An Interactive Intelligent Language Tutor Over The Internet Trude Heift Linguistics Department and Language Learning Centre Simon Fraser University, B.C. Canada V5A1S6 E-mail: heift@sfu.ca Abstract: This paper describes an Intelligent Language Tutoring System (ILTS) for German. The German Tutor is the grammar component of a comprehensive multi-media course, specifically adapted for distance education over the World Wide Web. The Intelligent Tutor parses student s answers to introductory language exercises. The system analyzes the linguistic source of an error and reports its description to the student. Through this analysis it infers an approximation of the learner s competence and adapts the instructional strategy accordingly. The main advantage of the system is its ability to individualize instruction and to provide insight to the student as an inherent feature of learning exercises. 1. Introduction The project describes an Intelligent Language Tutoring System (ILTS) for Distance Education over the World Wide Web. The predominant characteristic of Distance Education is the separation of student and instructor in space, and consequently, time. One-way communication media, such as radio, television, film and video came to dominate the field of distance learning because of their ability to deal at least partially with this difficulty. While broadcasting technologies using a single-source/multiple-receiver model, greatly enhanced print-based instructional media, feedback within the model was awkward or non-existent. It is, however, commonly accepted that the most effective education, particularly in second-language learning, requires two-way communication. For this reason, Computer Mediated Communication (CMC) has become to play an ever more important role in both traditional and distance education. The German Tutor, an Intelligent Language Tutoring System is conceived of as the grammar practice module in a distance education course over the Internet. An Intelligent Language Tutor over the World Wide Web combines the flexibility and intelligence of an ILTS with world-wide availability of WWW applications [Brusilovsky et. al, 1996]. Predominantly, existing second language learning applications use simple grammar practice and their feedback mechanism is more restricted than their off-line counterparts. The goal of the German Tutor is to provide meaningful grammar practice for foreign language learners through computer-mediated communication. Meaningful grammar practice requires intelligence on part of the computer program. Without intelligence the system is merely a method of presenting information, one not especially preferable to a static medium like print. In order to go beyond multiple choice questions, relatively uninformative answer keys, and gross mainstreaming of students characteristic of workbooks, the German Tutor emulates two significant aspects of a student teacher interaction: it provides error-contingent feedback and allows for individualization of the learning process. Generally, Intelligent Tutoring Systems consist of 3 components: the domain knowledge, the student model, and the teaching module [Wenger, 1987].
Natural Language parsers model the domain knowledge in Intelligent Language Tutoring Systems. The strength of Natural Language Processing is that the student tasks can go beyond multiple-choice questions and/or fill-inthe-blanks while still allowing for a sophisticated error analysis. Simple drills are based on string matching algorithms, that is the student response is compared letter by letter against an answer key. With meaningful grammar practice, however, one obviously cannot enter arbitrarily many sentences into memory for purposes of comparison. Natural Language Processing allows for more than a mere indication that an error has occurred: it can give a description of the error, and go to an even deeper linguistics analysis in order to isolate the source of the error. Thus a Natural Language parser provides the analytical complexity in an Intelligent Language Tutoring System. Students all learn at their own pace and the student model provides the key to individualized knowledge-based instruction [McCalla & Greer, 1992]. The student model keeps track of the learner history and provides learner model updates. It is used by the system to inform the teaching module and allows the system to adapt to individual student needs and alter the instructional process accordingly. 2. Grammatical Formalism The German Tutor parses students answers to introductory German exercises. Typically Intelligent Language Tutoring Systems augment grammars that parse grammatical input in one of three ways. They may augment rules such that if a particular rule does not succeed, specific error routines (meta-rules) force application of the rule by systematically relaxing its constraints, or they may augment the grammar with rules which are capable of parsing ill-formed input (buggy rules) and which apply if the grammatical rules fail (see, for example, [Liou, 1991], [Carbonell and Hayes, 1981], [Weischedel, 1983]). With feature grammar formalisms, they may also alter the unification algorithm itself such that in the event of conflicting feature values the parse does not fail, but instead applies a different set of procedures [Hagen, 1994]. In particular, parsers designed for language instruction typically contain components which search for errors in the event that the grammatical rules are not successful. Unlike other systems, the German Tutor does not seek errors, but instead records whether or not grammatical constraints are met. The system returns structures which provide possible feedback and student model updates of different level of specificity. The computational analysis reflects the underlying pedagogy of the system. The goal is to analyze students language input rather than the more modest requirement to recognize ill-formed construction. The German Tutor is written in ALE (The Attributed Logic Engine), an extension of Prolog. ALE is an integrated phrase structure parsing and definite clause programming system in which the terms are typed feature structures. Typed feature structures combine type inheritance and appropriateness specifications for features and their values [Carpenter and Penn, 1994]. The grammatical formalism used is derived from Head Phrase Structure Grammar [Pollard & Sag, 1987 and 1994], a unification-based grammar theory. It is one of a family which share several properties. Linguistic information is presented as feature/value matrices. Theories in this family are to varying degrees radically lexicalist, that is a considerable amount of grammatical information is located in the lexicon rather than in the grammar rules. 3. Program Architecture The pedagogical goal behind the German Tutor is to provide error-contingent feedback and allow for individualization of the learning process. For example, if a student chooses a wrong article in German the error might be either incorrect gender, number, or case. In such an instance the program must be capable of distinguishing between the three error types since... for almost all cognitive learning, instruction is enhanced by evaluative feedback [Venetzky, R. & L.Osin, 1991, p. 9]. In addition, inexperienced students might require detailed instruction while experienced students benefit from higher level reminders and explanations. To achieve this, the German Tutor, consists of four components: the Domain Knowledge, the Analysis Hierarchy, the Student Model, and the Filtering Module, given in Figure 1.
Domain Knowledge Analysis Hierarchy produces feedback parser and grammar messages of increasing abstraction STUDENT INPUT sends phrase descriptors of all parses to FM sends phrase descriptors of selected parse to AH sends feedback messages produced by the phrase descriptors Filtering Module 1) selects the desired parse 2) decides on the order of feedback messages sends feedback messages suited to the level of the learner Student Model keeps learner model updates and decides on student level ERROR-CONTINGENT FEEDBACK The Domain Knowledge consists of a parser with a grammar which parses sentences and phrases to produce sets of phrase descriptors [McFetridge & Heift, 1995]. A phrase descriptor is a model of a particular grammatical phenomenon such as case assignment or agreement. A phrase descriptor records whether or not the grammatical phenomenon is present in the input and correctly or incorrectly formed. The second component is an Analysis Hierarchy which takes a phrase descriptor as input and generates sets of possible responses to the learner s input. Since the phrase descriptors record both successful and unsuccessful attempts at practising grammatical knowledge, the student model can record mastery of grammatical structures as well as structures with which the learner has problems. The system tracks the student model. On the basis of the student model, the Filtering Module, and the Analysis Hierarchy the program interacts with the learner. 3.1. The Domain Knowledge The goal of the parser and the grammar is the generation of phrase descriptors. A phrase descriptor is implemented as a frame structure that models a grammatical phenomenon. Each member of the frame consists of a name followed by a value. For example, number agreement of subject-verb in German is modeled by the frame [number,_] where the underscore represents a value for each number. Consider examples (1) and (2): (1) *Die Familie gehen nach Paris. (2) Die Familie geht nach Paris. The family is going to Paris. Figure 1: The German Tutor The phrase descriptor for (1) is [number,error], while that for the phrase in (2) is [number,none]. A system
presented with (1) will instruct the learner on the nature of number agreement while the successful number application in (2) will be recorded in the student model. In addition to the grammatical features defined in HPSG the grammar uses a type descriptor representing the description of the phrase that the parser builds up. This type is set-valued and is initially underspecified in each lexical entry. Thus descriptor not only tracks whether the input is ill-formed but also records whether the sentence is grammatical. During parsing, the values of the features of descriptor are specified. For example, one of the members of descriptor, vp_num in Figure 2, tracks the number agreement of subject-verb in a main-clause. Its value is inherited from the sg feature specified in the verb geht. cat head v content index num sg 1 descriptor main_clause vp_num 1 3.2. The Analysis Hierarchy The second component of the system is an Analysis Hierarchy. The purpose of the Analysis Hierarchy is to take phrase descriptors as input and generate possible responses that the instruction system can use when interacting with the learner. A response is a pair that contains a message the system will use to inform the learner if a phrase descriptor indicates there has been an error and a student model update. The student model update is a name of an error category in the student model with an increment or decrement. The Analysis Hierarchy generates sets of responses of increasing abstraction. As an example consider the ungrammatical phrase in (3). An experienced student should be informed that Ball is a masculine noun, that the preposition mit is a dative preposition and that the determiner der is incorrect. A student who has mastered case assignment (as indicated by the student model) may be informed only that the case of the prepositional phrase is incorrect. (3) *mit der Ball mit dem Ball with the ball The Analysis Hierarchy is implemented in DATR [Evans and Gazdar, 1990], a language designed for representing multiple inheritance. The language is well suited to constructing responses from phrase descriptors. The Sussex version of DATR is implemented in Prolog and consequently the interface between the parser and the Analysis Hierarchy is a natural one. 3.3. The Student Model and the Filtering Module Figure 2: Descriptor vp_num of the verb geht The two remaining modules of the German Tutor are the student model and the Filtering Module. The student model keeps track of the learner history and provides learner model updates. On the basis of the learner model, it decides on the specificity of the feedback messages. There are three different kinds of learner levels considered in the system: the novice, the intermediate learner, and the expert. The student model passes feedback messages to the Filtering Module suited to the level of the learner.
The Filtering Module has two tasks: it filters all parses and feedback messages. Due to language ambiguity parsers generally produce more than one interpretation. The first task of the Filtering Module is to decide on the desired parse. The second task is to determine the order of the feedback messages showed to the learner. The Filtering Module displays one error at a time and once the student makes the required correction, the whole evaluation process starts from the beginning. 4. Conclusions The system described is the grammar component of a Distance Education course for German on the Internet. The German Tutor creates error-contingent feedback and learner model updates of different levels of specificity thus allowing for individualization of the learning process. The German Tutor contains grammatical constructions of an introductory course for German although at this point the lexicon is limited and needs to be expanded to fully encompass students input in such course. The suitability of the computational analysis used in the German Tutor is two-fold: first, the grammar is sufficiently general that it does not distinguish between grammatical and ungrammatical input for the phenomena it is designed to handle. This generality has the advantage of reducing the number of rules required by the grammar. Second, the decoupling of the parsing system from analysis of whether or not the input is grammatical has the practical advantage that development of each can proceed independently. 5. References [Brusilovsky, Schwarz, and Weber 1996] Brusilovsky, P., Schwarz E. and Weber G.. (1996). ELM-ART: Intelligent Tutoring Systems on World Wide Web., in Frasson, C., Gauthier G., Desgold, A. (Eds.). Intelligent Tutoring Systems. 259-268. Springer Verlag. [Carbonell and Hayes 1981] Carbonell, J. G. and Hayes, P. J. (1981). Dynamic strategy selection in flexible parsing. Proceedings of ACL 19th Annual Meeting. 143-147. [Carpenter and Penn 1994] Carpenter, B. and Penn, G. (1994). The Attribute Logic Engine. User s Guide, Carnegie Mellon University. [Evans and Gazdar 1990] Evans, R. and Gazdar, G. (1990). The DATR Papers, Volume 1. Brighton, School of Cognitive and Computing Sciences, The University of Sussex. [Liou 1991] Liou, H.-C. (1991). Development of an English grammar checker: a progress report. CALICO. 9, 27 56. [McCalla and Greer 1992] McCalla, G., and Greer, J. E. (1992). Student Modelling: The Key to Individualized Knowledge- Based Instruction, in Greer, J. E., G., McCalla, (Eds.).NATO ASI Series. Series F: Computer and Systems Sciences. Springer Verlag. Vol: 125: 39-62. [McFetridge and Heift 1995] McFetridge, P. and Heift, T. (1995). Grammar and Analysis for Implementing Granularity in Computer-Assisted Language Instruction. Educational Multimedia and Hypermedia, 1995. Graz, Austria: Association for the Advancement in Computing in Education, 1995. pp. 460-465. [Pollard and Sag 1987] Pollard, C. and Sag, I. (1987). Information-based Syntax and Semantics: Fundamentals. CSLI Lecture Notes. Stanford, Center for the Study of Language and Information. [Pollard and Sag 1987] Pollard, C. and Sag, I. (1994). Head-Driven Phrase Structure Grammar. Chicago, Chicago University Press. [Venetzky and Osin 1991] Venetzky, R. and Osin, L. (1991). The Intelligent Design of Computer-Assisted Instruction. London, Longman Publishing Group. [Weischedel 1983] Weischedel, R. M. (1983). Handling ill-formed input, in Proceedings of the Conference on Applied Natural Language Processing, 89-92. [Wenger 1987] Wenger, E. (1987). Artificial Intelligence and Tutoring Systems. Computational and Cognitive Approaches to the Communication of Knowledge. Los Altos, Morgan Kaufmann Publishers, Inc.