistic Perspectives. Mahwah, NJ: Lawrence Erlbaum, 113–142. McDonough, K. & Kim, Y. 2009. “Syntactic priming, type frequency, and EFL learners’ production of Wh-questions”. Modern Language Journal, 93 (3), 386–398. DOI: 10.1111/j.1540- 4781.2009.00897.x Mintz, T. H., Newport, E. L. & Bever, T. G. 2002. “The distributional structure of grammatical categories in speech to young children”. Cognitive Science, 26 (4), 393–424. DOI: 10.1207/ s15516709cog2604_1 Pica, T. 1983. “Methods of morpheme quantification: Their effect on the interpretation of second language data”. Studies in Second Language Acquisition, 6 (1), 69–79. DOI: 10.1017/ S0272263100000309 Robinson, P. & Ellis, N. 2008. A Handbook of Cognitive Linguistics and Second Language Acquisition. London: Routledge. Shockey, L. 2003. Sound Patterns of Spoken English. Malden, MA: Blackwell. DOI: 10.1002/ 9780470758397 Singleton, D. 1999. Vocabulary Learning in Another Language. Cambridge: Cambridge University Press. Swain, M. 1985. “Communicative competence: Some roles of comprehensible input and comprehensible output in its development”. In S. Gass & C. Madden (Eds.), Input in Second Language Acquisition. Rowley, MA: Newbury House, 235–253. Swain, M. 1995. “Three functions of output in second language learning”. In G. Cook & B. Seidlhofer (Eds.), Principle and Practice in Applied Linguistics: Studies in Honour of H. G. Widdowson. Oxford: Oxford University Press, 125–144. Toth, P. D. 2006. “Processing instruction and a role for output in second language acquisition”. Language Learning, 56 (2), 319–385. DOI: 10.1111/j.0023-8333.2006.00349.x Wulff, S., Ellis, N., Römer, U., Bardovi-Harlig, K. & LeBlanc, C. 2009. “The acquisition of tense– aspect: Converging evidence from corpora, cognition, and learner constructions”. Modern Language Journal, 93 (3), 354–369. DOI: 10.1111/j.1540-4781.2009.00895.x Zipf, G. K. (1935). The Psycho-biology of Language. New York: Houghton-Mifflin.

Authors’ addresses Scott Andrew Crossley Tom Salsbury Department of Applied Linguistics Department of Teaching and Learning Georgia State University Washington State University160 Cleveland 34 Peachtree St. Suite 1200, One Park Tower Hall Building Pullman, WA 99164-2132 Atlanta, GA 30303 USA USA [email protected] [email protected] 332 Scott Crossley, Tom Salsbury, Ashley Titak and Danielle McNamara

Ashley Titak Danielle S. McNamara Department of Applied Linguistics Department of Psychology Georgia State University Arizona State University 34 Peachtree St. Suite 1200, One Park Tower 950 S. McAllister Room 237 Building Tempe, AZ 85287-1104 Atlanta, GA 30303 USA USA [email protected] [email protected]

Copyright of International Journal of Corpus Linguistics is the property of John Benjamins Publishing Co. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.