19.07.1999   ©1999 - Patrick Boylan – patrick boylan.it

www.boylan.it – Return to Publications Page or Home Page

In: S. Lajoie & M. Vivet (Eds.), Artificial Intelligence in Education, IOS Press, Amsterdam, 1999, pp. 305-312. [ISBN:9051994524]

(To read a related unpublished article,   Business Letter writing with CALL,    press here: )



Metacognition in Epistolary Rhetoric:

A Case-Based System for Writing Effective Business Letters
in a Foreign Language


Patrick Boylan, Carla Vergaro
Dipartimento di Linguistica, Università di Roma Tre
Via del Castro Pretorio, 20 I-00185 Roma, Italia

Alessandro Micarelli, Filippo Sciarrone
Dipartimento di Informatica e Automazione, Università di Roma Tre
Via della Vasca Navale, 79 I-00146 Roma, Italia


1. Introduction

The Business Letter Tutor discussed here and described more fully in [10] helps office staff learn to compose effective business letters in English by getting them to define their goals and then retrieving and displaying appropriate excerpts from a database of letters of proven value appropriately tagged paragraph by paragraph. In cutting and pasting the excerpts together, users learn through example what effective letter writing means. Instruction is "self-directed" in that it is up to the users to judge how suitable the retrieved excerpts are and how they should be pieced together. The assumption is that, in a business environment, users will want to produce (and have a way of recognising) satisfactory end products. Their final choices may therefore be considered "expertise" which the Tutor can use to furnish ever more useful excerpts to study in similar circumstances in the future.

2. The Educational Philosophy behind the Project

The project is based on the dictum according to which teaching less favours learning more [12]. Conventional wisdom has it that training is only as effective as the trainer is competent -- thus the attempt to create tutoring systems able to formalise and control the entire learning process, systems which can then (it is hoped) be fine-tuned to perfection [3]. Implicit in this Faustian attitude is its Pygmalion counterpart: learners are seen as empty vases to be filled with "knowledge" by a deft trainer/system, i.e. one capable of pouring information into their heads without spilling a drop. Or, to use an image more consonant with computer science, learners are seen as inert silicon chips to be programmed by a skilled human/machine tutor so that, for a given input, a specified output is regularly obtained (learners are said to "know" a "subject" when they regularly output expected answers to test questions). This philosophy has such ancient roots that it has withstood decades of research clearly demonstrating the fundamental creativity -- and thus uncontrollability -- of the learning [13]. Cognitive scientists have in fact shown that students are neither empty vases nor inert chips: they are, indeed, the engineers of their own learning process [2]. Teachers or trainers, like books or audio-visual aids or computer programs, are simply one of the tools offered to students by the educational environment; they can help or hinder learning but cannot cause it to happen (nor prevent it from happening). The first premise of this paper, therefore, is that learners are experimenters who want (or who can be led to want) to investigate a domain and who, in forging tools for this purpose, end up (re-)creating a "subject". This view of learning radically changes the role of the tutor (human or electronic) who, from "depository of knowledge", becomes the agent responsible for creating a stimulating environment in which the learner can come to grips with a given domain by conducting successful experiments on it.

3. The First Implementation and the Educational Benefits

The Tutoring System presented here reflects this philosophy. It aims at harnessing the power of an AI-based engine to give users control over the System (specifically, over a data base of business letters) instead of giving the System control over them. The program first gets the user to define the circumstances of the letter to be written; then it searches a Case Library of past (successful) company correspondence for excerpts that match as closely as possible both the present circumstances and the user profile. Hits appear in a Model Letters Window. To produce an appropriate letter, users cut and paste excerpts into an Edit Window and then adjust the collage to fit the current situation. How are these processes handled by the System? In the original design (see Figure 1), letter excerpt retrieval is handled by three components:

1a) a User Model -- i.e., the set of attributes describing the person writing the letter or the person for whom the letter is being written (e.g., in the case of a secretary using the Tutor, the boss). In an office-pool situation, there would be many "bosses" and therefore the system would contain many User Models, each with an idiosyncratic way of handling such writing tasks as ADDRESSING_COMPLAINTS, REQUESTING_INFORMATION or DISPUTING_ FINDINGS. The User Model is built up initially by selecting a User_Stereotype [11] (from among a set number of Stereotypes programmed into the system) on the basis of the answers which a new user gives to a brief questionnaire. In other words, the User defines himself (or is defined by a secretary) in terms of ATTITUDES and EXPRESSIVE STYLE and the system picks the closest match from among the already existing Stereotypes. A Stereotype is a set of attributes/values, i.e. weighted ATTITUDES and EXPRESSIVE STYLES typically associated with letter-writing MOVES and STRATEGIES, i.e. with an inventory of the thematic development devices (or "rhetoric") contained in the various letters stored in the Business Letter-Writing Component.

1b) a Recipient Model -- i.e., the person to whom the letter is addressed. As in the case of the creation of the User Model, the recipient is initially defined by means of a short questionnaire; the system then enhances the definition by associating it with one of several Recipient_Stereotypes programmed into the system.








Figure 1. The Architecture of the System.

Both the Recipient and the User Models are then refined by the system as time goes by. That is to say, the attributes and values associated with successfully retrieved letters, in relation to domain knowledge and goals, are incorporated into the Stereotype to form an original Recipient Model (or User Model as the case may be). In other words, the attributes and values are not stored as such but are linked to data in the other system components to form an associative network.

2) a Domain Model, i.e. the savvy that a writer of good business letters has and which can be formalised as the set of possible links among the MOVES and STRATEGIES, associated with particular goals, characterising the business letters in the data base.

In line with the educational philosophy mentioned above, the System does not attempt to tell users whether their selections are "right" or "wrong". The burden of judging the suitability of excerpts and collages is left entirely to the users themselves. In a business environment, it can be assumed that they will want to produce -- and have some means of recognising -- satisfactory end products; their choices may therefore be used as "expertise" from which the System can learn to furnish increasingly appropriate excerpts to cut and paste, customised according to each user's history of choices.

"But isn't such a system simply a data base?" one might object at this point; "After all, it doesn't teach". The reply is simple: True, our System doesn't "teach" in the traditional sense. But then, systems don't necessarily have to, in order get users to learn. Indeed, self-directed learning presupposes non-directive teaching. Research shows that educational practices based on this philosophy can be extremely effective. "Learning by example" -- provided the examples are self-explanatory -- has shown its value in acquiring procedural knowledge (like writing skills, for instance [1]) while "learning-by-doing" -- provided the tools truly permit users to get a hold on their object of study -- has proven to be highly effective in getting learners to internalise procedural knowledge and make it automatic (id.). Even a seemingly mechanical activity such as "recopying" a model letter retrieved by the System -- provided the recopying is done with intent, i.e. to achieve a specific communicative aim -- is an extremely effective way for non-native speakers of English to acquire (and to get practice in using) the "hard-to-teach" lexical-grammatical and pragmatic subtleties that characterise well-written English and that cognitive methods tend to sweep under the carpet (id.). The Tutoring System we built fosters all three of these activities.

Of course, since non-directive "teaching" is by definition unobtrusive, most lay users may have the impression that our computer program is simply a "word-processing aid", not an educational product. But, in point of fact, such an impression is simply the proof that the learning going on is genuinely "contextual" (i.e., that these users perceive a connection between the texts to be produced and some real-life function). It also proves that user motivation is "intrinsic", i.e., that these users feel they are accomplishing something and not simply "doing exercises".

4. Lessons Learned

As mentioned, two problems were encountered in developing the prototype of the system. They were:

(a.) the sophisticated meta-language used to tag the paragraphs of the letters stored in the database cannot be utilised to furnish the user with categories of letter writing styles and strategies from which to choose; this is because users simply balk at having to choose from a myriad of menu selections. Another "meta-meta-language" must be created, couched in everyday language and based not on conceptual categories but on pragmatic goals. That language will be discussed in Section 5 of this paper and illustrated in the Appendix;

(b.) the user will, with time, tend to get into a rut and choose the same rhetorical strategies and even the same letters, simply because they work, however minimally; thus, to guarantee the educational value of the Tutor and to get the user to explore new ways of writing effectively, the System must incorporate an additional component in the form of an Overseer with a "personality" of its own, capable of suggesting different approaches to a specific writing task.

The Overseer Model is a new component consisting of a set of goal-setting and goal-attaining heuristics which enable the system to improve its performance creatively. In the original specifications, system performance improved over time through the continual refinement of the USER, RECIPIENT and DOMAIN models on the basis of successful hits (i.e. on the basis of which bits of stored correspondence are actually chosen and used by the secretary day by day). The addition of an Overseer Model, now underway, will give the system an "independent personality" in overseeing the operations and suggesting improvements in writing style and effectiveness.

The Overseer Model intervenes with "helpful" suggestions as users go about composing their letters. The purpose of these suggestions is to get users to consider MOVES and STRATEGIES not necessarily consonant with their past choices but perhaps better suited to obtaining the desired effect. In fact, the MOVES and STRATEGIES suggested correspond to the Overseer’s developing "personality", not the User’s attitudes and style. In a word, it is though the user receives, in addition to the help furnished by the retrieval system, the advice of an outside agent (perceived something like an office colleague trying to lend a hand). The intrusions may at times appear meddlesome (since the "colleague's" way of doing things is not necessarily consonant with the user’s way of handling correspondence); at other times they may be truly illuminating. This is because suggestions made by an "outsider" may help the user get out of the rut her/his past choices have placed her/him in. The Overseer Model is still in the project stage and will not be discussed in this paper.

5. A Simplified Meta-Language

As discussed previously [8, 9], what makes our letter-writing tutor efficient in retrieving the "right" bits and pieces of letters is the use of User Model built dynamically by means of a hybrid architecture in which an artificial neural network is embedded into a case-based reasoner (each combination of USER_STEREOTYPE + RECIPIENT_STEREOTYPE + LETTER_ BITS_SELECTED_FROM_DOMAIN constitute a case). This solves the indexing problem created by the continual updating of system every time the user pastes together a letter and thereby "approves" a certain set of letter fragments and a certain order of presenting them, which is linked to certain goals.

Where our system proved most lacking, however, was in the user interface. Our original design failed to take sufficiently into account the negative effect on the menu displays caused by the "explosion" of attribute types associated with new letters added to the data base as time goes by.

To keep things simple, our previous work was based on a handful of letters. As soon as we extended the number beyond 50, however, it became impossible to incorporate new categories of MOVES and STRATEGIES (as well as CONTENT, LETTER TYPE, KEYWORDS) into easily legible screen menus. Internally, the system had no problem in indexing the new attribute types. But the overburdened menus made the system decidedly unfriendly for an unsophisticated user and therefore, with run of the mill office personnel, practically unusable. It became increasingly clear that only a very highly motivated and extremely intelligent student (such as we had employed for testing) could possibly handle the conceptual challenge of deciding between such subtle choices in writing a letter as, for instance, ESTABLISHING CREDENTIALS in order to DEFEND THE FIRM'S IMAGE, instead of DESCRIBING CAPABILITIES in order to REASSURE THE ADDRESSEE [5, 6]. In most any office, harried personnel would balk at having to choose from menus filled with such intricate and apparently arcane options [7].

What we needed, then, was a way to simplify the secretary's task by automating the choices she had to make and yet, at the same time, maintain the complexity of the tagging system used to characterise letter fragments, in order to enable the Tutor to deliver a small number of extremely appropriate choices. Our solution has been to display only summary menus, get the user to choose from them, and then let the program guess what specific choices the secretary would have made had the full set of options been presented. In other words, the same mechanism used to choose the letters -- i.e., the combined heuristics provided by the USER MODEL, the RECIPIENT MODEL, the DOMAIN MODEL in specifying the letter selections to be displayed -- is now being implemented to permit the system to choose the from a complex list of attributes on the basis of a few keywords. The simple keywords are in effect translated into a complex series of attributes on the basis of past performance. In other words, the non sequitur is avoided by having the user approve (thus, "choose") the options selected automatically by the system before they then become the basis of selection of the bits and pieces of correspondence: the System learns to guess what THIS particular user means by OBTAIN_ACTION on the basis of what OBTAIN_ACTION has usually meant for the user in the past.

In practice, the system offers the user a short-list of letter goals. Then, the system relates the goals chosen by the user to the characteristics of the parties involved in the correspondence (USER, RECIPIENT, DOMAIN EXPERT, OVERSEER), coming up with a second short-list of very specific, highly articulated goals. The user may approve the list without even reading it, choose from among the tags, or reject the entire list and call for another. In that case, a second short-list is displayed. If the AI retrieval mechanism is efficient, no more than one or two further attempts should be necessary.

Thus, the "overburdened menus" problem has been eliminated by making the choice of letter description options a problem of intelligent data retrieval, just as the choice of letters was. This is not a particularly new idea: intelligent Help Messages in a Word Processor Application operate on a similar principle (albeit without a User Model). What makes the present implementation interesting is the entity of the matching process: the list of letter tags can run into one or two hundred items, next to the dozen or so items that an intelligent Help Message algorithm must handle in a given context.

It is clear however that, for this solution to work, the system requires a carefully worked out architecture of METAKNOWLEDGE. It must, in other words, map its internal resources and translate them into even more general categories (a "meta-meta-language") which capture given sets of pragmatic relationships (USER - RECIPIENT - DOMAIN EXPERTISE - CONTROL FUNCTION) and which, at the same time, can be expressed in everyday language any office staff member can understand. The system's META-META-KNOWLEDGE may therefore be described as a simplified network within the network represented physically by the links between GOALS (contained in the Domain Model) and BEHAVIOUR (as expressed by the USER and RECIPIENT MODELS). It is represented communicatively in "everyday language" based on "actions to get things done". Examples are given in the appendix.

As in knowledge-based systems in general this metacognitive apparatus does not have to be entirely spelled out. The associations established between the various program components and the tagged letters stored in the data base, are due to weightings assigned on the basis of past user choices. In a word, the system builds up its expertise through the sedimentation of historical events (acts of will). This is indeed, it may be argued, what the intelligence of human beings comes down to. A general METACOGNITIVE framework is still necessary, of course, in order to provide the system with a basis for its initial operations. In the case at hand, that framework defines specific epistolary communicative acts and relates them to a network of GOALS and BEHAVIOUR typical of the business world [4]. This enables the system to second-guess the user's intents, thus saving her/him the tedium of making her/his will fully explicit through introspection. The METAKNOWLEDGE thus represented is linked to the META-META-TERMINOLOGY displayed in the menus dynamically, on the basis of who is writing to whom, for what purpose and using what kind of strategy. In other words, using the metacognitive framework, the system is able to choose the most likely candidates from the host of MOVES and STRATEGIES, related to specific ATTITUDES and EXPRESSIVE STYLES aimed at obtaining specific GOALS and producing specific BEHAVIOUR, and then present them to the user with a simplified terminology that corresponds, if the neural network has been sufficiently trained; to what the user desires.

6. Future Developments

We are currently examining the prospect of making our letter-writing tutor a more sophisticated tool by enabling it to learn from other machines. "Borrowing" tagged letters from other data bases can help reduce the considerable costs of updating the system by tagging new, additional letters by hand and introducing them manually into the system, to increase its scope. It would be much cheaper for a firm to take advantage of the work already done by tagging experts working for firms in the same domain. (The correspondence exchanged among firms would, of course, have to manually purged of specific references to clients, etc.) It is clear, however, that no system can be updated simply by copying the data base of tagged letters stored in another system. This is because the tags (MOVES, STRATEGIES) applied to the paragraphs of letters in other data bases must be "translated" into ATTITUDES and EXPRESSIVE STYLE corresponding to specific GOALS and BEHAVIOUR, stored in the various components of the target system. It would, however, be possible to interconnect computers running the program and let the Overseer of one unit take turns in interrogating the other units, as would a user. The bits and pieces of letters furnished by the other computers at the request of the Overseer, instead of being cut and pasted into a letter to send, would be stored as additional correspondence in the data-base. The unit doing the interrogating would establish, within its own system, the associations and weightings relative to the letter called up on the host computers' screens; the host computers would furnish the letters and the tags associated with them. All this could be done at night, while office staff are not using the machines, and would eliminate the need for high-priced consultants to update and enlarge a given system.

[1] Boylan, P. (1995). "What does it mean to 'learn a language' in today's world; what role can present-day computer technology play?". In Proceedings of the Symposium on Language and Technology, Florence: Editrice CUSL, pp. 92-114.
[2] Dalin, A. (1975). Towards self-management of learning processes? Strasbourg, Council of Europe: CCC/EES 75(9).
[3] Dickinson, L. (1978). "Autonomy, self-directed learning and individualization". In Self-directed learning and autonomy, Cambridge: University of Cambridge.
[4] Jenkins, S. and Hinds, J. (1987). "Business Letter Writing: English, French and Japanese". TESOL Quarterly, 21(2), pp. 327-349.
[5] Kong, K.C.C. (1998). "Are Simple Business Letters Really Simple? A Comparison of Chinese and English Business Request Letters". Text, 18(1), pp. 103-141.
[6] Maier, P. (1992). "Politeness Strategies in Business Letters by Native and Non-Native English Speakers". English for Specific Purposes, 11, pp. 189-205.
[7] Mauranen, P. (1993). Cultural Differences in Academic Rhetoric, Peter Lang Verlag, Frankfurt am Main.
[8] Micarelli, A., Sciarrone, F., Ambrosini, L. and Cirillo, V. (1998). "A Case-Based Approach to User Modeling". In: B. Smyth and P. Cunningham (eds.) Advances in Case-Based Reasoning, Lecture Notes in Artificial Intelligence, 1488, Springer-Verlag, pp. 310-321.
[9] Papagni, M., Cirillo, V. and Micarelli, A. (1997). " Ocram-CBR: A Shell for Case-Based Educational Systems". In: D.B. Leake and E. Plaza (eds.) Case-Based Reasoning - Research and Development, Lecture Notes in Artificial Intelligence, 1266, Springer-Verlag, pp. 104-113.
[10] Papagni, M., Cirillo, V., Micarelli, A. and Boylan, P. (1997). "Teaching through Case-Based Reasoning: An ITS Engine Applied to Business Communication". In Proceedings of the 8th World Conference on Artificial Intelligence in Education AI-ED 97, Kobe, Japan, pp. 111-118.
[11] Rich, E. (1983). "Users are individuals: individualizing user models". International Journal of Man-Machine Studies, 18, pp. 199-214.
[12] Schank, R. (1998). "Horses for Courses". Communications of the ACM, 41(7), pp. 23-25.
[13] Stevick, E.W. (1976): Memory, Meaning and Method. Rowley, Mass.: Newbury House Publishers.
[14] Toulmin, S. E. (1958): The Uses of Argument. Cambridge: Cambridge University Press.
[15] Wainright, G. (1993). Tricky Business Letters. Pitman Publishing, London.


Metacognitive System Applied to a Random Sample of Business Letters


given in the form of Value+[attribute]


dealership / 1, 4 specify problem / 12, 15,

[new] / 4, [insufficient overdraft] / 12,

payment / 2, 8, [errors in press report] / 15,

[overdue] / 2, solicit compliance/ 6, 8, 13, 14, 12, 15,

[impossibility to pay [unexpected emergency]] / 8, [retraction] /15,


request / 8, 11, 12, 13, warn / 6, 7, 13,

[extension for payment] / 8, [discretely] / 6,

[explanation] / 11, [order placement with another supplier]/ 7,

[increase in overdraft facility] / 12, [move elsewhere] / 13,

[rent review] / 13, support the request / 8, 12,


(level(s) of) service / 3,

account(s) / 2, 11,

aircraft / 5,

amount(s) / 11,

apolog(y)(ies) / 3,


Example of Meta-Language: Argumentative Discourse Flow (based on Toulmin 1958:104)

[D] You have not taken up your option yet,

[W] although not taking up option means relinquishing it

[B] as is stated in our signed Agreement; thus

[Q] it is reasonable to infer that

[C] you accept that we are free to act now and therefore

[R] while you may be just buying time,

[C’] we shall consider Agreement null.


(D = Datum, Q=Qualifier, C=Claim, C’=corollary to claim, W=Warrant, B=Backing, R=Rebuttal)


Transofrmation into Meta-Meta-Language: same Argumentative Discourse Flow

[D] No reply (option)

[C’] Invalidate (agreement)

[W] "As per" (agreement)

(Qualifier, Claim, Backing, Rebuttal reconstructed by system from previous similar correspondence)


Other examples of Meta-Meta-Language:

Apologise (failure to comply)
Buy time (decision)
Get clarification (shipping)
Complain (quality)


(letter taken from [15])

"A special promotion"

6 August 1994

Dear Mr Baker

[Title of special product or service]


If you could increase your efficiency and productivity and at the same time lower your costs, would you be interested?

It is because we think you would that we want you to know about [product or service]. *

* strategy: grab attention

At the present time, all businesses are looking for methods of reducing costs without harming profitability. [Product or service] helps you to do this.


[Briefly describe how product or service will benefit the customer]


In addition to all this, [product or service] possesses a unique additional benefit for your business. It is available for a limited period at a special discount price of [state price].


If this were not enough, we further guarantee that if, after trying [product or service] for seven days you are in any way dissatisfied with its performance, you can return it undamaged in its original packaging for a full, no-questions-asked refund.


Please return the enclosed post paid card today and you will receive

[product or service] by return.


Yours sincerely


related article (from Leeds Conference, 1999):


Business Letter writing with CALL

Patrick Boylan
Dipartimento di Linguistica, Università di Roma Tre
Via del Castro Pretorio, 20
I-00185 Roma, Italia

Alessandro Micarelli
Dipartimento di Informatica e Automazione, Università di Roma Tre
via della Vasca Navale, 79
I-00146 Roma, Italia

Carla Vergaro
Centro Linguistico di Ateneo, Università di Roma Tre
Viale Ostiense, 139
I-00154 Roma, Italia



1. Introduction

Post-communicative language teaching and learning is characterised by the heterogeneity of hypotheses and practices. Nonetheless, there is considerable consensus as to the major features of any successful language learning program. Catchwords such as "engaged", "meaningful", "task-based", "content-based", "process-oriented" and "self-directed" are inevitably used and may be considered extensions of the concepts behind the audio-oral or communicative approaches of the past: "learning by example", "learning by doing", "situational learning". What underlies these themes is the idea (not always put into practice, however) that language learning is a holistic and experiential process in which students construct their knowledge by interacting creatively with richly varied (yet somehow structured) linguistic input. In this perspective, the best learners are those who take responsibility for their own learning. In a recent paper Wolff (1994) suggests that post-communicative language learning and teaching falls under a constructivist paradigm. Constructivism may thus be said to represent, at the dawn of the 21st century, a new (or rather, a newly rediscovered) paradigm in language learning and teaching. In a word, human learners are not seen as passive absorbers of data: empty vases to be filled or inert silicon chips to be programmed. They are seen as explorers. Human knowledge is the explorer’s attempt to make sense of acquired experience.

If this is so, it follows that teaching aimed at developing our distinctively human capacities should not seek to fragment knowledge into elementary components and then systematically present those fragments so that only minimal cognitive effort is required to store them in memory (the hallmark of atomic or behavioural teaching methodologies). Instead it should provide students with the means and tools for actively constructing their own interpretation of a particular problem. This means offering them a richly varied (yet somehow structured) environment in which responsibly to "learn to make sense" of things.

How can this view of knowledge acquisition be implemented in a Computer-assisted Language Learning (CALL) system? Unfortunately, there are not many examples to go by. Much of the CALL software on the market today is based on the atomic and/or behavioural learning paradigms and is therefore of modest educational value (Garrett 1991). As Jonassen, Mayes and Mc Aleese (1997) write:

Taking a cue from such calls for "open learning systems", the authors of this paper have designed a CBT programme to teach Applied Written Business Communication using a holistic, self-directed, constructivist learning paradigm. The System makes use of a case-based reasoner integrated with a neural network to handle indexing. Students learn to compose effective business letters in English by being prodded to define their goals (for which they are the sole judges) and then provided with the materials needed to attain those goals. In practice, students are furnished with appropriate excerpts from a database of letters of proven value appropriately tagged paragraph by paragraph. In cutting and pasting the excerpts together, students learn through example what effective letter writing means. If the System manages to furnish truly appropriate cases that satisfy the students’ need to know in the particular communicative situation they are in, learning will take place without instruction.

As mentioned, the System does not propose the judge the quality of the students’ efforts. It is up to the students to judge how suitable the retrieved excerpts are, in what manner they should be pieced together, and whether the final product is indeed effective writing. This assumes, of course, that students are in an educational environment which allows them to use the System for real-world interaction (e.g., composing a job application or a letter to get information really needed). In other words, for the constructivist learning paradigm to work, students must be intrinsically motivated to piece together the most effective letter possible and to verify the effectiveness of the final product (e.g., by seeking the opinion of the teacher, seen as an aid instead of a judge; by seeking the opinion of native speakers through Internet chat sessions; and, most importantly, by getting feedback from the recipient of the letter, useful for future occasions and in updating the System correctly). For example, if the System is used in offices as part of a Continuing Education programme, secretaries will surely be motivated (their job depends on producing satisfactory correspondence) and will probably have some way of evaluating the effectiveness of the letters they produce: the opinion of their bosses, the opinion of the native-speaker house translator, feedback from sales representatives in contact with the recipients of the letters, etc.

After giving an overview of the program, this paper will concentrate on the research carried out by the domain expert to label the sample business letters that form the database. The concepts of "move" and "strategy" will be proposed to define the rhetorical development underlying such intent and a metacognitive analysis of the tags will be proposed in dialoguing with the User.


2. The Program: an overview

The System is composed of three main components:

1a) a User Modelling Component which selects the set of attributes describing the person writing the letter or the person for whom the letter is being written (the "boss" in the case of a letter being written by a secretary). The User Model is built up initially by selecting a User Stereotype from among a set number of Stereotypes programmed into the system on the basis of the answers which a new user gives to a brief questionnaire. In other words, Users define themselves and the system picks the closest match from among the already existing Stereotypes. A Stereotype is a set of attributes/values, i.e. typical ATTITUDES and EXPRESSIVE STYLES associated with letter writing STRATEGIES and MOVES, i.e. with an inventory of the "thematic development tags" contained in the various letters stored in the database which the program consults.

1b) a Recipient Model which selects the attributes describing the person to whom the letter is being addressed. As in the case of the creation of the User Model, the Recipient is initially defined by means of a short questionnaire; the system then enhances the definition by associating it with one of several Recipient Stereotypes programmed into the system.

The Current User and Recipient Models are stored in a User/Recipient Database where they are refined by the System as time goes by;

2) a Domain Model and Database; the Database is the actual store of business letters with their tags and links. The linking process is called "indexing the letters";

3) a User Interface that manages Input/Output with the User (initial interviews, user requests, browsing or editing, letter retrieval).

One might object at this point that the System described here is simply a database and not a tutor or CBT program. And in fact, the System does not "teach" — at least in the traditional sense of the term. But the point of constructivism (often missed) is that it is not necessary to "teach" in order to cause learning to happen. A human or electronic tutor can be said truly to enable students to "learn by example" if it simply provides sufficiently appropriate self-explanatory examples in the precise moment they are needed; likewise, it can be said to enable students to "learn by doing" if it provides them, upon request, with tools that are both appropriate for them and for the job at hand. In such cases, the best teaching is the least teaching (Boylan 1995).


3. Selecting "cases"

The System Database contains, as just mentioned, sample business letters appropriately tagged. The System indexes these tags together with the links to the User and Recipient Stereotypes active when the letter was originally inserted into the Database and with the Stereotypes (or updated Models) active each time parts of the letter are selected. Thus, each single letter extract (usually, a paragraph), together with its tags and links, represents a "case" which the System matches with the needs of the current User.

Clearly, tagging (which an expert does initially by hand, paragraph after paragraph) must be sufficiently sophisticated to allow the retrieval of genuinely appropriate cases; moreover, it must take into account is the mental processes Users are likely to go through in composing a letter (Mantovani 1995). The tagging procedure we developed uses the following categories:


The above categories permit retrieval of specific letters, as in any traditional database. But the System also permits the User to write original letters by choosing from a palette of discourse strategies, each one representing a rhetorically effective:


In this case, what appears on the screen is a series of discourse steps, invisibly linked to paragraphs in letters of various kinds stored in the. By selecting and ordering the discourse steps deemed appropriate, the User gets the System to create a patch-work letter, i.e. a new letter made up of bits and pieces of many other letters. For each move chosen by the User — e.g., "State_Problem", "Threaten_Action" — the System furnishes a series of model paragraphs (extracted from letters in the Database) from which to choose. The System will, of course, have put on top of the pile the extracts best fitting the present case (i.e., the weighted combination of tags and User/Recipient attributes which correspond to a given rhetorical ploy). Thus, an extract which has zero correlation as to Content or Letter Type if searched using typical SQL procedure, may end up topmost on the pile because it correlates highly with the rhetorical ploy the User has decided to use. The rhetorical structure of a letter is therefore considered the key element for the student to observe when learning to write effectively; content and format can be easily modified to fit current circumstances.


4. Defining business communication

How does our paragraph-tagging procedure work? We start with the tenet that all communication is purposeful, i.e. it is the manifestation of an intent (Parisi & Castelfanchi 1976, Poggi & Magno-Caldognetto 1999). However ritual it may be, communication expresses a given individual’s way of making things happen through words. Moreover, it is psychologically and culturally determined, as Callow and Callow (1992, p.5) note:

Moreover, given the manipulative character of business dealings in general, it may be claimed that what is made to happen through words in effective business communication is obtaining some kind of compliance. In other words, the prevailing communicative function of any business letter may be said to be persuasion. Even a letter notifying a change of address is arguably an attempt to persuade Recipients to note down the new address: effective letters of notification will obtain this kind of compliance, non effective ones will not. See Ghadessy (1993), Maier (1992), Bargiela Chiappini and Harris (1996), Jenkins and Hinds (1987), Yeung (1997) and Kong (1998) for more on purposeful discourse and business discourse in particular. See also Halliday (1978) for a framework (field, tenor and mode) useful in discussing style.

To conclude, research into communication and business discourse would suggest that learning to write effective business letters means learning to implement general rhetorical strategies in order to obtain compliance by a specific reader in a specific culture.


5. Tagging moves

Text patterning (and thus letter tags) may be conveniently described using the notion of move (Coulthard 1991, Longacre 1992, Bhatia 1993, Mauranen 1993).

In formalising our sample database of letters, we have drawn mainly on the work of Mauranen (1993) with regards to both the notion of move and the criteria used in identifying moves in business discourse. Mauranen (1993, p.225) gives the following definition of rhetorical moves:

Mauranen then lists six criteria which we applied to our corpus to identify the moves characterising each business letter. In addition, we added the category of strategy to indicate overall goals, i.e. typical combinations of moves characterising the letter. Briefly put, moves may be detected through a combination of the following clues.

Typographic indicators

Although in business letters one-sentence paragraphs representing a single move are extremely common, there is no one-to-one relationship between a move and a sentence, and a move and a paragraph, as the following exceptions will show.


(10) I therefore wish to apply for the post of [title] with [name of company]
and hereby enclose a copy of my CV in support of my application.

Here a sentence (admittedly co-ordinate) consists of two moves whereas in the following excerpt a single move spans two sentences.

(13) You will be aware, of course, that many businesses have ceased to trade recently. One factor in the high rate of business failure has undoubtedly been the level of rents that businesses have been asked to play.


Introduction of a new referent can correspond to a new move. Referents indicating the main point of each sentence are underlined.


(8) Due to an unexpected emergency that has occurred, I will be unable to make this payment by then.
If you review my file, I believe you will find that my payments have generally been made on time.*
My relationship with your company is very important to me, as well as my desire to maintain a good credit rating.*


Thematisation of connectives, adverbs, fixed expressions, etc. is a good indicator of the functional relationships among the parts.


(14) [Briefly describe how product or service will benefit the customer]
In addition to all this, [product or service] possesses a unique additional benefit for your business. It is available for a limited period at a special discount price of [state price].

The thematised expression signals that something new and different is being added to what has already been communicated.

(20) Unless his/her attitude towards customers improves, it will cost you business, the first of which could be ours.
As you know, in the course of the year, we send a considerable amount of business your way.


A change of tense can signal a different move. In the example given, the shift from a narrative past tense to an imperative and to the final future underlines the transition from move to move.


(7) When the order was delivered, the following items were defective: [list items]
Please arrange for speedy replacement of the defective items.
Unless we receive them by [date] we shall be compelled to cancel the order and return items already delivered.

Text reflexivity

This term refers to the use of connectives as indicators of functional boundaries among text patterns.


(9) Thank you very much for your application for this post.
I very much regret, however, that on this occasion you have not been successful.

"However" signals that the information in the second sentence moderates that in the first ("Even if I thank you for your letter, I am discarding it")


The following letter (from Wainwright 1993) has been tagged using the preceding principles. Categories will be listed only if they vehicle changes.






We have received your letter of [date], notifying us that, as a result of the recent rent review, our monthly rent in future will be [amount].


STRATEGY: businesslike response [opposition]

MOVE: reference to previous communication


LETTER TYPE: request [rent review]

DATE/KEYWORDS: 28.11.96 rent(s), review, monthly



This seems to us to be excessive, bearing in mind the state of repair of the property, the facilities available and the current state of the rental market for commercial properties.


MOVE: state position (rejection) + give reason

KEYWORDS: excessive, repair, propert(y)(ies), market(s) [rental], commercial, rental


You will be aware, of course, that many businesses have ceased to trade recently. One factor in the high rate of business failures has undoubtedly been the level of rents that businesses have been asked to pay.


STRATEGY: veiled threat [terminate relationship]

MOVE: state consequences

KEYWORDS: failure(s), rent(s)


In the light of all the above factors and in the light of the adverse effect on our business at this time, we would ask you to further review the rent you wish to charge us with a view to setting it at a more reasonable level which we would have a better chance of being able to afford.


MOVES: state consequences (continuation of move in previous paragraph) + solicit compliance

KEYWORDS: review, charge with





6. Conclusion

The educational philosophy behind the CBT System described here may be described as constructivist. According to this view, technology is deemed useful only insofar as it contributes to creating environments that allow a learner to satisfy some "need to know" by exploring the resources offered, formulating hypotheses, experimenting them and then evaluating the pertinence and utility of the knowledge thus gained. In the System described here, creative experimentation is carried out on cases (i.e. business letters) selected in order to give the student the "right" information just when it is needed. The business letters (tagged to indicate discourse moves, strategies, content, type, etc.) are stored in a Domain Database linked to a User/Recipient Models Database: the correlation of specific tags with specific User and Recipient profiles constitute cases. The capacity of the System to select the most appropriate cases for the letter-writing assignment undertaken by the User clearly depends on the sophistication of the tagging scheme. That scheme was developed according to: (1.) usability criteria and (2.) a pragmatic and intercultural view of business discourse. To tag or label letters appropriately, an investigation of business letter discourse was carried out. It was discovered that, in spite of a common opinion, this genre is extremely articulated. Nonetheless, formalisation is possible using the notions of type, content, move and strategy. Texts thus labelled are easily retrievable by the System and offer Users, who wish to write an effective business letter in English, suitable cases to observe and learn from.



[1] Bargiela-Chiappini F. & Harris S. J. (1996), "Request and Status in Business Correspondence", in Journal of Pragmatics, 28, pp. 635-62.
[2] Bhatia V. K. (1993), Analysing Genre: Language Use in Professional Settings, Longman, London.
[3] Boylan P. (1995), "What does it Mean to Learn a Language in Today's World; what Role can Present-day Computer Technology Play?", in Proceedings of the Symposium on Language and Technology, Editrice CUSL, Firenze, pp. 92-114.
[4] Boylan P. & Micarelli A. (1999), "Metacognition in Epistolary Rhetoric: A Case-Based System for Writing Effective Business Letters in a Foreign Language", Paper delivered at AI-ED99, Le Mans, 7/99.
[5] Callow K. & Callow J. C. (1992), "Text as Purposive Communication: a Meaning-based Analysis", in c. mann s. a. thompson (ed.) (1992), Discourse Description: Diverse Linguistic Analyses of a Fund-raising Text, John Benjamins, Amsterdam, pp. 5-37.
[6] Clyne M. (1981), "Culture and Discourse Structure", in Journal of Pragmatics, 5, pp. 61-6.
[7] Coulthard M. (1991), An Introduction to Discourse Analysis, Longman, London.
[8] Garrett N. (1991), "Language Pedagogy and Effective Technology Use", in Applied Language Learning, 2, pp. 1-14.
[9] Ghadessy M. (1993), "On the Nature of Written Business Communication", in m. Ghadessy (1993), Register Analysis. Theory and Practice, Pinter, London and New York, pp. 149-64.
[10] Halliday M.A.K. (1978), Language as Social Semiotic, Edward Arnold, London.
[11] Hoecklin L. (1993), Managing Cultural Differences for Competitive Advantage, Special report No. P656, The Economist Intelligence Unit, London.
[12] Jenkins S. & Hinds J. (198), "Business Letter Writing: English, French and Japanese", in tesol quarterly, 21, 2, pp. 327-49.
[13] Jonassen D.,  Mayes T., Mcalesee R. (1997), A Manifesto for a Constructivist Approach in Higher Education, http://led.gcal.ac.uk/clti/papers/TMPaper11.html.
[14] Kong K. C. C. (1998), "Are Simple Business Letters Really Simple? A Comparison of Chinese and English Business Request Letters", in Text, 18(1), pp. 103-41.
[15] Longacre R. E. (1992), "The Discourse Strategy of an Appeals Letter", in C. Mann & S. A. Thompson (1992).
[16] Maier P. (1992), "Politeness Strategies in Business Letters by Native and Non-native English Speakers", in English for Specific Purposes, 11, pp. 189-205.
[17] Mantovani G. (1995), L'interazione uomo-computer, Il Mulino, Bologna.
[18] Mauranen A. (1993), Cultural Differences in Academic Rhetoric, Peter Lang Verlag, Frankfurt am Main.
[19] Papagni M., Cirillo V., Micarelli A., Boylan P. (1997), "Teaching through Case-based Reasoning: an ITS Engine Applied to Business Communication", in Proceedings of the 8th World Conference on Artificial Intelligence in Education AI-ED, Kobe, Japan, pp. 111-8.
[20] Parisi D., Castelfranchi C. (1979), Discourse as a Hierarchy of Goals, Working Papers, University of Urbino, Urbino.
[21] Poggi I., Magno-Caldognetto E. (1999), The Core of Multi-Modal Communication and the Goals of Political Discourse, Quaderni dell’ IFED (CNR), vol. XVIII, Edizioni Libreria Progetto, Padova.
[22] Wainwright G. (1993), Tricky Business Letters, Pitman, London.
[23] Wolff D. (1994), "Der Konstruktivismus: Ein Neues Paradigma in der Fremdsprachenunterricht?", in Die Neueren Sprachen, 93, 5, pp. 407-29.
[24] Yeung L. N. T. (1997), "Polite Requests in English and Chinese Business Correspondence in Hong Kong", in Journal of Pragmatics, 2, pp. 505-27.

Cliccare per andare in su / Click to go to top>