2(4), August 1985, page 45

Computer-Assisted ESL Research and Courseware Development

Gerard M. Dalgish


This paper describes two applications of computers in a program for 350 ESL students whose essays on standardized Writing Skills Assessment Tests (WATS) placed them in ESL writing courses in Baruch College, City University of New York (CUNY). The first application involves using computer-assisted research to determine the kinds of errors most commonly exhibited by these students, to document the effects of first language interference if any, and to individualize instruction for the students by first language where appropriate. The second application involves developing courseware based on this research, integrating the research data into a CAI writing program that focuses on the common errors and particular first-language interfering structures of our students. This paper has two main sections. The first section contains a brief discussion of the research: the data-gathering strategy and an analysis of the results for two error patterns (subject-verb agreement and prepositions). The second section

COMPUTERS and COMPOSITION 2(4), August 1985, page 46

contains a description and demonstration of some of the CAI courseware developed specifically for those areas and for our student population. Two important characteristics of this software are its ability to randomly generate sentences (by this I mean it can create sentences from randomly selected structures, nouns, verbs, phrases, etc.--thus ensuring that all students do not see the same "canned" exercises), and its ability to branch to particular sentence types to individualize instruction according to students' first languages. To my knowledge, no commercially available ESL software combines both features.


The data for our research consisted of sentences taken from the WATS and also from students' essays written in class, to ensure the authenticity of the written sample. Nearly 3,000 sentences from these sources were entered in the CUNY system's mainframe IBM VM360-370 and accessed with a database. The totals of sentences in each language are listed in Table 1.

COMPUTERS and COMPOSITION 2(4), August 1985, page 47

Table 1 Sentence Totals ------------------------ Language Total Number of Sentences ------------------------ Chinese 1061 Farsi 55 French 134 Greek 60 Hindi 44 Indonesian 18 Japanese 14 Korean 279 Malay 25 Polish 21 Russian 315 Spanish 395 Tagalog 81 Vietnamese 85 Yoruba 11 ------------------------
The number of sentences we selected for our study reflect the percentages of students speaking individual languages enrolled in ESL classes. The largest percentage of students is Chinese, followed by Spanish, Russian, Korean and Greek. Since the number of sentences for languages other than these top five is rather small, we decided to use only these five languages as the bases for analyzing and subcategorizing the data. We also tailored our courseware more closely to these languages.

COMPUTERS and COMPOSITION 2(4), August 1985, page 48

The database files we compiled for this study contained a field for first language, one for error type or types, and another for sentences. The standard kinds of searches, counts and reports could be made, but because of the shortcomings of the system--difficulty of access to terminals, frequent and unexplained crashes--we supplemented this work with a microcomputer database. The biggest impediment in this kind of analysis was not to have access to a computer-driven parser like WRITER'S WORKBENCH for EPISTLE.1 We had to analyze the error type and enter the errors, first language, and text (without making the correction unconsciously) by ourselves on a microcomputer. And, although this process was slower, we found humans could handle the data better--especially in the semantic domain. Table 2 shows examples of the microcomputer file structure we used.

Table 2 Example of Sentences
------------------------------------------------------------------------- Language: Chinese Error: voc/idiom Sentence: We consist the New Year as an important day. Language: Japanese Error: prep Sentence: Regions really answer to these questions.

Language: Greek Error: adj-n Sentence: Some of their children become violences.

COMPUTERS and COMPOSITION 2(4), August 1985, page 49

Language:  Polish					Error:  article

Sentence:  In order to decrease death rate, the speed limit was imposed.


Currently, we have begun to analyze data on such errors as prepositions, subject-verb agreement, part-of-speech confusion, articles, number, verb tenses, etc. (Dalgish, 1985). For all languages, sentences containing vocabulary or idiom mistakes and preposition mistakes due to idiomaticity outnumbered all other grammatical error categories. We are continuing to look at the data gathered in our files and analyze other error types, but we have gone on to develop some courseware based on our preliminary findings. What follows are some analyses of the preposition and verb-agreement data we complied which indicated to us the directions to take in designing our software.


In terms of percentages, prepositional errors remained fairly constant cross-linguistically no matter how we broke up or dissected the data. The number and percentage of preposition errors are given in Table 3. Table 4 lists some representative preposition errors.

COMPUTERS and COMPOSITION 2(4), August 1985, page 50

Table 3 Prepositions ------------------------------------------------------------------------- Number of Percentage of Language Preposition Preposition Errors Errors ------------------------------------------------------------------------- Chinese 201 (19%) Greek 47 (18%) Korean 64 (23%) Russian 41 (13%) Spanish 69 (17%) ------------------------------------------------------------------------- Table 4 Preposition Error Sentences ----------------------------------------------------------------------------- 1. Chinese: Since this society is concerned time very much, . . . 2. Korean: Upon my experience, I know this is good. 3. Greek: They never complain for their occupation. 4. Russian: I can see you, regardless where you live. 5. Spanish: In here the trains and buses are cheaper. -----------------------------------------------------------------------------
When we scrutinized the preposition errors according to the criteria of omission vs. con-

COMPUTERS and COMPOSITION 2(4), August 1985, page 51

fusion, or by percentages of idiomatic prepositional errors, we found similar patterns. Errors in which prepositions were incorrectly omitted (see Table 4, #1 for an example) ranged from 17% for Greek to 37% for Korean (Table 5). Errors in which prepositions were confused (See Table 4, #2) ranged from 83% to 63%, and were roughly similar cross linguistically, with speakers of Greek as a first language tending to confuse prepositions more often than native speakers of other languages (Table 5).

Table 5 

Omission vs. Confusion of Preposition Errors Totals %


Language	% of Omitted-Prep	% of Confused-Prep	

		    Error Types		    Error Types		


Chinese			30%			70%		

Greek			17%			83%		

Korean			37%			63%		

Russian			27%			73%		

Spanish			29%			71%		


When we considered all the preposition errors involving idiomatic usage (errors of commission, as in Table 4, #3; or of omission, as in Table 4, #4), the percentages were also close cross-linguistically, from 47% to 54 % (Table 6).

COMPUTERS and COMPOSITION 2(4), August 1985, page 52

Table 6 

Idiomatic Type Preposition Errors Totals %


Language			% of Preposition Mistakes

				   Due to Idiom Errors	 


Chinese					54%		 

Greek					49%		 

Korean					47%		 

Russian					49%		 

Spanish					52%		 


The bottom line, then for prepositional errors was that there was a heavy concentration of errors due to idiomaticity, mostly in sequences involving VERB, ADJECTIVE, or VERBAL ADJECTIVE + PREPOSITION. This indicated that software development for our ESL students should focus on those sequences, emphasizing work on sentences involving verbs or adjectives and related prepositions. We decided that the software should also test students on verbs that typically do not take prepositions, since errors of commission and confusion in idiomaticity were so high.

We did not feel it necessary to incorporate the only language-specific finding--that native speakers of Greek omitted fewer prepositions and confused more prepositions than the native speakers of other languages--into the ESL courseware we created because all the other errors were non-specific with regard to language. Further research, at the individual lexical

COMPUTERS and COMPOSITION 2(4), August 1985, page 53

level, may show that certain verbs or adjectives are more confused for speakers of certain languages, but at the moment the data does not indicate that this is so to a significant degree. If this turns out to be the case, we will incorporate these findings into future courseware for ESL students. For now, our preposition software concentrates on VERB and ADJECTIVE + PREPOSITION idioms without regard to first language of the speaker.


For subject-verb agreement, there was again some similarity cross-linguistically for percentages as Table 7 indicates.

Table 7

Verb Agreement Totals


Language	% of TOTAL Number	

		   of Sentences     	


Chinese			11%		

Greek			 7%		

Korean			 7%		

Russian			 6%		

Spanish			14%		


Speakers of languages with a verbal agreement system did not seem to find this syntactic structure any easier than speakers without such a system; note the similarity between the Spanish

COMPUTERS and COMPOSITION 2(4), August 1985, page 54

and Chinese totals. And for all languages, percentages of verb-agreement error totals for the improper use of the base form instead of the -s form were similar. All groups used the base form more than they should have; the range is displayed in Table 8.

Table 8

Agreement:  Improper Use of Base Form


Language	% of Agreement Sentences


Chinese			77%		

Greek			82%		

Korean			70%		

Russian			83%		

Spanish			77%		


There were, however, cross-linguistic differences of certain verb-related errors. For instance, speakers of Greek, Korean and Spanish had problems with prepositional phrases interfering with determination of the subject and, hence, verb-form choice; Chinese and Russian Speakers had more troubles when adverbs appeared between subject and verb (Table 9).

COMPUTERS and COMPOSITION 2(4), August 1985, page 55


Table 9

Agreement:  Comparison of Adverbs vs. Other

Modifier Phrases Between Subject and Verb


 			 % of Subject-Verb	

Language		 Agreement Errors	

		      Adverbs	     Modifiers	


Chinese			18%		11%	

Greek			11%		24%	

Korean			20%		35%	

Russian			11%		 5%	

Spanish			 8%		30%	


Agreement mistakes also occurred with existential there as subject; this error was prominent in our Chinese and Spanish speakers and less so for speakers of other languages. Finally, determiners and pronouns as subjects featured very highly in verb-agreement errors for Greek Speakers, less highly (but still significantly) for speakers of other languages.

I do not have time to address the theoretical issues of how these findings relate to the current controversy over the dominance of developmental vs. first-language interference as the source or errors for second language learners, here.2 In a practical sense, however, these findings can tell us much about the directions our instruction should take and how we might design more appropriate courseware.

COMPUTERS and COMPOSITION 2(4), August 1985, page 56


One immediate use for the data we had gathered was simply to call up by computer all examples of certain errors and print them out, grouped by first language, for students to use as correctional or proofreading exercises. Both students and faculty found the material to be enormously useful. Faculty explained that the sentences in these exercises were less static than those found in textbooks, more authentic (after all, they were straight from the students themselves), and more relevant because the students recognized idiomatic phrases, terminology, and vocabulary that were featured in the topics on the essays. Students were the first to note that "the errors could have been our own"; they seemed to recognize that certain patterns of errors made by their linguistic counterparts could have been ones they themselves would have made or did, in fact, make. This use of the data was immediately and successfully implemented.

Some time after the research had begun, the college was able to support the purchase of equipment for an ESL microcomputer lab. At this point, we recognized that it was time to take the results of the research and apply it to grammar software we could write ourselves.


This is not the place to expound a philosophy of what constitutes good ESL courseware.3 Instead, I will focus on some points that have not received much attention, either in courseware reviews or the literature about computer-assisted instruction in ESL.

COMPUTERS and COMPOSITION 2(4), August 1985, page 57

The software we wrote attempts to strike a balance between the marvelous potential and the amazing shortcomings of computers--and of software--in dealing with human language.

For our purposes, computerized parsers were out; cost, memory, and inability to deal with our students' mistakes were the prime reasons for that.4 Canned software was also rejected because in most cases it was nothing more than a textbook on a computer screen, with the same sentences or phrases repetitively given to all students along with the minimal animation and sound effects deemed necessary by publishers to sell the product. All of this is inappropriate in a university setting.

No software can call for completely open-ended responses by the students, since the computer is so limited in what it can decipher. In our own courseware, we restricted input, limiting ourselves to word, phrase and sentence levels.


To remedy the problems posed by canned software, the software we wrote relies on a random sentence generator.5 As illustrated in Table 10, this sentence generator works by interrupting a loop with a GET statement, ending a run through the random generator.

COMPUTERS and COMPOSITION 2(4), August 1985, page 58

Table 10

Listing of Randomization Subroutine




45010 	IF PEEK ( -16384) <128 THEN 					

	XX=RND(l);YY-RND(l):ZZ=RND(l):AA=RND(l):GO TO 45010 		


45020	LET X=INT ((. . . . .*XX)+1)					


The random numbers generated by this process are then used to select among the various syntactic patterns appropriate to the lesson, thereby varying the linguistic structure of sentences among such parameters as transitives, intransitives, modals or operators, existential there, various kinds of adverbs, modifier phrases, final phrases, etc. The random numbers are used to select various nouns, verbs, phrases, prepositions, etc. which have been loaded from text files into arrays; the preposition program loads some 2000 items. All of this forces students using the program to focus more intently on the structure (and, we hope, on the principle behind the structure being tested) and less on any one individual, idiosyncratic example.

In each program written for sentence-level work, students must then manipulate the randomly-generated sentences in specified ways. The agreement program, for instance, asks

COMPUTERS and COMPOSITION 2(4), August 1985, page 59

students to determine if the subject and verb agree, then asks students to identify both the subject and the verb and to modify and incorrect sentence by changing either the subject or the verb. At the end of this sequence, students produce or are given the corrected sentence. The prepositional program is similar, requiring students to provide the correct preposition if necessary, define the differences between certain prepositional pairs (talk to vs. talk about, for example), and insert the verb and preposition as a unit in the original context of the sentence.

Students are frequently given choices in the direction of the program. They are, for example, asked if they wish to loop back, create part of a sentence, or test themselves on a particular verb or adjective.

The lexicon of the program contains the vocabulary and idiom of our students, particularly the error areas they exhibited according to our research. In our preposition program, for example, we include those verbal and adjectival idioms that our research showed occurring most frequently in the students' writing. Memory limitations prevent our programs from containing every noun, verb, or prepositional idiom, but we have included common vocabulary and idioms taken from basic or beginners' dictionaries.

In addition, the lessons we have designed branch according to the first language of our students: the patterns of errors found most prevalent for each particular linguistic group are emphasized in different branches. This was accomplished by weighting various random numbers depending on the users' first language. For example, because we saw that Greek, Korean, and

COMPUTERS and COMPOSITION 2(4), August 1985, page 60

Spanish speakers made more errors when prepositional phrases were interposed between subject and verb, these students are given more sentences with this structure, in contrast to the Chinese and Russian speakers who get more sentences with adverbs appearing between subject and verb. Existential there as subject is featured more prominently when users speak Chinese or Spanish, and less so for users speaking other languages.


It is much too soon to determine if the CAI software we have developed is successful; we have had it in place for only half a semester. But early indications are promising; students have worked for long periods on these exercises, some spending hours on the same program. Additional programs for infinitive vs. gerund choice, articles, and plurals are being written as supplements to classwork. Students may take home their work and study their problem areas at their leisure, and their teacher can also examine their progress.6

Currently, we feel we have provided a solid basis for improving our students' writing by grounding our lessons firmly on the results of error analysis in our research. We expect that more exposure to these problem areas will help students and teachers focus more closely on errors.

In the future, we plan to build more language-specific features into our program if our research shows this to be viable. We may develop, for example, lists of individual lexemes

COMPUTERS and COMPOSITION 2(4), August 1985, page 61

that are more difficult for speakers of language X than language Y and let students branch to these more often. And beyond the lexical level, we may learn that other syntactic structures might be tested more often for speakers of language X than language Y. This application of branching in ESL CAI has largely been overlooked but could be extremely useful and helpful to the ESL student.


1 See Heidorn (1982) for a description of the EPISTLE program, which can handle some types of subject-verb agreement errors.

2 See Krashen (1977).

3 See Dalgish (1984, Higgins) among many others for discussions of what constitutes good ESL CAI.

4 Professor Martin Chodorov of Hunter College has worked on the EPISTLE project and in personal communication reports on the difficulty of dealing with ESL-type errors. A humorous and entertaining account of the problems with human language for a sophisticated program called RACTER is given in Dewdney (1985) and Chamberlain (1984).

5 Personal communication, Professor Geoffrey Akst, Dept. of Mathematics, Borough of Manhattan Community College.

COMPUTERS and COMPOSITION 2(4), August 1985, page 62

6 We also use word processing for essay writing, in and out of class periods.

GERARD M. DALGISH teaches at Baruch College, CUNY.


Burt, Marina, Heidi Dulay and Mary Finocchiaro (eds)., (1977) Viewpoints on English as a second language. New York, Regents.

Chamberlain, William. (1984) The policeman's beard is half-constructed. New York, Warner Books, Inc.

Dalgish, Gerard M. ( 1985) Computer-assisted ESL research. Calico Journal 2(2), 32-33.

Dalgish, Gerard M. ( 1984) Microcomputers and teaching english as a second language: Issues and some CUNY applications. Research Monograph Series Report No. 7, Instructional Resources Center, The City University of New York.

Dewdney, A. K. (1985 January) Computer recreations. Scientific American 252(1), 14-20.

Heidorn, G. E. (1982) et al. The EPISTLE text-critiquing system. IBM Systems Journal 21(3), 305-321.

Krashen, Stephen D. The monitor model for adult second language performance. In Burt, Dulay and Finocchiaro (eds).