Serendipitous 1 acquisition of Web Knowledge by Agents in the context of Human Learning

Transcription

1 THAI ETIS, Varese, Italy; June page 1of e25 Serendipitous 1 acquisition of Web Knowledge by Agents in the context of Human Learning Stefano A. Cerri 1, Vincenzo Loia 2, Sergio Maffioletti 1, Pierluigi Fontanesi 3 and Alberto Bettinelli 3 1 Università di Milano, Dipartimento di Scienze dell Informazione; Via Comelico, 39; I Milano, Italy; cerri@dsi.unimi.it 2 Università di Salerno, Dipartimento di Matematica ed Informatica; Baronissi, Salerno, Italy; loia@dia.unisa.it 3 Università di Milano, Polo Didattico e di Ricerca di Crema; Via Bramante, 65; I Crema, Italy; <DIV ALIGN=right> fontanes@weblab.crema.unimi.it, abettinelli@crema.unimi.it Abstract. Web Information may currently be acquired by activating search engines such as Northern Light. However, our daily experience is not only that Web pages are often either redundant or missing but also that there is a mismatch between information needs and the Web s responses. If we wish to satisfy more complex needs, such as produce Educational material and manage Educational dialogues we need to extract part of the Information and transform it into new interactive documents. The transformation to Knowledge useful for our purposes may either be performed by hand or automatically. In the paper we describe the preliminary choices for our system that will generate local Web pages and their associated interactive processes useful for learning foreign language terminology. Documents and dialogues are designed exploiting domain dependent classification principles (expressed as Abstract Data Types) and Agents. These "autonomous" software modules will be implemented in our own Agent Languages and operate conversationally both as Information seekers and as 1 From : a definition of serendipitous in an Agent s glossary: Serendipitous matches Serendipitous matches between users consist of matches made without the user necessarily intending to look for a match. A serendipitous match may involve a user who has not even thought to look for someone with similar interests on some topic, and may be surprised to find that there is any other user out there with such an interest. See also Brian La Macchia s thesis ( [1] ) about Internet Fishes. See also the on line Webster dictionary at w.com/cgi bin/dictionary : Main Entry: ser en dip i ty; Pronunciation: di p& te; Function: noun; Etymology: from its possession by the heroes of the Persian fairy tale The Three Princes of Serendip; Date: 1754: the faculty or phenomenon of finding valuable or agreeable things not sought for.

2 THAI ETIS, Varese, Italy; June page 2of e25 Knowledge constructors. Our work is part of a recently started European INCO COPERNICUS project called LARFLAST. Table of Contents Table of Contents Introduction LeARning Foreign LAnguage Scientific Terminology Agents Summary of the following sections Information Agents: Serendipitous Knowledge Acquisition from the Web Information acquisition The use of JIT Web Information by "traditional" Computer Assisted Instructional (CAI 11 ) programs Knowledge construction Dynamic classification criteria Mastery of Terminology: a linguistic or a domain dependent, conceptual skill? Communication Agents: Dialogues stimulating learning The model: STROBE Scientific technical competencies The bottom up approach to competencies The tools The Agent Communication Language(s) : AL and Jaskemal The content language The architecture An example of a simple interaction Conclusions References...23

3 THAI ETIS, Varese, Italy; June page 3of e25 1. Introduction The availability of the Web today has definitely changed the Information scenario. As Information pervades most of our activities, we have to be aware of the new opportunities even if we have not yet been attracted by them or forced to accept them. In Education, we are even more convinced that [2] : "One fundamental reflection for anybody interested in Education is that the goal of Education is that learners learn, i.e. change state during / after a communicative process. The process does not per se need to be "educational". That term applies eventually after an evaluation of the new state reached by the learner as a result of communicating. Communication is the real issue for learning and therefore for Education; learning may occur as a side effect (as it was agreed in the workshop reported in [3]. Educational software, then, is nothing else as interactive software. Whether or not communication stimulates learning in the learner is not primarily a property of the software managing the communicative process but a relation between the process and its effects on the learner." Traditional Information Technologies for Learning may be classified according to the "initiatives" taken in the Human System dialogues. When the System, playing 2 the role of a teacher, takes most of the initiatives, we may speak of "Tutoring Systems", while when the Learner is allowed to "explore" freely the system s knowledge, we call the system "a Learning Environment". The spectrum is quite broad, as well as its applicability to any kind of technological support to learning. The " I " that precedes the various acronyms (e.g. in ITS, or ILE, or else, for the French community, EIAO etc.) was previously bound to "Intelligent" (in the Artificial Intelligence sense) while later denoted more the "Interactive" property of the learning processes. Interaction, on turn, was previously considered to occur locally (Interaction with Information or Knowledge locally available) while, in the 90ties, the term extended its scope to the whole WWW. There are pro s and con s with respect to a design of Technologies for Learning such that the initiative in dialogues is taken by the system or by the learner or, alternatively, by each of the two. The available literature documents experiences in different domains, with different target learners and various technological resources that offer guidelines, if not recipes, for the successful choice. However, one aspect of the design process is considered a starting point in most application efforts: the "Knowledge" is in the group of producers of the 2 The seminal work of Carbonell [4], considering the need for mixed initiative instructional dialogues, pioneered the area of AI in Education.

4 THAI ETIS, Varese, Italy; June page 4of e25 Technologies for Learning. The "owner" of the Knowledge is considered to be the Teacher or the Domain Expert that, in the process, may also become a programmer, a system integrator, an author or the like; or either may be accompanied by other professionals that will help him (her) to transfer the Knowledge on the domain into a form to be used by the System in future dialogues with the Learners. Since the 80ties, research on Authoring Systems has focussed on the issue of facilitating Domain Experts, or Teachers to transfer their Knowledge to the machine. The term "Authoring" inherited often the meaning traditionally associated to the production of documents. These documents were electronic versions of books, including or not multimedia material according to the available technologies and the needs of the domain. Even the name of the Authoring Systems recalled books (e.g. Toolbook). A few researchers, instead, were pointing out that the objective of Authoring was not just to produce electronic books (recall the "million dollar page turner" criticism to Computer Assisted Instructional material in the early years). On the contrary: Authoring was a design and production process such that the outcome (the System) is able to guide learners to learn through dialogues. The output of Authoring, therefore, was considered a "dialogue management system", quite different in nature from a simple book [5, 6]. One of the most successful scientific and technical efforts in advanced Authoring that we know of has produced the Authoring System GTE, reported extensively in [7]. The output of Authoring is a Dialogue Management System that operates on an explicit representation of Instructional Knowledge partially separated from the representation of the Instructional domain. Instructional or Pedagogical knowledge emerges from studies and the practice of Instructional Design. The assumption of approaches such as GTE is that Pedagogical Knowledge or expertise may be at least partially independent from the pedagogical domain. At the other extreme of the spectrum (Learning Environments) we have two types of systems for learning: simulators (including games, programming languages for learning by doing, such as LOGO?) and hypertexts hypermedia. Certainly the most impressive and successful system that includes these software tools is the Web. We may consider the Web as a kind of gigantic repository of Information 3. An archive that also includes ways to obtain Information from connected people (think to the use of mailgroups) and even a relevant set of well designed Tutoring Systems that may be downloaded or used interactively 4. When compared to other settings, Instructional Design and the use of tools such as GTE requires significant human resources 5. It is therefore natural to ask if, given a learning purpose, are we able to shortcut any "specific design" phase and send the 3 In [8] it is reported that in the year 2000 the Information available in the Web will be around 30 terabytes (30 x bytes), i.e. more than the US Library of Congress. 4 In we find an example of current free Web services for distance learning. Another one may be found at

5 THAI ETIS, Varese, Italy; June page 5of e25 learners to the Web with the suggestion to "go and look for their own material, support, opportunities" without a significant loss of the learning effect. It is clear that there are in between solutions, such as a teacher s (or expert s) guidance for the student to search in the Web what may be interesting for a specific learning purpose. The choice among these different alternatives and the principles guiding the decision process are the issues presented in this paper. In the European funded INCO COPERNICUS project we have recently started, called LARFLAST 6, we are expected to build cooperatively a series of integrated tools that will promote in East European Bulgarian, Russian and Rumanian learners the acquisition of foreign language (English) scientific technical terminology [10]. We present hereafter the reflections of one partner in the project, the University of Milan. In brief, we believe that in the application domain of LARFLAST, and within the limited resources available, we may take the Web as the dynamic repository of Information for most of the Knowledge needed to build the tools required. However, instead of handcrafting new software tools tailored for LARFLAST, we adopt an approach of semi automatic construction of Knowledge from Web Information consisting of two partially overlapping activities: "serendipitous" acquisition and construction of Knowledge from Web Information and management of dialogues for LARFLAST perspective learners. Both activities use Agent technologies. The methodological questions we address in LARFLAST are fundamentally two: 1. Assuming a constructive view of learning, i.e. that learners may eventually learn as a side effect of communication: is it possible and how to generate communicative situations for learners that require a limited effort in Instructional or Software design, yet with a significant learning effect? 2. Assuming that the Web is an archive of Information, while learners need Knowledge, i.e. the outcome of transformations of Web Information mapping their learning needs: how is it possible to transform available Web Information to Knowledge in the simplest, most flexible and dynamic way? 5 In [9]we see a recent example of use of meta CASE tools for Authors to design their own material. Even if the trend to consider Courseware Engineering a kind of software engineering is certainly important, nevertheless the approach requires a significant investment, in particular concerning the "standards" or "ontology" to define for a useful collaboration in the developments. 6 Prime contractor is the University of Leeds, CBLU; Scientific coordinator the Bulgarian Academy of Sciences and the University of Sofia (BU). Partners include: VIRTECH (a SME from Sofia), the Rumanian Academy of Sciences, the University of Simferopol (Crimea), the University of Manchester and the University of Milan.

6 THAI ETIS, Varese, Italy; June page 6of e LeARning Foreign LAnguage Scientific Terminology The project LARFLAST concerns the study, realization and experimentation of a set of tools for learning foreign language scientific terminology. "Terminology" denotes all terms in a (natural) language that are not to be found in a dictionary. Indeed, in traditional dictionaries we may find usually terms that are commonly used for long times, not terms that represent specific interests of very restricted groups for a short period of time. What occurs in several areas nowadays, is that the time delay between the "invention" of some term by technologists and the use, or desired use of the term by many people is much shorter as it was the case before. Technologies diffuse rapidly, be it welcome or not. And more and more people needs to master new terms (e.g. acronyms) in order to use the technologies 7. This is particularly the case for English technical terms, because scientific and technical communities have historically chosen English as a common language 8. 7 Examples taken from a real translation process (thanks to Chiara Sottocorona professional journalist on technologies and Lucia Sottocorona professional translator ) of an American semi technical book about the Web include the following American English sentences hard to understand and translate, even if the reader has a deep knowledge both of English and Technologies:? Today s Internet hustlers invade our communities with computers, not concrete?? The key ingredients of their Silicon snake oil is a technocratic belief?? devices that create antisocial dundes out of nice, clean kids?? "Jolts per minute" strategies game implemented in programming to help maintain the zapper s attention?? run control abilities of software supported meta design?? their "read only" nature?? critics are prompt to take big business to task for overhyping interactivity social critic more of the same?? this is done either directly, as favored routings of nerve impulse?? the user as an essential partner in sense making and meaning creation?? available off the shelf?? silicon snake oil?? he did allow as how he d read a lot of Artaud in college, so that who knows, you know??? arm wrestling experiment?? it sortof worked?? vivogenic cyberspace?? media holosthetic?? living through our teeth?? station identification devices?? robotic tinkering? 8 A counterexample may be seen in Italian, where the international community in the 19th Century chose Italian for expressing musical terms (such as coro, alto, soprano, allegro ma non troppo, andante con moto...) but the limited number of these terms (and the limited economic value associated to the mastery of them) do not really call for an effort, such as LARFLAST, to let more people understand and use those terms.

7 THAI ETIS, Varese, Italy; June page 7of e Agents The group that is currently active at the University of Milan will contribute mainly with research activities that concern the "Agent Technologies" that may be used for building, maintaining and using the tools. In particular we have adopted two complementary research approaches described briefly hereafter and temporarily called Information Agents and Communication Agents. Information agents are software programs that search for Information on the Web and transform that Information into a form that maps better with the learner s needs. Subsequent transformations of Web Information into domain and user dependent forms may be considered steps of Knowledge construction. Knowledge is not all or nothing, rather the outcome of transformations of Information into a form suitable for deciding or acting by humans or machines with goals, plans etc. in a specific context. We suggest the process to be "serendipitous" in the sense that not all the Knowledge we will construct will depend on intentions (e.g. pedagogical plans) explicitly expressed before starting the process. Communication agents are software programs that represent, activate, monitor, control and report about communication events (dialogues) in a distributed environment. Communication agents include Information Agents, Pedagogical Agents and other types of agents supporting generic communication (e.g. on line assistants for various kinds of tasks, such as design, monitoring...). The main reason for proposing Agents as a basic component of our work, is that Agents are software systems that offer, at the moment, properties of "autonomy" and "evolution" in their behavior that allow to think about distributed asynchronous collaborative problem solving in an unconstrained way. These properties are not similarly available in Actors or Objects [11], even if Agents may be considered "autonomous" Actors, and Actors may be considered concurrent Objects 9. As Agents are software components with an evolutionary behavior, mostly based on asynchronous communication, they seem to us suitable to represent evolving scenarios such as those that we envision in the project, where a more traditional top down approach to software (courseware) engineering seems less convenient, or perhaps unfeasible. 9 Notice that the three notions reflect the functional properties of the artifacts, not the implementation conventions. For instance, Agents may be implemented wit Object Oriented Languages, such as Java or Smalltalk or STklos or with Actor Languages, such as ABCL or Rosette or Actalk, but this does not mean that a Java program is an Agent unless it shows typical Agent s behaviors.

8 THAI ETIS, Varese, Italy; June page 8of e Summary of the following sections In section 2 (Information Agents) we will outline how we think to use Information Agents to construct and maintain the Knowledge about scientific terminology that will be later accessed, used and improved by learners and other users. In section 3 (Communication Agents) we will describe how we think to apply Agents to facilitate multi agent conversations among various LARFLAST users and improve progressively their effectiveness for learning. In section 4 (Tools) we will briefly mention the tools, a preliminary architecture and a trace of an extremely simple user system dialogue currently realized. 2. Information Agents: Serendipitous Knowledge Acquisition from the Web 2.1. Information acquisition It is clear to us that learning of the terms should be linked with the life cycle of the terms themselves. It does not make much sense to spend time and efforts to represent in any detailed way the meaning of a term, if we know that the term will become obsolete in a few months, or that another term perhaps a specialization or a generalization of the first one will become more important. Unfortunately, we don t know how persistent will be each term, neither if it will be more or less important in the future. Therefore, we think that a continuous monitoring of "the most relevant terms" should allow to concentrate on terms that are relevant at the moment our learner will need them and try to learn their meaning. This observation implies that we cannot compile a list of "meaningful terms" (according to some criterion) today and assume that the same list will have the property of being "meaningful" in the future. Probably, even the criterion of choice will have to be modified! We therefore think to associate a criterion of "meaningfulness" to the compilation of a preliminary list that in itself is not fixed. We will have a set of initial terms in a domain selected (e.g. electronic commerce or finance) and enrich that set according to a search of the most commonly used (technical scientific) terms in the area. The operation of searching for the most commonly used terms may occur asking widely used crawlers (such as Metacrawler, Altavista, Lycos...) to search in the net documents that include those "initial terms" and store them in one or more databases. One may then use traditional

9 THAI ETIS, Varese, Italy; June page 9of e25 techniques of computational linguistics (indexing, frequencies, concordances, etc.) to have a scenario of the usage (in the WWW) of a set of terms in an area chosen, and run the sequence of operations defined in order to update the scenario regularly. The linguistic corpus obtained may be partially stable, but not necessarily fixed. It will be interesting to see how stable it is, and how terms reach higher positions (become widely cited) with what speed. Software tools for accessing, viewing and using the corpus should become part of the results of the project. Yet, no formal representation, nor a classification, nor a specific learning training technology has been conceived or identified at this stage. We call this phase: Information acquisition. It is asynchronous in the sense that the Information Agents, once launched by their owners, will continue to collect terms, contexts, URL s, pictures, etc. from the Web and store them in databases for further processing until the owner stops their "spidering" activities through the Web. These Agents will report regularly, and eventually ask for further orders, requests, etc.; they will search mainly when the Web is less used (e.g. by night). One of the major values added by Information Agents is to mirror locally selections of Information available in principle in the net, but inaccessible in practice for most people for technical reasons (slow connectivity, limited experience in the use of Web tools etc.). A second advantage is to have "just in time (JIT)" Information 10. A third one consists in the opportunity to improve dynamically the ability of Information Agents to be selective by using even a minimal representation of Knowledge, as it was reported extensively in [1] and, with a different approach, in [12]. Whether improvements in the behavior of Information Agents require "reprogramming and recompiling" or may be performed dynamically (while agents run) is a crucial issue that we will discuss later in the paper The use of JIT Web Information by "traditional" Computer Assisted Instructional (CAI 11 ) programs If the terms (and their contexts, i.e. the sentences, or the Web pages that denote their meaning) are variable, the teaching programs using the corpus can, at most, work with presentations, pattern matching and similar traditional CAI techniques. As the Information is variable, these CAI programs are "generative 12 " and may in themselves become a useful initiative. Even if the "Instructional Design" of these educational dialogues may not be optimal, nor 10 Among the earliest projects adopting this approach: JITOL (Just In Time Open Learning) was an EU DELTA project running from 1992 to 1994 (see for instance ). 11 Acronyms such as CAI, CBT (Computer Based Training), CBL (Learning) and the like will be considered synonyms in this paper.

10 THAI ETIS, Varese, Italy; June page 10of e25 highly "intelligent", yet the continuously available new information collected by our Information Agents may feed traditional CAI software tools in such a way that the user may profit from the generated presentations, exercises and tests. These CAI modules would possibly be produced with a limited effort, yet have a significant effect on the learners, as these would be motivated to use them knowing that they have been generated "just in time" from the Web. In order to implement these tools, one has to define a few "CAI oriented" textual patterns and fill in these patterns from databases suitably fed, on turn, by the Information acquired from the Web. As the criteria for generating even the simplest database should emerge from the learner s needs, we do not commit ourselves at the moment to any choice. Rather, we indicate the simplest examples we may think of as hypothetical first steps for generative CAI in our project: a. Synonyms, antonyms b. Paraphrases c. Automatic multilingual translations 13 and wait for the suggestions of the partners active in the "applied linguistics" area to formulate clear requirements in this, we believe, simple, yet useful direction. Once requirements are defined, the specifications will follow and the corresponding DBMS generation. This phase of the project will require at least three kinds of tools: for searching in the Information extracted by Information Agents; for generating each DBMS for merging templates with DBMS data An example of "historical report" of CAI including "generative CAI" can be found at 13 An example of automatic translation performed using : English: this is done either directly, as favored routings of nerve impulse Italian: ciò è fatta uno direttamente, come percorsi favoriti dell impulso nervoso French: ceci est fait l un ou l autre directement, en tant que routages favorisés d impulsion de nerf Clearly, translations are quite primitive, and possibly incorrect. Further: the cited Altavista tool does not allow translations to our East European languages (other ones do). But an Italian or a French native speaker is helped quite significantly in the task of understanding the meaning of the English original word "routing" applied to nerve impulses. 14 Recent developments such as those reported in tln.fr/~gect/simm/sgmlql seem promising for each of the three tasks outlined here, and other ones reported in the paper. For the generation of Web pages and the Scheme based Literate Programming tool see also

11 THAI ETIS, Varese, Italy; June page 11of e Knowledge construction From Information to Knowledge (e.g. in terms of Conceptual Graphs 15 ) is the necessary next phase for Information Agents. Traditional Knowledge acquisition assumes that the Knowledge engineer imposes a structure to the natural language texts and produces formal representations, suitable to be visualized (and processed) in a variety of ways by programs. These Knowledge bases become then a basic component of systems that engage in dialogues with various purposes, in particular teaching (or, better, stimulating learning in) the students users. The key problem is classification 16, and the problem of classifications is that the classification criterion and the classification primitives are viewpoint dependent, non standard, etc. For instance, in [8] we have one of the best recent examples of Web technologies for classifying documents in order to produce catalogues. The methods and tools chosen by the authors represent an important example of automatic construction of Knowledge from Web information. The goal of the classification general purpose catalogues drives the classification criteria and thus the algorithms applied to the classification process. In our case, the Knowledge we should construct has a different purpose, i.e. should become a set of integrated software components for managing dialogues by humans with learning purposes. This entails a goal dependent definition of classification criteria before performing any classification. Not only we will have to design classification criteria for the "passive" components of our system (e.g. Web pages representing terminological knowledge) but also for the "active" components (e.g. exercises, simulations, etc.). Ontologies for learning seem to us a suitable starting point. 15 Other partners in the LARFLAST project (Sofia, Leeds, Manchester) have adopted a top down approach of Knowledge design by using Conceptual Graphs [Sowa, 1984 #239]. We consider their studies as potentially useful and complementary to ours, assuming that some effort is devoted to viewpoint dependent representations in CGs, such as those supported by a recent thesis defended at INRIA, Sophia Antipolis, for the Université de Nice by Myriam Ribière, with the title: Représentation et gestion de multiples points de vue dans le formalisme des graphes conceptuels (April 1999). 16 Even sequencing of concepts is a classification issue. For instance: the choice of Java (objects) or Scheme (functions) as a first programming language for students of Computing depends on the priorities one assigns in classifying concepts useful for learning Computing. Further: for each choice of a concept, one may present examples first and definitions afterwards, or the reverse. Assuming that the goal is that learners learn not only the concept s definitions, but particularly how to use each concept in each situation, one strategy may be preferred to the other one.

12 THAI ETIS, Varese, Italy; June page 12of e Dynamic classification criteria We are convinced that ad hoc coding by humans of structured Knowledge serves various purposes (e.g. in Expert Systems) where the knowledge is relatively stable as well as the goal of the system to be built (e.g. a diagnostic or a monitoring system for complex equipment). Our Knowledge (emerging from our Information) will not be stable, as well as the learning purposes, and therefore it will be more convenient to consider Knowledge construction as a semi automatic process guided by variable, parametric criteria. Concerning the instability of our Knowledge about relevant technical terminology, we have expressed above a few reasons that we may expand further. Concerning the instability of our learning purposes, we believe these purposes to be identified primarily starting from the "competencies" desired by prospective users. These competencies will be the major motivational force behind the use of our tools, so that we cannot avoid considering them of primary importance. Competencies relate Knowledge with skills, in particular the skills required by professions and jobs. They are therefore variable geographically, in time, among individual potential employers, etc. If competencies are variable, then the purpose of learners will be variable and therefore a traditional "curriculum based" system will be less interesting than a more dynamic set of customizable learning tools. A minimal description of competencies desired by perspective learners will have to guide the selection of an initial, possibly simple set of classification criteria that may help in the quasi automatic construction of Knowledge. For instance, in Electronic Commerce, if the competence required by a learner is to be able to buy electronic or other special purpose technologies from the net, then the skill of comparing technical performances among products offered may be privileged. On the contrary, if the skill is to sell via the net, then the competence of constructing convincing "commercially oriented" interfaces may be privileged. For the learner buyer the Knowledge base will include comparative evaluations (e.g. the Information Agent will be asked for URL s where the electronic products are compared by third parties). For the learner seller the Knowledge base will have to include "best site awards" that may interest the visitor. An example of tool for learning terminology is the one, for instance, that allows for the learner seller to learn the meaning of the term "award" and "site" and "best site award", while for the learner buyer technical terms may include acronyms abbreviations such as "bps" or "Mflops". The criterion for inclusion of a new term may be automatic as soon as a learner asks for an explanation for a term. The dictionaries should be customized and continuously updated for each learner. Minimal, but increasingly powerful relational structures may emerge, such as "Mflops and Mips are related terms", or "an award may be synonym with a prize" or "leasing is a kind of rental with a promise to sell the good leased". These relational structures yet in natural language may become candidates for simple conceptual representations.

13 THAI ETIS, Varese, Italy; June page 13of e Mastery of Terminology: a linguistic or a domain dependent, conceptual skill? The last examples suggest us to consider seriously the nature of learning we should envision, namely how understanding and using terminology in a domain requires linguistic and domain dependent competencies. Most of the terms, for instance, may not be translated at all, because in the target language where the learner is a native reader the concept denoted by the term is simply not present. In traditional linguistics, we recall the example of translation from Lappon, where a dozen of different terms denotating "types of snow" are commonly used, and recall quite different concepts in the hearer 17. A non native will need to know concepts about snow before learning the terms denoting those concepts 18. Similarly, the word "thread", synonym of "light weight process" will not be understood by anyone unless quite competent in Operating Systems and computational processes. A non trivial objective, the consequence of an apparently simple translation task. The notions of minimalist, dynamic, quasi automatic, interactive, collaborative, evolutionary Knowledge Acquisition were explored in the NAT*LAB project years ago (see, for instance, in [14] a retrospective view) when the Web was not available. The NAT*LAB acronym stood for "Student Model Acquisition in a Natural Laboratory": the community of users students and users teachers were supposed to collaborate through the NAT*LAB tools for enhancing student models, i.e. representations of "incorrect" Knowledge. We believe that the basic choices made in that project may be reused now in a more fruitful way by exploiting the Web. Concerning student models and the issue of individualizing the educational dialogues from the outcome of a diagnosis of the learner s incorrect knowledge, we have reported our efforts in [15; 16]. Due to the complexity of student 17 For a difference between connotation and denotation, see 201/denoconn.htm. 18 The classical example of "because" and "why" corresponding to the French "parce que" and "pourquoi" but to the single Italian "perchè", both in causal and interrogative contexts, forced us to propose in : S. A. Cerri, "ALICE: Acquisition of Linguistic Items in the Context of Examples," Instructional Science, vol. 18, pp. 1 30, 1989 to denote the difference between causal and interrogative sentences for Italians either by means of an introduction to linguistic analysis or by presenting differentially causal and interrogative contexts including each word. Two ways to give meaning to foreign language words that are not discriminated in Italian in such a way that you may not expect an Italian native speaker to be aware of the conceptual differences and thus be able to translate correctly "perchè" from Italian to French or English.

14 THAI ETIS, Varese, Italy; June page 14of e25 modeling, we will not give a priority to this approach in the LARFLAST project Communication Agents: Dialogues stimulating learning We have assumed that the stream of dialogues occurring at learning time is the stimulus to let people learn. But: what is learning time? Can an interaction with an indexed database of Web pages about English terms generated by an Information Agent be the only learning time we may consider useful? We will try to consider any kind of dialogues as potential stimuli for learning, not only those related to intentional activities for learning English terms 20. We envision dialogues in LARFLAST initially engaged by LARFLAST members the "producers" of the learning environment and gradually expanding also to "learners" the "users" with no real solution of a continuity. In building LARFLAST tools, we will have to consult electronic dictionaries, set up mailing lists, exchange electronic mail, send each other Web pages, build sites, criticize each other papers, statements or initiatives, suggest new directions etc. We do not see dialogues for learning as consisting only of interactions with ready to use materials. Learners, as well as experts, connected to the Web with a common interest (even if, a priori, with different roles) may also engage in dialogues for maintaining or improving the available learning tools offered by LARFLAST. Differently from a "product" that may be sold from a seller to a buyer, and therefore physically transferred from the first to the second, Knowledge is "constructed" by both partners in a dialogue as a result of the communicative process. The role of partners influences the construction process, but communication modifies the state of both agents. Artificial agents will "learn" from interactions with human agents, and the reverse. Among all properties of Agent technologies, the learning feature seems to us the most significant in the context of our project 21. Therefore we envision an evolutionary scenario that is centered around the agent to agent (human and artificial agents) communicative processes and not 19 The lessons learned from those experiences in diagnostic reasoning advice us to envision the use of diagnostic modules only in the cases where knowledge (and misconceptions) are stable. In a few of those cases, emerging from different domains, we have shown that it is relatively easy to design diagnostic modules by adopting concurrent, asynchronous software tools (Actors). 20 The "serendipitous" acquisition, thus, concerns for us not only the spidering but also the knowledge construction phase. 21 As we will see, the integration of Scheme and Java allows us to compile Scheme code at run time. This enables us modify the Agent s behavior dynamically as a result of agent to agent interactions.

15 THAI ETIS, Varese, Italy; June page 15of e25 just the specifically designed or automatically constructed tools, be them learning materials or active Computer Based Training modules. In order to understand, analyze and improve the effectiveness of the multi agent conversations we need to start with a formal model and experiment its properties in real conversations. This is the purpose of the activities concerning "Communication Agents" that we will hereafter outline The model: STROBE We have described elsewhere our model for representing and exploiting multi agent dialogues (e.g. [17; 18; 19; 20]). At the moment, two Agent Communication Languages have been build on the basis of the STROBE model (see, e.g. [21; 22; 23]). Notice that even if the STROBE model emerged originally from the area of Educational Technologies, it is generic with respect to any application domain where Agents and conversations are desirable. The model proposes two innovative components. Both represent components of reflective behavior that are inherent in realistic dialogues. The first is the notion of "first class environment" to model cognitive environments, i.e. viewpoints. These are necessary for reasoning about ambiguous classifications, and for engaging in dialogues with multiple autonomous agents. The second is the notion of Stream, i.e. delayed evaluation. One cannot plan ahead a dialogue beyond the next move, because the "program" to be executed by the next move may have to be constructed after receiving the partner s reaction to the first move, so that the underlying evaluation mechanism must allow the execution of code constructed at run time. As messages may be requests as well as definitions or orders, any Agent, upon reception of a qualified (yet possibly unexpected) "definition or order" may need to include the definition of a new procedure in its environment and subsequently run the (compiled) procedure, thus modifying dynamically its behaviour. The "Communication Agent" research issue that we wish to explore with our experiments concerns the set of concatenation rules that represent improve dialogues generated by the LARFLAST project, be it simply in the "domain of conversation for perspective learners (e.g: Electronic Commerce)", or possibly "in the course of the project itself". The result of this study will consist of software modules based on Agents that hopefully will enhance the effectiveness of LARFLAST conversations both in terms of learning by learners and in terms of improving the process of construction of tools for learning. In principle, the approach will again be evolutionary and bottom up. From a minimal set of basic "speech acts" we wish to show the value added by respecting concatenation rules (sometimes called conversational protocols) in

16 THAI ETIS, Varese, Italy; June page 16of e25 multi agent dialogues. Assume that the application scenario chosen will be that of Electronic Commerce. An example of well formalized auction (the "Spanish Fishmarket") has already been simulated with our Agent Collaboration Language AL [21] by combining acts belonging to three simple classes. Other, more complex situations in Electronic Commerce seem to us interesting sources of experimentally supported concatenation rules = dialogue protocols (in negotiations, argumentation: see, e.g. [24] that we wish to implement in our systems (and include in the corresponding version of the Agent Language). A formal starting point for those rules may be found in recent theses, (e.g. [25] 22 ) but experiments specific for learning have previously been described in [26;27;28] Scientific technical competencies We believe that learning scientific terminology should be intended not just as an exercise in foreign language acquisition, but instead as a process necessary for acquiring competencies (in the technologies) that will eventually be used in the professional life. Learning the meaning of "Archive" or "Directory" or "Thread" should not be intended for the Italian counterpart just learning "Archivio", "Cartella" or "Filo". On the contrary, we should focus on learning properties, examples, counterexamples, relations etc. of the above mentioned concepts when they are used in the technical contexts where they are going to be learned. Ambiguities and negotiations, explanations and justifications become components of the learning dialogues that cannot be underestimated. If the goal of learning is primarily "learning to use" instead of "learning to translate", the experimental set up becomes rather different and, possibly, more stimulating The bottom up approach to competencies At the moment, about "competencies" there is a rising concern in the industrialized world. Usually, in vocational training, people want to know, for instance, if "there is anybody competent in Java and Oracle". A rather "top down" approach. How can we define somebody "competent in Java"? Will (s)he know the syntax of Java, or, at the extreme, have been in the Java development group at SUN, or finally, have written several Java application programs industrially successful? We prefer to privilege an approach to "competencies" that is bottom up. We should define a few primitive competencies concerning aspects of our 22 In hohenheim.de/~ludwig/promotion.htm we may find another recent contribution to the same area.

17 THAI ETIS, Varese, Italy; June page 17of e25 "technical vocabulary" and define combination rules so those new competencies are defined (learned and used) as a function of the primitive ones. Like in Programming Languages: primitives combine to form higher level programming structures, and applications. When eventually these combinations of primitives will be "run" on data, they will perform according to the wishes of the programming team. From learning the technical vocabulary, e.g. in one of the sub areas of Computing, one should perhaps acquire (and demonstrate to have acquired) competencies in Computing. Such a choice shifts the acquisition process to the mastery of procedural skills that provide an operational context to the concepts denoted by the foreign language terms. 4. The tools 4.1. The Agent Communication Language(s) : AL and Jaskemal At the moment, two prototypical languages are available: AL (Alice Language) [21] and Jaskemal [20;23;22] 23. Improving the power of other Agent Communication Languages 24, these do have "cognitive environments" and allow to design "dialogue protocols" by defining simple transitions in the finite state automata that represent Agents engaged in conversations The content language We believe that there is a need for a meta level description language for the documents that will be generated by the Agents to construct the knowledge for the learners (see, for instance, a recent report on the use of SGML for learning Japanese [31]). Our viewpoint [32] is that markup languages, such as SGML, XML or other ones, reflect all a kind of Data Abstraction: tags are Data Types, and the functions methods processing these data types are "in the standard". We wish on the one side to be compliant with the emerging standards, on the other to be as free as possible to define our own abstraction level that will necessarily be application dependent. 23 The implementation language of AL is Scheme [29; 30] while Jaskemal integrates Scheme and Java. The documentation of Jaskemal is in progress. Both AL and Jaskemal are freely distributed writing to one of the authors. 24 E.g. KQML

18 THAI ETIS, Varese, Italy; June page 18of e25 In a quite limited experiment [33] we have noticed that it is possible and relatively easy to build a query update system for a DBMS associated to a Library such that messages are XML compliant and the operations may be performed via Web. An extension of this system is likely to become the kernel for constructing the Knowledge base associated to the English terminology The architecture In the following figure we present an overall architecture of the STyLE System 25 such as it is currently conceived by the Agents team at "delivery" time. It outlines the Information acquisition phase, the Knowledge construction phase and the Knowledge use phase, when Interfaces to the User allow the STyLE system to access via Web to the Knowledge available. However, it does not account for the idea that the STyLE system, as it is conceived by us, it will be improved continuously by the interactions between artificial Agents generating the system and human Agents using the system. Human agents considered will certainly have different roles, but at present we do not differentiate between roles: human agents (be them students, experts, teachers, administrators?) will have to be able not only to interact with the STyLE system by querying, but also by ordering or informing STyLE at run time. These orders, technically, result in two fundamental classes of operations: updating the DBMS of keywords that activate the acquisition and updating the archive of specifications that define the transformation from Information to Knowledge; updating the behavior (the Finite State Automata) of Agents in STyLE; i.e. modifying procedures at run time. Currently, Jaskemal agents allow both operations: a User connected to an Interface Agent may send requests to the Agent to access a remote DBMS, but can also update the DBMS content. Further, a User may also send an Agent a message that includes new procedures for the Agent to be performed under conditions. The new behavior is dynamically compiled (from Scheme to Java byte code) and the new Agent generated. The Information Acquisition phase has been implemented [34, 35] by using Northern Light and the DBMS MSAccess. The F.I.T.K phase is awaiting for the 26 specifications. The Knowledge Use phase kernel, based on a question answering conversation between Agents, has also been realized, as shown briefly by the example interaction in 4.4. The dynamic update from remote Agents of other 25 STyLE is the acronym chosen to denote the system constructed by the LARFLAST project. It stands for Scientific Terminology Learning Environment. 26 One of the search engines available

19 THAI ETIS, Varese, Italy; June page 19of e25 Agent s behaviors has been experimentally realized, but is not described further here. In order to interpret correctly the following figure, notice that the User StyLE interaction depicted is one of the many potential ones available. Users may take the initiative, or either respond to StyLE s initiatives, and the two phases of the dialogue may be intertwined. Either agent, taking an initiative, may essentially query, inform or order ; expected reactions by the partner will respectively be an answer, an acknowledgement (and the side effect of memorizing the information) and the execution of the action indicated by the order. When the User interacts with StyLE s DBMS, we may talk of "Knowledge Use"; when the User interacts with the DBMS of KeyWords or the Learning Dependent Specifications, then we may better talk of "Knowledge Construction".

20 THAI ETIS, Varese, Italy; June page 20of e25 Phase 1: Serendipitous Information Acquisition from the Web DBMS of KeyWords Web search (spidering) WEB Learning dependent Specifications DBMS of Sequential Files Documents Phase 2: F.I.T.K. From Information To Knowledge DBMS of Docs with Semantic Mark ups Phase 3: Knowledge Use (Web page generation) FILTER INTERFACE QUERY FILTER FILTER INTERFACE ANSWER USER

21 THAI ETIS, Varese, Italy; June page 21of e An example of a simple interaction In the figure above, we show an extremely simple dialogue between two agents I_Agent and D_Agent, through the Net. The first one represent an Interface Agent (connected to a User) and the second one a DBMS Agent, i.e. the Agent responsible for accessing a DBMS. The numbers identify the sequence of exchanges among Agents during the conversation. Messages in the "question answering" conversation generate English texts for a trace of events while other messages include Italian texts. D_Agent sends 4 records about the words thread and light_weight_process that refer to URL s where the words occur. On the left, the evidence of three other Agents present in the system (an Agent Name Server, a Policy Agent and a Service Agent) with the number of messages exchanged during the connection.