The types of texts in the corpus correspond to three different levels of complexity for the learning situation. The textbook is placed in the first category of introductory and informative types, in agreement with Johns (1997, p. 46). At a middle plane, we have reports and e-discussions (i.e., electronic sets of messages in academic forums), also categorized as descriptive types by Henry (2000). At the third (higher) stadium, we identify journal articles discussing research results and thus presenting a more specialized academic discourse type (Conrad, 1996). In relation to subject matter, as can be checked in Figure 1, five main subjects are explored, seen by students along their respective majors (albeit in different years – e.g., Computer Science people learning Statistics in third year, and Business students in first and second years). Contrary to being an inconvenience, such slight variations can contribute to the inclusion of different genres and text types for each separate subject.

            To select the various sources, advises from co-workers and graduate students can orientate our search. Some items, such as Díaz Martín, Veloso & Rodríguez García (1999), written by colleagues in Computer Science, were pertinent, accessible via the professors’ websites (at University of Extremadura, the Moodle platform http://campusvirtual.unex.es). Also, electronic databases allows for the identification and classification of the types of texts sought. One example is the Kluwer engine for academic material related to BIT (www.kluweronline.com) and the site `Global Edge´ for electronic files and digital forums (www.globaledge.com / http://globaledge.msu.edu/index.asp), where the corpus designer may examine documents by topic and text type. 

            These corpus sources in Figure 1 were selected and stored in the year 2000. As a result, many are no longer available on the Web or electronically (only as electronic corpus sources in University of Extremadura Moodle platform mentioned above). In these sources, the textual and visual elements that do not contribute significant lexical information were discarded. For instance, in the e-discussions, retrieved from an academic forum on the web site ‘Global Edge’, such items as addressers, addresses, and proper names (e.g., university, cities, etc) were omitted. Thus, only plain text is compiled, at times with just a few lines per message, as in this e-discussion example for the topic of Accounting:

Other e-discussion messages were as long as 700-800 words. A key factor in the corpus distribution and design was to keep in mind the need for balance during text retrieval and the compilation of the corpus. This meant that, in the overall corpus design, a similar number of words (ranging between 30,000 and 45,000 words) should be met for each subject category displayed in Figure 1 (e.g., from 32,265 words in M.I.S. to 48,340 in Statistics). This distribution is done so that no particular group of texts in one subject, when fed into the wordlist tool of the concordance software (WordSmith Tools, Scott, 2000), may yield a much lower or higher amount of words than the rest. The DCL (Detailed Consistency List) that takes the five subject-driven wordlists and contrasts items in them should, in fact, contain similar numerical data for lexical analysis.   

            The lexical analysis is performed by following the DCL for the whole BIT corpus. As a result, the main goal is to display the top items that most frequently co-occur across the subject categories. The lexical core would represent the highest degree of general or common academic items shared by the two disciplines.

            The academic elements, regarded as semi-technical in many cases (e.g., Thurstun and Candlin, 1998; Nation, 2001) tend to contain a significant semantic basis. The total amount of such items should be around 500 node words – a basic-to-intermediate lexical knowledge to be expected for our learning context. To obtain this estimate, the total number of words or tokens in the corpus (652,034) is divided into the total number of types (distinct words in the corpus – 21,963). The result is then multiplied by the standardized token-to-type ratio established by Scott (2000): the number of distinct types per 1000 tokens (i.e., 37.12 in the corpus). The resulting figure is then divided in half due to the need to adapt to a medium level of learning (i.e., intermediate; should it have been lower, the amount would have been divided into three, whereas for a high or high/intermediate level, the core academic vocabulary should be the whole figure, about 1000 words according to this computation). The result (551) is rounded up as the amount mentioned above. Some scholars (e.g., Nation, 2001; Flowerdew, 2001) also describe concrete lexical amounts according to such different language commands.

            The 500 words selected are listed in the detailed consistency list, which specifies word repetition according to each of the five subjects of the corpus. By examining high frequency and dispersion of the items on the list, common use of such elements in the corpus is made a primary factor. An example is the use of the verb lead as a common core word (in agreement with other academic lists – e.g., Coxhead, 1998), since it appears in all five subjects; however, a synonym like the verb cause, also considered common core academic by Coxhead (1998), would not be included in our case, since it does not appear in all five categories made.

            The corpus-based 500-word glossary is organized by listing key word combinations and expressions found across various text categories. The arrangement of the collocational items is based on frequency and dispersion by using their t-score, which indicates that the given combination is not due to chance. This statistical score should be above required measurements (see, for instance, Church, Gale, Hanks & Hindle, 1991) to consider the lexical elements as common core. For instance, to take the same example as above, the verb lead to combines with the adjectives better, improved, greater, and different. In all cases, given their co-occurrence frequencies, the resulting t-scores are higher than 2 (minimum required). As a result, in the entry for the verb lead, the glossary should contain different examples of constructions found with those adjectives (e.g., lead to a better environment, lead to a greater understanding, etc). In contrast, a combination like lead to + higher appears with a t-score below the minimum value. Thus, the phrase can lead to a higher turnover would be discarded for the glossary. The same thing applies to collocates preceding the node word: for instance, expected to lead to is included (t-score above 2.0), while other forms like necessarily lead to and automatically lead to are excluded (t-scores below 2).

ESP teaching and glossary development

The glossary and activities based on it were made available on the web (at our university links, e.g., http://epcc.unex.es, www.unex.es/lengingles/ALejandto.htm, and the Moodle platform mentioned above). The undergraduate university students could thus access the alphabetical and frequency lists of academic words with which to answer various on-line questions on word order and formation. For example, given a set of words such as accounting, account, accounted, accounts, they have to find the most frequent item and determine its collocations (with verbs, nouns, adjectives, and so forth); then, to translate such constructions into Spanish. The aim of these weekly activities is to familiarize students with the glossary, but not to build lexical knowledge. In fact, according to most students’ answers in the post-task questionnaire (see Appendix) given out at the end of the course, the wordlist-based activities are fairly easy but do not contribute to their assimilation of lexical items (questions 4 and 5).

            Instead, what seems to qualify as worthwhile lexical development has been a direct approach to the glossary, even though students are divided concerning the type of lexical material explored and how it was organized (see questions 7 and 8 in the post-task questionnaire). They seem to particularly favour the use of lexical skimming and scanning in order to figure out expressions and write them in different exercises, especially translation (see questions 6, 9 and 10).

            The most common type of activities based on the corpus tends to involve the use of academic collocates (see Appendix for the test given to learners, where matching, fill-in-the-gap, cloze, translation, and writing exercises are included). Learners tend to value as highly positive to be able to decode academic collocations from Spanish into English (question 12). In addition, related to this is the notion of having to pair up lexical items derived from common corpus-driven expressions (e.g. exchange + information with). Collocations were examined extensively in the glossary. For instance, matching exercises were developed as a need to ‘visualize’ the company that node words in the glossary keep. Both commonly used and restrained combinations were explored in this way. In the Appendix (in the test activity A), see, for instance, a more general expression like may increasingly be in contrast with do a significance test, more typical within Statistics texts. In this test, nonetheless, some types of activities involving collocation work were evaluated as more convenient than others, according to the students. For instance, the activity of matching collocates is seen as less productive for lexical development.

            In turn, Fill-in-the-gap exercises (e.g., activity B in test) derive from corpus analysis focusing on lexical development within co-texts or concordance lines. The basis for this type of activity is given by key concordance lines where explicit use of collocations is shown. This type of exercise is valued more highly by students in the questionnaires. Its development corresponds to the need for actual co-texts where node words are found. The selection of these concordance lines is made in agreement with semantic units unfolded for each expression; for instance, for the item run on, the key structure in this corpus is run on + computerized device. Students seem to favour this kind of help in the exercise, as it apparently improves their recognition of the expressions.

            Thirdly, cloze exercises (exercise C in test) also originate as a consequence of corpus-based approaches. In this case, textual chunks containing words that typically co-occur with a given node may be easily spotted and provided to students (e.g., by conducting concordance searches of words in context). This type of lexical gapping is similar to Coxhead’s AWL exercises on the web (http://gpoulard.tripod.com/index.htm), where core items from the lists can be automatically removed so that learners can work with them in the paragraphs. Nonetheless, the Cloze activity with whole paragraphs is not evaluated as important by students – see question 15 in Appendix – even though more context for the lexical items is provided.

            The other three activities given in the test are two translations of paragraphs (direct and reversed translations) and a short written composition. According to the questionnaires, in terms of communicative skills, students judge that it was their writing which benefited most (question 16 in the Appendix). Then, according to their answers given for question 17, their favourite topics to write about are socio-technical (e.g., giving opinions about advantages and disadvantages on the use of technology in society). Regarding the least improved skill, most answers refer to listening (questions 18) because in many cases (question 19), students claim that there should have been more audios and videos in class or on the Web (see activities on the webpages referred above). Finally, the majority of the students (more than 50 percent) favour the integration of the glossary in the ESP class (question 20), reasoning in some cases that it provides clear evidence about the important language that they should know for their academic work.

            Given the percentages of the 40 surveyed students (see Appendix), in order to either corroborate or contradict the stated ideas and opinions, the post-test described above was evaluated by the teacher. This evaluation consisted of the six types of exercises named, focusing on glossary-based lexical items; each activity was evaluated from 1 to 10 (a 5 as a passing grade). In the test, the activity done best by students is the English-to-Spanish translation, as can be examined in Figure 2, followed by the Matching exercise. The one done with the most mistakes is the Cloze exercise.