Corpus-Based Frequency Profiling:
Migration to a Word List Based on the British National Corpus
Leah
Gilner* and Franc Morales
The
selection and assessment of ELT materials involve multiple criteria. The use of
frequency word lists to profile the vocabulary makeup of a text is one such
criterion. It provides a quantifiable characterization and classification of
lexical material in terms of corpus-based frequency measures. The process of
vocabulary profiling is not without challenges, first among which is the
identification of a word list adequate for ELT. The choice will determine the
amount of information, if any, that can be derived from a text. This paper
provides an appraisal of a frequency word list based on the British National
Corpus (BNC) and shows the benefits that can be gained by profiling with this
list rather than with the long-established General Service List (West, 1953).
There
are two basic appeals to moving on from the GSL. First, the GSL was compiled
based on data from corpora tallying up to 10 million tokens while the BNC is
ten times larger. The differential in size makes it possible to obtain a more
accurate account of the frequency organization of the English lexicon. Second,
the GSL leaves uninformative gaps when deployed in profiling, ranging from 10%
to 25% depending on the text (Nation and Waring, 1997). Due to the way in which
the GSL was manufactured, expanding this word list is nearly impossible if one
is to follow the original directives and criteria (Faucett et al, 1936; Lorge,
1949; West, 1953). While necessarily observing different criteria, word lists
that supplement the GSL have been proposed (Coxhead, 2000; Xue and Nation,
1984) and investigated (Hyland and Tse, 2007). It can be said that
expandability is, perhaps, the weakest point of the GSL and the best reason for
seeking a replacement. The BNC affords the possibility of addressing this
issue.
However,
since the BNC has not been designed to inform ELT, knowledge of its origin and
composition is an important factor to take into consideration when seeking to
derive information and insight from it. The BNC is a 100-million-word sample,
synchronic, general, monolingual, mixed corpus of present-day British English.
The compilation of the BNC was a collaborative undertaking carried out by
dictionary publishers (Oxford University Press, Longman, Chambers Harrap) and
academic institutions (Oxford University, Lancaster University, the British
Library) with financial backing from British government agencies (Leech et al.,
2001). About 90% of the corpus is comprised of written language, categorized as
imaginative (i.e. fiction) or informative (i.e. non-fiction, expository); most
of the written texts date from 1975 or later (20% of imaginative texts date
from 1960). The 10-million-word subcorpus of spoken language contains samples
recorded between 1991 and 1993; 40% of the samples represent conversational
language use, that is to say, “spontaneous interactions engaged in by some 127
adults aged 15 and over” (Leech et al., 2001, p.2); 60% of the samples
represent task-oriented speech (lectures, sermons, TV/radio programs,
consultations), “those types of […]
spoken activity that were unlikely to be recorded by the conversational
volunteers during a typical day of their lives (Leech et al., 2001, p. 3).
It
stands to reason that not every frequency word list based on the BNC has the
potential to be equally informative in ELT. Nation (2004) inspected the 3,000
most frequent word families in the BNC and found that they contain material
from the GSL as well as the Academic Word List (Coxhead, 2000). The presence of
this academic vocabulary interwoven with words of general service - which are
thought to provide the foundation for subsequent learning (Nation, 2001) - made
it difficult to “decide how the GSL could be replaced” (Nation, 2004, p.12)
when considering, precisely, core vocabulary. Following a different line of
inquiry, Nation (2006) produced 14,000 word families, organized according to
frequency data from the BNC, in order to asses “how large a receptive
vocabulary is needed for typical language use” (p. 59). Of interest, his
analyses reveal important information in those gaps uncharacterized by the GSL.
For example, ‘topic words’ were identified in the 4,000 word family and beyond.
The evolution of this list of 14,000 word families (hereafter, BNC-ELT word list) was not limited to expandability (over the 3,000 originally used) but subsequently included the reorganization of the word families according to the spoken subcorpus of the BNC. As Nation explains in the “readme” file of the RANGE software available from his website, “previously the lists had been sequenced using figures from the whole BNC but because of the overwhelming amount of formal written material this resulted in lists that did not satisfactorily represent informal spoken uses of English”. It should be noted that range was the main criterion used in the creation of the list and that frequency was the second criterion (I.S.P. Nation, personal communication, February 13, 2008).
Although
it is reasonable to assume that further refinement of the BNC-ELT list might
take place, we have adopted this latest revision of the list and have used it
in the analyses carried out in this investigation in order to characterize a
relatively large and varied sample of the kinds of authentic materials that can
be used in ELT. As findings will show, this characterization gives powerful
reason to migrate from the GSL to the BNC-ELT.
We
begin with a closer look at the BNC by inspecting the raw frequency measures it
yields and how these metrics characterize the whole corpus, a generalizable
insight into lexical choice in language use. Following we present an analysis
that illustrates how these same metrics characterize a sample of materials used
in ELT. Some structural properties of the BNC-ELT list will then be presented
together with a description of the ELT corpus compiled for this investigation.
We bring the two together - the BNC-ELT list and the ELT corpus – in a series
of analyses that will allow us to explore the extent and manner in which
profiling can be used to characterize the lexical content of ELT materials. We
close by providing comparative measures between the GSL and the BNC-ELT for
referential purposes that may facilitate the transition from one to the other.
The
importance of lexical frequency in language use cannot be overstated. Analysis
of any reasonable amount of language in use reveals substantial uniformity with
regards to lexical content. The numbers are quite impressive. Nearly 50% of all
language used is confined to 100 words, 75% to 2,000 words, and 85% to 5,000
words (Leech et al., 2001). These numbers show an extremely sloped curve of
distribution where very few items in the vocabulary account for most discourse while
the majority of items in the vocabulary occur with severe, even extreme,
infrequency (Ellis, 2002). The implications for language instruction are clear.
Speakers demonstrate marked preferences when it comes to lexical choice, making
it feasible to isolate a vocabulary of objective value.
Table
1 presents data on the lexical coverage that the most frequent words of the
English language provide for the British National Corpus. According to the data
obtained from Leech et al.
(2001), 397,041 of the 757,087 words (52.44%) in the corpus account for only
0.0039% of the occurrences in the corpus, while 100 of the 757,087 words
(0.0132%) account for 45.8786% of the occurrences in the corpus. The analysis
we provide in Table 1 illustrates this phenomenon in greater detail.
It
is evident that exceedingly few types (words) account for the vast majority of
tokens (occurrences). While the 10 most frequent types occur over 21 million
times, there are 397,041 types (52.44%) that occur only once in the entire
corpus (Leech et al., 2001).
Table
1 shows that the first 100 words (types) occur 45,878,600 times (tokens) in the
BNC. The next set of 100 types contributes 6,373,600 tokens (or 6.3736% of the
corpus) while the next 300 types add 8,358,400 tokens. From here, the column
labeled ‘Difference’ shows how the amount of tokens contributed by the
subsequent addition of types gradually diminishes. The column labeled
‘Cumulative’ grows as the amount of types increases, yet
|
Types |
Tokens |
% of BNC types |
% of BNC corpus |
Difference |
Cumulative |
|
100 |
45,878,600 |
0.0132% |
45.8786% |
|
|
|
200 |
52,252,200 |
0.0264% |
52.2522% |
6.3736% |
6.3736% |
|
500 |
60,610,600 |
0.0660% |
60.6106% |
8.3584% |
14.7320% |
|
1,000 |
67,569,500 |
0.1321% |
67.5695% |
6.9589% |
21.6909% |
|
1,500 |
71,864,900 |
0.1981% |
71.8649% |
4.2954% |
25.9863% |
|
2,000 |
74,950,900 |
0.2642% |
74.9509% |
3.0860% |
29.0723% |
|
2,500 |
77,332,500 |
0.3302% |
77.3325% |
2.3816% |
31.4539% |
|
3,000 |
79,255,000 |
0.3963% |
79.2550% |
1.9225% |
33.3764% |
|
3,500 |
80,828,900 |
0.4623% |
80.8289% |
1.5739% |
34.9503% |
|
4,000 |
82,144,700 |
0.5283% |
82.1447% |
1.3158% |
36.2661% |
|
4,500 |
83,254,100 |
0.5944% |
83.2541% |
1.1094% |
37.3755% |
|
5,000 |
84,214,800 |
0.6604% |
84.2148% |
0.9607% |
38.3362% |
|
5,500 |
85,060,900 |
0.7265% |
85.0609% |
0.8461% |
39.1823% |
|
6,000 |
85,809,200 |
0.7925% |
85.8092% |
0.7483% |
39.9306% |
|
6,500 |
86,480,400 |
0.8586% |
86.4804% |
0.6712% |
40.6018% |
|
7,000 |
87,088,700 |
0.9246% |
87.0887% |
0.6083% |
41.2101% |
Table 1. Breakdown of the
most frequent words in English.
showing that the amount of tokens does not quite double
even after the inclusion of tokens corresponding to 7,000 types. Summing up,
the 100 most frequent types (words) of the language amount to more tokens
(occurrences) than the following 6,900 types (words) combined and, together,
the 7,000 most frequent types account for 87.09% of all tokens in the BNC.
It
is relevant to question if the observed frequency distributions are limited to
large corpora and whether texts of smaller size exhibit similar metrics.
Furthermore, it is pertinent to ask if ELT materials are equally served by this
information. Table 2 answers this question affirmatively by analyzing eight
texts of different sizes and types of discourse (spoken and written). These
texts have been randomly chosen from the ELT corpus that was compiled for this
study and that will be introduced later on. For now, note that these texts have
been arranged so that “beginner” materials are displayed in the upper rows and
“advanced” materials in the lower rows, that is, the Interview script is a transcription of a dialogue of little
complexity while the NYT article is a
piece of news from the New York Times and, therefore, of reasonable difficulty
for advanced learners.
The first observation is that the coverage provided by the
most frequent types (words) in English is superior for the ELT materials than
it is for the BNC (82.65% over 79.25%). The second observation is that while
values fluctuate, trends are uniform and correlate. Approximately half (or
more) of all tokens of any text are confined to the 100 most frequent types in
the language while approximately three-fourths (or more) of all tokens of any
text are confined to the 2,000 most frequent types. The third observation is
that, together, the 3,000 most frequent types offer better coverage of the Interview script than of the NYT article. The fourth observation is
that, as we
|
Source |
Tokens |
100 |
500 |
1,000 |
2,000 |
3,000 |
|
BNC |
100,000,000 |
45.89% |
60.61% |
67.57% |
74.95% |
79.25% |
|
All
texts |
52,169 |
49.91% |
63.97% |
69.63% |
78.29% |
82.65% |
|
Interview
script |
560 |
62.11% |
80.33% |
84.88% |
88.89% |
91.62% |
|
Short
story |
15,935 |
54.95% |
69.65% |
74.97% |
80.95% |
84.76% |
|
Family movie script |
17,730 |
45.81% |
58.65% |
63.39% |
75.93% |
80.48% |
|
News article (intermediate) |
448 |
43.95% |
61.16% |
69.30% |
80.00% |
85.81% |
|
Novel |
10,542 |
51.77% |
66.85% |
72.74% |
79.70% |
84.16% |
|
ESP reading (technology) |
1,602 |
45.56% |
58.78% |
65.15% |
72.14% |
77.02% |
|
IHT
article |
1,555 |
45.71% |
58.80% |
67.64% |
75.61% |
79.00% |
|
NYT
article |
1,265 |
42.28% |
54.68% |
64.45% |
72.25% |
79.89% |
Table 2. Profile of a
variety of texts based on BNC raw frequencies.
look from row to row in descending order, we can see that
profiling does not provide a definite correlation between frequency and
difficulty even though the data shows a tendency in that direction. In other
words, better coverage (in this case, fewer infrequent words) does not necessarily imply lesser difficulty.
The
word lists employed in the two previous analyses were manufactured by
identifying the most frequent unlemmatized types (words) in the English
language (for example, the first ten types are the, of, and, a,
in, to, it, is, to,
and was). For the purposes of ELT,
such a list is not as useful as a list where items are lemmatized and, more
importantly, clustered into word families. A word family refers to a grouping
containing a headword, its inflections, and its closest derivations (Nation,
2001). It is posited that awareness of word family relationships can “greatly
decrease the learning burden of derived words containing known base forms,”
(Nation, 2001, p. 8) although the extent to which this is true will depend on a
given learner’s experience and linguistic background among other things
(Mochizuki and Aizawa, 2000; Sakata, 2007).
One
of the strengths of the BNC-ELT list is that types are clustered into families.
The BNC-ELT is comprised of a total of 50,598 types (words), grouped into 14
sublists of 1,000 families each which in turn are ranked by descending
frequency. That is, the first sublist contains the 1,000 most frequent word
families in the English language, the second sublist contains the following
1,000 most frequent families, and so on. The detail and scope of the BNC-ELT
list makes this contribution an unparalleled resource for the identification of
lexical distributions in ELT materials.
Table
3 shows some of the structural characteristics of the BNC-ELT sublists together
with the code that will be used to refer to them in the discussion that
follows. As mentioned, each sublist contains 1,000 word families. An example of
a family from the sublist SL-01 is ABLE:
able, ability, abler, ablest, ably, abilities, unable, and inability while an example of a family from the sublist SL-14 is ALLURE: allure, allured, allures, alluring, and alluringly.
|
Code |
Families |
Words |
Average |
|
SL-01 |
1,000 |
6,348 |
6.35 |
|
SL-02 |
1,000 |
5,593 |
5.59 |
|
SL-03 |
1,000 |
4,517 |
4.52 |
|
SL-04 |
1,000 |
4,287 |
4.29 |
|
SL-05 |
1,000 |
3,992 |
3.99 |
|
SL-06 |
1,000 |
3,494 |
3.49 |
|
SL-07 |
1,000 |
3,272 |
3.27 |
|
SL-08 |
1,000 |
3,192 |
3.19 |
|
SL-09 |
1,000 |
3,050 |
3.05 |
|
SL-10 |
1,000 |
2,840 |
2.84 |
|
SL-11 |
1,000 |
2,794 |
2.79 |
|
SL-12 |
1,000 |
2,568 |
2.57 |
|
SL-13 |
1,000 |
2,426 |
2.43 |
|
SL-14 |
1,000 |
2,225 |
2.23 |
|
BNC-ELT |
14,000 |
50,598 |
3.61 |
Table 3. Characteristics of
the BNC-ELT list.
We
can see from Table 3 that less frequent families have fewer members (Nation,
2007). The column labeled ‘Average’ quantifies this trend by presenting the
average number of types per family for each of the sublists and for the BNC-ELT
list as a whole.
The
ELT corpus compiled for this investigation consists of eight collections of
texts, each taken from the kinds of sources (newspaper articles, movie scripts,
short stories, novels, etc.) generally referred to as authentic material
(Gilmore, 2007). Table 4 presents some general statistics which serve to inform
on the composition of the ELT corpus compiled for this study. The collections
were compiled from the kind of sources that are often used when the desire is to
provide students with authentic models of naturally-occurring, fluent language
use (Brown and Yule, 1983) and that, in our experience teaching at university
level, often find their way into the classroom. These collections can be said
to illustrate a natural grading (Gilmore, 2007) in terms of content,
presentation, and register (Carter and McCarthy, 1994), ranging from Interview scripts and Short stories which are likely to be
used with less experienced students (i.e. first-year university) to articles from
the International Herald Tribune
(IHT) and the New York Times (NYT)
which might be selected for more experienced and advanced students.
|
Code |
Description |
Items |
Tokens |
Avg. length |
|
CLT-01 |
Interview
scripts |
116 |
49,613 |
453 |
|
CLT-02 |
Short
stories |
18 |
52,413 |
2,958 |
|
CLT-03 |
Family
movie scripts |
19 |
325,581 |
18,937 |
|
CLT-04 |
News
articles (intermediate) |
134 |
72,187 |
538 |
|
CLT-05 |
Novels |
14 |
398,854 |
29,614 |
|
CLT-06 |
ESP
readings (technology) |
31 |
55,486 |
1,866 |
|
CLT-07 |
International
Herald Tribune |
164 |
154,658 |
1,010 |
|
CLT-08 |
New York
Times |
58 |
48,701 |
903 |
|
ELT
Corpus |
All collections |
554 |
1,157,493 |
2,226 |
Table 4. Characteristics of
the ELT corpus.
The
design of the ELT corpus was approached from an ELT practitioner’s perspective
in as much as we wanted to compile an assortment of material that might reflect
the choices made by colleagues in the field. The ELT corpus deliberately
contains collections that have markedly more tokens than others (i.e. Family movie scripts vs. Interview scripts) in order to highlight
the effect text length has on profiling results. We now describe each
collection in turn.
The
Interview scripts collection is
comprised of transcriptions of 116 interviews taking place between speakers of
different backgrounds. The materials can be described as modeling naturally-occurring
interactions, as when people are getting to know each other, and in which
English is used as an international language. The Interview scripts have an average length of about 450 tokens
(occurrences), making them relatively short. From the ELT practitioner’s
perspective, the length, breadth, and depth of the interviews make them
appropriate for less experienced students. Topics are discussed in general
terms and speakers often provide narratives about personal experiences.
The
Short stories collection contains 18
children’s stories (i.e. The Tale of
Peter Rabbit, The Emperor’s New
Clothes, Rapunzel). The stories
are written for a young L1 reading audience but the fictional, imaginative
aspects of the texts can make them entertaining and engaging for L2 learners of
an older age. The stories are lengthy, on average about 3,000 tokens, and could
be a resource for extensive reading in lower and intermediate levels.
The
Family movie scripts collection
contains 19 scripts, on average about 19,000 tokens long, from movies that seem
to be widely-recognized, such as E.T.
and Back to the Future. The nature of
the genre implies that a substantial amount of the discourse comes in the form
of dialogues. To a large extent, the structural and conceptual complexity of
the material is bound by the target audience (parents and children). As family
movies are less likely to involve in-depth development of ideas or elaborate
argumentation, they are deemed most appropriate for intermediate-level
students.
The
News articles (intermediate)
collection includes 134 newspaper articles from Voice of America and the
English version of a well-known Japanese newspaper. Topics vary from economics
and politics to health and education. They are relatively short, on average
about 500 tokens long, and differ from the articles in the IHT and the NYT
collections in terms of lexical and structural complexity as well as depth of
exposition; these factors combine to make this collection accessible to
intermediate-level students.
The
Novels collection contains 14
full-length fiction stories written for adult audiences, each averaging 30,000
tokens in length. The nature of these kinds of texts implies a more in-depth
development of plot and characters than the other collections and is likely to
include examples of spoken discourse in the form of dialogue interwoven in the
narrative. Full-length novels are also likely to make use of a larger and more
varied vocabulary than shorter texts. Given these characteristics, the texts in
this collection represent extensive reading material for students at
advanced-levels.
The
ESP readings (technology)
collections contain 31 texts which describe aspects and constructs related to
computers, the Internet, and electronics. The texts average about 2,000 tokens,
are procedural in nature, and make use of domain-specific, specialized
vocabulary. This collection is deemed to represent intermediate- and
advanced-level material for Science and Engineering majors.
The
International Herald Tribune (IHT)
and New York Times (NYT) collections
contain 164 and 58 news articles and special reports, respectively. In both
cases, the average length is about 1,000 tokens. Topics vary widely and
include: politics, economics, culture, society, travel, sports, health,
fashion, etc. As the name suggests, the IHT targets an international audience
while the NYT, although of international repute, is thought of as a newspaper
of record in the U.S. Thus, even though the two newspapers are owned by the
same company, the role of each may influence not only the treatment and
perspective provided in the texts, but also the style of discourse and
expression. Either collection might be used with advanced-level students.
With
this outline of the collections and corpus in mind, we move on to the analysis
of the ELT corpus by means of the BNC-ELT list. Results for all eight
collections are presented throughout. However, we will focus the discussion on
CLT-01 and CLT-08 as these
collections are similar in size while clearly distinct in difficulty. This
narrow focus will allow us to formulate (weak) propositions regarding the
insights into “difficulty” that profiling affords. Once the data has been
presented and discussed, we will proceed to take into account the results
obtained from the other collections. The reader is encouraged to consider all
results as analyses are presented.
The
reason why the discussion elaborates on the assessment of difficulty by means
of frequency profiling is because we find that it is a relationship that is
established intuitively yet not addressed in the literature on vocabulary
frequency lists. There seems to be an assumption or tendency to assume, for
instance, that frequent words are “easier” than infrequent ones. Intuitively,
it makes sense to consider a word that is rarely used as being, one, of very
specialized application and/or, two, of such low occurrence that it is
difficult for a learner to obtain repeated exposure to it. The interpretation
of results presented hereafter hopes to provide some observations regarding the
equivocal relationship between frequency profiling and learning burden.
A
profile of the ELT corpus using the BNC-ELT list is presented in Table 5. The
lexical material in the ELT corpus belonging to the BNC-ELT list is 98.63%
(last column, bottom row). By collection, the extremes are 99.64% coverage of
CLT-01 (‘lower-level’ texts) and 97.42% coverage of CLT-8 (‘advanced-level’
texts).
|
Code |
CLT-01 |
CLT-02 |
CLT-03 |
CLT-04 |
CLT-05 |
CLT-06 |
CLT-07 |
CLT-08 |
ELT Corpus |
|
SL01
|
91.45% |
82.23% |
79.27% |
80.04% |
82.25% |
76.18% |
76.84% |
76.94% |
80.43% |
|
SL02
|
4.23% |
7.01% |
7.07% |
9.59% |
7.01% |
10.28% |
9.47% |
9.35% |
7.65% |
|
SL03
|
1.62% |
3.72% |
4.67% |
2.93% |
3.94% |
3.62% |
3.18% |
3.32% |
3.83% |
|
SL04
|
0.80% |
1.78% |
2.43% |
2.24% |
1.63% |
2.71% |
2.80% |
2.95% |
2.13% |
|
SL05
|
0.42% |
1.36% |
1.39% |
1.10% |
1.34% |
1.36% |
1.47% |
1.36% |
1.32% |
|
SL06
|
0.34% |
0.76% |
0.97% |
0.94% |
0.72% |
0.96% |
1.03% |
0.86% |
0.85% |
|
SL07
|
0.17% |
0.51% |
0.63% |
0.46% |
0.53% |
0.57% |
0.63% |
0.64% |
0.56% |
|
SL08
|
0.15% |
0.46% |
0.39% |
0.44% |
0.35% |
0.82% |
0.54% |
0.53% |
0.42% |
|
SL09
|
0.13% |
0.23% |
0.42% |
0.21% |
0.30% |
0.34% |
0.41% |
0.41% |
0.34% |
|
SL10
|
0.11% |
0.33% |
0.58% |
0.25% |
0.26% |
0.59% |
0.37% |
0.27% |
0.38% |
|
SL11
|
0.09% |
0.15% |
0.28% |
0.16% |
0.28% |
0.22% |
0.26% |
0.27% |
0.25% |
|
SL12
|
0.09% |
0.16% |
0.17% |
0.15% |
0.15% |
0.25% |
0.21% |
0.17% |
0.17% |
|
SL13
|
0.02% |
0.15% |
0.34% |
0.13% |
0.17% |
0.17% |
0.21% |
0.22% |
0.22% |
|
SL14
|
0.02% |
0.16% |
0.08% |
0.06% |
0.07% |
0.23% |
0.15% |
0.15% |
0.09% |
|
BNC-ELT |
99.64% |
99.03% |
98.70% |
98.69% |
98.99% |
98.28% |
97.58% |
97.42% |
98.63% |
Table 5. Token coverage of
the ELT corpus by the BNC-ELT list and sublists.
Inspection
of the column labeled ‘ELT Corpus’ shows that sublist SL-01 (the first 1,000
families and, thus, the most frequent in the language) accounts for 80.43% of
words in the ELT corpus. The second sublist of families, SL-02, accounts for
7.65% of the vocabulary and, together with the first sublist, amounts to 88.08%
of all words in the ELT corpus. A clear drop in use is evident as families
become more infrequent; a trend that is disrupted in only two occasions (SL-10
and SL-13).
The
coverage analysis also reveals that collections deemed more adequate for
lower-level learners have higher concentrations of vocabulary in the first
sublist (SL-01) than those collections containing advanced material. The data
for SL-01 indicates that this sublist accounts for 91.45% of CLT-01 yet for a
much smaller share of CLT-06, CLT-07, and CLT-08 (76.18%, 76.84%, and 76.94%
respectively). From this data, it is possible to propose that, in general,
there might be a connection between lexical frequency and level of difficulty.
In other words, a characteristic of advanced texts might reside in the use of a
comparatively infrequent vocabulary and, conversely, that a characteristic of
beginner texts might reside in the limitation of vocabulary to frequently
occurring - i.e. more common - words. The proposition might be intuitively
correct but it is of importance to note that the BNC-ELT list provides a means
for the quantification of this characteristic.
We
now examine the amount of families (and types) from each sublist in the BNC-ELT
list that is used by each of the collections in the ELT corpus. The data is
first presented globally, that is, regarding the ELT corpus as a whole and
without detailing use by collection.
|
Code |
Families |
Percent |
Types |
Percent |
Tokens |
|
SL-01 |
1,000 |
100.00% |
4,563 |
71.88% |
930,981 |
|
SL-02 |
994 |
99.40% |
3,822 |
68.34% |
88,559 |
|
SL-03 |
981 |
98.10% |
2,972 |
65.80% |
44,325 |
|
SL-04 |
954 |
95.40% |
2,486 |
57.99% |
24,642 |
|
SL-05 |
902 |
90.20% |
2,018 |
50.55% |
15,271 |
|
SL-06 |
860 |
86.00% |
1,640 |
46.94% |
9,817 |
|
SL-07 |
763 |
76.30% |
1,284 |
39.24% |
6,430 |
|
SL-08 |
698 |
69.80% |
1,149 |
36.00% |
4,837 |
|
SL-09 |
654 |
65.40% |
1,030 |
33.77% |
3,943 |
|
SL-10 |
626 |
62.60% |
925 |
32.57% |
4,386 |
|
SL-11 |
588 |
58.80% |
820 |
29.35% |
2,914 |
|
SL-12 |
481 |
48.10% |
633 |
24.65% |
1,958 |
|
SL-13 |
469 |
46.90% |
627 |
25.85% |
2,493 |
|
SL-14 |
313 |
31.30% |
388 |
17.44% |
1,097 |
|
BNC-ELT |
10,283 |
73.45% |
24,357 |
48.14% |
1,141,653 |
Table 6. Use of BNC-ELT
lists in ELT corpus (global data).
Table
6 shows that all families in sublist SL-01 (the most frequent in the language)
are used in the ELT corpus. As a family is a collection of inflected and
derived forms, it is important to note that 71.88% of all types in SL-01 are
found in the ELT corpus and that these types account for 930,981 of all of the
tokens (80.43% as shown in Table 5). As in all other analyses, sublists of more
infrequent families are used progressively less both at the family and type
(word) level, a fact that correlates with the amount of tokens they account for
in the ELT corpus. The data, again, supports the validity and adequacy of the
BNC-ELT list for the purpose of assessing ELT materials beyond the scope of the
GSL.
From
this data, it is possible to formulate a second proposition, namely, that there
might be a connection between lexical variety and level of difficulty. In other
words, one way to characterize advanced texts might be in terms of the use of a
comparatively rich vocabulary. Table 7 makes a clear case for this proposition
|
|
CLT-01 |
CLT-02 |
CLT-03 |
CLT-04 |
CLT-05 |
CLT-06 |
CLT-07 |
CLT-08 |
ELT Corpus |
|
SL-01
|
85.00% |
83.60% |
99.00% |
96.50% |
98.60% |
92.80% |
98.40% |
95.90% |
100.00% |
|
SL-02
|
47.80% |
57.50% |
92.40% |
78.80% |
91.80% |
77.50% |
94.70% |
79.70% |
99.40% |
|
SL-03
|
23.10% |
43.70% |
83.70% |
49.30% |
82.30% |
50.80% |
80.20% |
52.90% |
98.10% |
|
SL-04
|
14.20% |
29.20% |
69.90% |
39.10% |
67.20% |
39.30% |
71.80% |
42.60% |
95.40% |
|
SL-05
|
7.90% |
21.40% |
59.40% |
26.70% |
54.50% |
27.70% |
58.30% |
29.80% |
90.20% |
|
SL-06
|
6.10% |
17.30% |
48.80% |
19.60% |
45.70% |
20.40% |
45.90% |
21.40% |
86.00% |
|
SL-07
|
3.60% |
11.10% |
37.60% |
11.70% |
36.10% |
14.30% |
38.00% |
17.00% |
76.30% |
|
SL-08
|
3.00% |
8.60% |
33.60% |
12.10% |
32.50% |
11.70% |
30.90% |
12.90% |
69.80% |
|
SL-09
|
2.20% |
7.40% |
31.60% |
8.30% |
30.40% |
8.40% |
26.20% |
11.80% |
65.40% |
|
SL-10
|
1.80% |
6.30% |
28.40% |
8.60% |
26.50% |
7.70% |
24.30% |
9.70% |
62.60% |
|
SL-11
|
1.10% |
4.80% |
25.50% |
6.60% |
24.80% |
6.70% |
24.10% |
7.70% |
58.80% |
|
SL-12
|
1.00% |
4.20% |
17.60% |
4.10% |
18.70% |
5.70% |
15.20% |
6.00% |
48.10% |
|
SL-13
|
0.70% |
5.10% |
18.70% |
3.80% |
18.90% |
5.50% |
15.20% |
6.30% |
46.90% |
|
SL-14
|
0.50% |
2.40% |
10.50% |
2.60% |
10.70% |
3.80% |
11.30% |
4.50% |
31.30% |
|
BNC-ELT |
14.14% |
21.61% |
46.91% |
26.27% |
45.62% |
26.59% |
45.32% |
28.44% |
73.45% |
Table 7. Use of BNC-ELT
words in ELT corpus (family level data).
Collection
CLT-01 makes use of 85.00% of the families in sublist SL-01, 47.80% of the
families in SL-02, 23.10% of those in SL-03, 14.20% of those in SL-04, and,
beyond this point, from 7.90% to 0.50% of the families in the remaining
sublists. In contrast, collection CLT-08, makes use of significantly more
families in every one of the sublists in the BNC-ELT list, about double for
SL-02 and SL-03, triple for SL-04 through SL-06, quadruple for SL-07 through
SL-09, and so on.
When
considering the use of actual types (words) from each sublist in the BNC-ELT
list, the data shown in Table 8 becomes more uniform although it still
correlates with trends seen in Table 6 and 7.
From
the data shown in Table 3, we know that the amount of types per sublist
decreases according to the relative frequency of the sublist. Sublist SL-01
contains 6,384 types and sublist SL-02 contains 5,593 types while sublists
|
|
CLT-01 |
CLT-02 |
CLT-03 |
CLT-04 |
CLT-05 |
CLT-06 |
CLT-07 |
CLT-08 |
ELT Corpus |
|
SL-01 |
26.54% |
28.39% |
48.72% |
41.73% |
52.03% |
37.15% |
53.04% |
39.74% |
71.88% |
|
SL-02 |
12.05% |
17.63% |
40.78% |
27.96% |
44.06% |
25.73% |
44.09% |
27.71% |
68.34% |
|
SL-03 |
6.24% |
14.43% |
38.90% |
16.29% |
40.85% |
16.67% |
32.79% |
16.27% |
65.80% |
|
SL-04 |
3.76% |
9.07% |
28.57% |
13.32% |
29.44% |
12.74% |
28.97% |
13.58% |
57.99% |
|
SL-05 |
2.33% |
6.99% |
23.12% |
8.59% |
23.72% |
8.92% |
22.62% |
9.44% |
50.55% |
|
SL-06 |
2.12% |
6.01% |
20.89% |
6.78% |
20.78% |
7.04% |
18.20% |
7.33% |
46.94% |
|
SL-07 |
1.28% |
4.22% |
15.77% |
4.10% |
16.44% |
5.17% |
14.52% |
5.93% |
39.24% |
|
SL-08 |
1.07% |
3.07% |
14.10% |
4.42% |
14.25% |
4.23% |
12.06% |
4.73% |
36.00% |
|
SL-09 |
0.95% |
2.46% |
13.90% |
2.92% |
13.57% |
3.05% |
10.43% |
4.33% |
33.77% |
|
SL-10 |
0.63% |
2.64% |
12.68% |
3.35% |
12.43% |
3.20% |
9.89% |
3.70% |
32.57% |
|
SL-11 |
0.47% |
1.90% |
10.95% |
2.58% |
11.42% |
2.61% |
9.81% |
2.86% |
29.35% |
|
SL-12 |
0.39% |
1.75% |
8.14% |
1.64% |
9.27% |
2.49% |
6.50% |
2.45% |
24.65% |
|
SL-13 |
0.33% |
2.39% |
9.27% |
1.94% |
9.93% |
2.47% |
7.09% |
2.76% |
25.85% |
|
SL-14 |
0.22% |
1.17% |
5.44% |
1.30% |
5.48% |
1.89% |
5.53% |
2.25% |
17.44% |
|
BNC-ELT |
6.18% |
9.66% |
24.94% |
13.34% |
26.14% |
12.70% |
24.29% |
13.57% |
48.14% |
Table 8. Use of BNC-ELT
types in ELT corpus (type level data).
SL-13 and SL-14 contain 2,426 and 2,225 types,
respectively. Resolving the percentages shown in Table 7, we see that 1,694
types from sublist SL-01 are used in collection CLT-01 while 2,523 types from
the same sublist are used in collection CLT-08. Considering all 14 sublists
together, collection CLT-01 uses 3,127 types from the BNC-ELT list and this
number accounts for 99.64% of the collection (see Table 5). In contrast,
collection CLT-08 uses more than double that amount - specifically 6,866 types
- from the BNC-ELT list, this amount accounting for 97.42% of the collection.
It is easy to see that collection CLT-08 uses a larger vocabulary than
collection CLT-01.
As
we mentioned, the interpretation of results must be done with caution as the
relationship between the frequency of words and their “difficulty” is not
necessarily unequivocal. With this in mind, we now take into account the
results from all collections. The profiling information obtained does not show
a uniform evolution from CLT-01 to CLT-08 (Tables 5, 7, and 8). For example,
CLT-03 uses 24.94% of the types in the BNC-ELT list while CLT-8 uses a little
more than half that amount (Table 8). When one takes into consideration that
CLT-03 (Family movie scripts) makes
use of a more varied vocabulary than CLT-08 (New York Times articles), our second proposition proves invalid. It
couldn’t be any other way, a New York Times article is almost necessarily more
difficult for a learner than a family movie and while, this might not be so in
particular instances, it is certainly so in general. Since these two
collections contain not one but 19 scripts and 58 articles respectively, we
must find reason for the counter-intuitive results.
The
explanation and the reason for caution in interpretation reside in, at least,
two observations: first, frequency is a measure of probability of occurrence;
second, frequency can influence semantic precision and collocational variation.
Regarding probability, the word astronaut
has a frequency of less than one occurrence per million words (Leech et al.,
2001), meaning that it has about 1/1,000,000 chance of occurring in a randomly
chosen text (in contrast with the word people
which has a chance of occurrence above 1/1,000). If a text (ELT or otherwise)
of, say, 300 tokens in length contains several occurrences of the type astronaut, we know that this word is
behaving abnormally, i.e. it is defying its probability of occurrence.
Profiling allows us to automatically detect these divergences and as Nation
(2006) points out, there is cause to consider if the appearance of infrequent
types in a text might not imply they are ‘topic words’ or, simply, words of
particular import. A text that includes several occurrences of the word astronaut is likely to be related, in
some way or other, to space exploration.
The
flip side of probability of occurrence is that it increases in tandem with the
size of a text, that is, the larger the text the better are the chances that
infrequent words appear. Simply put, the collection CLT-03 contains 325,581
tokens while CLT-08 contains 48,701 tokens, that is, CLT-08 amounts to 14.95%
the size of CLT-03. And so, the probability of occurrence of infrequent words
is much larger in CLT-03 than it is in CLT-08 even though the former collection
is deemed to be easier for learners. The data shown in Tables 5, 7, and 8
reflects the effect of size.
Semantic
precision and collocational variation generally go hand in hand. A word such as
‘point’ can be used as a noun and as a (transitive and intransitive) verb, each
part of speech dictating different collocational partners. Furthermore, the
word point has 46 different meanings
(Webster’s New World, 2006) and its highly polysemous nature implies a wide
range of collocational relationships. In contrast, the word astronaut has but one single meaning and
part of speech (Webster’s New World, 2006) implying that its collocational
complexity will be limited. Frequency-wise, the word point occurs 484 times per million words as a noun and 142 times
per million as a verb, that is, the word point
(SL-01) is over 600 times more frequent than the word astronaut (SL-11). Comparatively speaking, therefore, a higher
frequency might imply a heavier learning burden. However, there are cases were
the converse may also hold. The precision of meaning that infrequent words
exhibit can also imply a greater degree of difficulty for a learner because of
the fine distinction of meaning they might convey. Examples can be fallacy (SL-08), assiduous (SL-11), adulation
(SL14), or maladroit (unlisted, 0.08
times per million) all of which have one or more approximate synonyms of
relatively general meaning and applicability in more frequent words, for
example, lie (SL-01), persistent (SL-04), praise (SL-03), or unskillful
(SL-02), respectively.
These
observations have greatly simplified the notion of learning burden as it is
outside the scope of the paper. The intention has been to show that profiling
should not be used to determine the difficulty of a text in general or its
vocabulary in particular unless the results are interpreted with caution. Granted,
results of this investigation have shown that, given similarly sized texts, it
is possible to use profiling to differentiate the extreme cases, i.e. the
easiest from the most difficult.
Summing
up, the BNC-ELT list provides detailed and exhaustive information about the
lexical composition of the ELT corpus. Unlike the GSL, it leaves minimal
uninformative gaps - regardless of the difficulty of the text - ranging from
0.46% to 2.58% for each collection and 1.47% for the entire ELT corpus of
1,157,493 tokens (Table 5). Results also make it possible to formulate two
(weak) propositions: first, comparatively more difficult texts demonstrate a
tendency to use a larger amount of infrequent vocabulary; second, comparatively
more difficult texts demonstrate a tendency to use a wider, more varied
vocabulary.
The
discussion so far has shown that the extension of coverage provided by the
BNC-ELT (over the GSL) is informative. We now turn to the first two sublists
(SL-01 and SL-02) to see what to expect when migrating from the GSL to the
BNC-ELT. As previously mentioned, Nation (2004) conducted a comparison of the
3,000 most frequent families in the BNC against the GSL and AWL. His results
showed that much of the content of the lists was shared and that the coverage
provided (of the corpora he employed) by each was quite similar. In general,
the 2,000 most frequent families from the BNC provided marginally better
coverage than the GSL with the exception of fiction texts, in which case, the
coverage provided by the GSL was superior (again marginally) than that of the
2,000 families from the BNC.
Code
|
CLT-01 |
CLT-02 |
CLT-03 |
CLT-04 |
CLT-05 |
CLT-06 |
CLT-07 |
CLT-08 |
ELT Corpus |
SL-01
|
91.45% |
82.23% |
79.27% |
80.04% |
82.25% |
76.18% |
76.84% |
76.94% |
80.43% |
|
SL-02 |
4.23% |
7.01% |
7.07% |
9.59% |
7.01% |
10.28% |
9.47% |
9.35% |
7.65% |
|
|
|||||||||
|
GSL-01 |
87.25% |
84.39% |
78.04% |
79.33% |
83.59% |
73.74% |
75.32% |
75.08% |
80.02% |
|
GSL-02 |
5.11% |
6.60% |
7.87% |
5.24% |
6.58% |
7.92% |
5.65% |
5.96% |
6.71% |
Table 9. Coverage of the ELT
corpus by the GSL and BNC-ELT first two sublists.
The
results shown in Table 9 concur with Nation’s. We used the first 2 sublists
(SL-01: 1,000 families; 6,348 types. SL-02: 1,000 families; 5,593 types) from
the BNC-ELT and the GSL (GSL-01: 998 families; 4,119 types. GSL-02: 988
families; 3,708 types). Overall coverage of the ELT corpus by both lists is
strikingly similar despite the fact that the GSL-01 amounts to only 64.88% the
size of the SL-01 and GSL-02 amounts to only 66.29% of the size of SL-02. This
will not come as a surprise to those familiar with the origin and content of
the GSL. It is a remarkably well-manufactured word list. Note that collections
CLT-02 (short stories) and CLT-05 (Novels) are marginally better served by the
GSL, again in accord with Nation’s data.
In
regards to range, it is of interest to see how these two lists (SL-01 + SL-02:
2,000 families; 11,941 types. GSL: 1,986 families; 7,827 types) work across
collections. Table 10 shows the percentage of each word list that appears in
all 8 collections (right-most column), only 7 collections, only 6 collections,
and so on, until we are left with the percentage of words that do not appear in
any collection (left-most column).
|
# of collections |
0 |
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
|
SL-01 + SL-02 |
0.30% |
0.60% |
0.90% |
2.05% |
3.80% |
7.45% |
15.25% |
24.70% |
44.95% |
|
GSL |
0.76% |
0.50% |
1.21% |
4.03% |
5.74% |
10.32% |
12.99% |
21.90% |
42.55% |
Table 10. Amount of words
from each word list that appears in up to 8 collections (range).
Again,
results are strikingly similar for both lists, especially when one takes into
consideration that only 3.4% of SL-03 is found in all 8 collections, the trend
continuing to decline sharply, 0.9% of SL-04, 0.3% of SL-05, 0.1% of SL-06, and
zero beyond this point. It is with this data in mind that we echo Nation’s
comment regarding the difficulty of finding a replacement for the GSL in
regards to a core vocabulary of general service.
Despite
the agreement in coverage and range between the GSL and SL-01+SL-02,
differences in content exist. Nation (2004) provided information from the GSL
perspective, that is, showing how many of its word families could be found in
the 3,000 most frequent BNC word families. His analysis also revealed, as we
mentioned previously, that 80% of the AWL was also present among these 3,000
word families. Unsurprisingly, our analyses again reveal corresponding results,
the only exception regarding the inclusion of academic vocabulary which in our
data is lowered to 67.36% (the remainder of the AWL is present in the BNC-ELT
but at lower frequency levels). It appears that the reorganization of the
14,000 word families according to the spoken subcorpus of the BNC has, indeed,
produced a word list more appropriate for ELT.
|
alright |
Christ |
kid |
score |
|
America |
client |
lad |
Scotland |
|
awful |
county |
London |
switch |
|
bet |
Europe |
minus |
television |
|
bloke |
feed |
non |
thou |
|
bother |
France |
okay |
traffic |
|
brilliant |
Germany |
pence |
video |
|
Britain |
guy |
pension |
wee |
|
budget |
hell |
quid |
x |
|
chap |
Jesus |
reckon |
|
Table 11. The 39 families
from SL-01 not present in the GSL + AWL.
For
those familiar with and migrating from the GSL, Table 11 shows the 39 families
in SL-01 that are not found in the GSL or AWL. The underlined words are proper
nouns that the GSL makers intentionally excluded as were words such as bloke, chap, wee, or pence, on account of lacking
universality (Faucett et al, 1936). Differences increase when considering SL-02
in which 212 new families are introduced.
In
conclusion, comparison of content and coverage between SL-01 + SL-02 and the
GSL do not make a clear case as to which word list might be “better” in regards
to the identification or isolation of a vocabulary of general service. However,
the scope of the BNC-ELT is so much larger that there is no question that it
provides a more detailed characterization and classification when used in
vocabulary profiling. Moreover, from the perspective of the ELT practitioner,
migrating from the GSL to the BNC-ELT does not involve a sacrifice of
established expertise or practices as the GSL can be considered a sublist of
the BNC-ELT in terms of content and application.
Bibliography
Brown,
G. and Yule, G. 1983. Teaching the spoken language. Cambridge: Cambridge
University Press.
Carter,
R. and McCarthy, M. 1988. Vocabulary and Language Teaching. New York: Longman.
Coxhead,
A. 2000. A new Academic Word List. TESOL Quarterly, 34, 2, pp. 213-238.
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.
Faucett,
L., H. Palmer, E.L.Thorndike, and M. West. 1936. Interim Report on Vocabulary Selection. London: P.S. King and
Son, Ltd.
Gilmore, A. 2007. Authentic materials and authenticity in foreign language learning. Language Teaching, 40, pp. 97-118.
Hyland, K. and Tse, P. 2007. Is there an “Academic Vocabulary”? TESOL Quarterly, 41, 2, pp. 235-253.
Leech, G., P. Rayson, and A. Wilson. 2001. Word Frequencies in Written and Spoken English. Harlow: Pearson Education Limited.
Lorge,
I. 1949. The Semantic Count of the 570
Commonest Words. New
York: Teachers College, Columbia University.
Mochizuki, M. and K. Aizawa. 2000. An affix acquisitional order for EFL learners: An exploratory study. System, 28 pp. 291-304.
Nation, P. and R. Waring. 1997. Vocabulary size, text coverage, and word lists. In Schmitt and McCarthy (eds.).
Nation, I.S.P. 2001. Learning vocabulary in another language. Cambridge: Cambridge University Press.
Nation, I.S.P. 2004. A study of the most frequent word families in the British National Corpus. In Bogaards and Laufer (eds.).
Nation,
I.S.P. 2006. How large a vocabulary is needed for reading and listening? The
Canadian Modern Language Review, 63, 1, pp. 59-82.
Sakata,
N. 2007. How do Japanese EFL learners comprehend derivatives?: A qualitative
analysis from the perspective of vocabulary expansion. JACET Journal,
45, pp. 15-29.
Webster’s New World College Dictionary (4th ed.) 2006. Cleveland, OH: Wiley Publishing, Inc.
West,
M. 1953. A General Service List of
English Words. London: Longman, Green and Co.
Xue,
G. and Nation, I.S.P. 1984. A university word list. Language Learning and
Communication, 3, pp. 215-229.