The NLP Role in Animated Conversation for CALL 
Michael Schoelles and Henry Hamburger 
Dept. of Computer Science 
George Mason University 
Fairfax, VA 22030 USA 
mikejs@cs.gmu.edu & henryh@cs.gmu.edu 
Abstract 
Language learning is a relatively new appli- 
cation for natural language processing (NLP) 
and for intelligent tutoring and learning 
environments (ITLEs). NLP has a crucial role 
to play in foreign language ITLEs, whether they 
are designed for explicit or implicit learning of 
the vocabulary and grammar. FLUENT is an 
implicit approach, in which NLP and shared- 
control animation support a two-medium 
conversation, designed to foster implicit 
learning of language. This report highlights 
specific ways that FLUENT already uses NLP, 
suggests potential benefits from additional use 
of NLP and grounds the method in widely used 
language learning pedagogy. It concludes by 
describing and evaluating the system's use in 
the classroom, with a particularly challenging 
kind of learner. 
1 Introduction 
Language learning is a relatively new application for 
natural language processing (NLP), compared to 
translation and database interfaces. The 1988 Trieste 
conference on foreign language intelligent tutoring 
systems did not manage a proceedings, so few were the 
functioning systems using NLP and other techniques of 
artificial intelligence (AI). Indeed the field of 
Computer-Assisted Language Learning (CALL) has 
been dominated by the work of language teachers who, 
tired of waiting for us, have proceeded to learn some 
programming or hire computer science students to make 
their own systems, most of them just for fairly 
inflexible drill-and-practice. The 1990s have witnessed 
increasing interest in bringing AI to language learning 
systems, reflected in Chanier's (1994) special issue, 
Holland's anthology (1995) and Zock's (1996) panel. 
The AI is principally of two kinds: NLP and what we 
will call ITLE, the overlapping categories of intelligent 
tutoring systems and interactive learning environments. 
In this paper, we report on a foreign language ITLE that 
has moved from the lab into the classroom. After 
noting alternative strategies for using NLP in CALL 
(section 2), we describe our system and its NLP require- 
ments (section 3), as well as its pedagogical foundations 
(section 4). Finally, we describe its graduation into the 
classroom, for use with a particularly challenging kind 
of learner: the highly motivated but computer- 
unfamiliar immigrant. 
2 A Crucial Role for NLP in CALL 
Developers of foreign language ITLEs have only 
recently begun to expand the use of NLP in language 
learning systems beyond relatively simple uses of 
syntax. More sophisticated and widespread use has 
awaited greater robustness and ease of use of NLP 
software as well as better speed and memory of the 
hardware. 
Recent improvements along these lines are playing 
key roles in meeting the special NLP requirements of 
foreign language ITLEs. Since the language is the 
subject matter to be learned, ITLEs must expect the 
input to be in error, and therefore require very robust 
NLP systems with error detection capabilities. In 
addition, conversational ITLEs must respond to student 
input in a time frame that parallels natural dialog. 
Advances in parsing technology and faster machines 
with more memory are making NLP attractive to the 
developers of ITLEs. The flexibility, increased coverage 
and other advantages of NLP are finding their way into 
the design of language learning systems. The issue is 
no longer whether language learning systems should use 
NLP, but how NLP can best be applied to language 
learning. 
One possible strategy for creating a language ITLE 
can be called the direct approach. A tenet for ITLEs in 
general is that they represent the knowledge to be 
learned in a domain expertise module. In the case of 
language learning, the knowledge to be learned already 
exists in computational form as the lexicon and 
grammar of a NLP system, so the idea would be to 
incorporate these NLP knowledge bases as the ITLE 
domain expertise. However, successful language use, in 
contrast to other ITLE domains, is not just a matter of 
overt stepwise reasoning. The goal of language 
127 
learning is not (only) to conjugate verbs or passivize 
sentences, but to understand and produce meaningful 
language in context. Even for the limited role of 
teaching grammar explicitly, a direct approach 
encounters two obstacles. First, a performance-oriented 
representation of domain knowledge is typically not 
structured in a way that is cognitively suitable for 
learners (Clancey, 1987). Moreover, NLP knowledge 
bases cannot be used directly in explanation since their 
representation language is incomprehensible to students 
(Chanier et al., 1992). 
To avoid these pedagogical problems, we turn to an 
alternative strategy that deploys NLP principally in the 
ITLE's learning environment, rather than as explicit 
expertise. This approach is more concerned with the 
performance of the NLP rather than its competence, 
ITLEs applying NLP in this manner immerse the 
student in the language by situating the student in a 
life-like conversational setting. 
Pedagogical motivation for this kind of foreign 
language learning environment comes from classroom 
methods like Total Physical Response or TPR (Asher, 
1977) and the Natural Approach (Krashen and Terrell, 
1983). These techniques require only comprehension at 
first, as students act out appropriate responses to the 
teacher's commands. As student actions indicate 
sufficient progress, constituents of the utterances are 
recombined to increase the skill at understanding. There 
are also questions, which TPR students answer by 
pointing. In the natural approach, production is 
included, beginning with questions that take very short 
answers. There is a gradual progression from yes-no 
questions to other single-word-answer questions and 
either-or questions, and then on to phrases and 
ultimately sentences. 
The flexibility, efficiency and error handling of NLP 
is crucial to implementing the above pedagogy in an 
ITLE. A system that puts NLP to crucial use in 
support of these pedagogical considerations is FLUENT, 
to which we now turn. 
3 FLUENT and NLP 
FLUENT uses NLP to converse with the student in the 
context of a realistic microworld situation. Actions 
performed by both the student and the system tutor 
appear as context-dependent animation and cause 
updates in the underlying representation. The 
generativity of both the NLP and the animation is 
crucial to the flexibility of this conversation, allowing 
it to be steered in pedagogically useful directions, yet 
also permitting student exploration. The NLP module 
generates descriptions, commands and questions that 
may be related to actions that the student carries out 
graphically with the mouse or may be related to the 
states of objects in the current situation. The generated 
language appears as both text and speech. Figure 1 
shows FLUENT's Kitchen Microworld. 
~.~ Quit Style Variety Activity Who Controls Tools I 1:54:10\[~ I~I (~ 
Conversational Styll¢.: Quizmaster Control: Tutor Activity: COOK POTATOES 
Tutor: 
Student: 
Kitchen World 
What did I turn on? 
Pause 1 
i i 
Figure 1. A FLUENT Microworld 
128 
/ ........   ects 
/ Grammars ~ ~agguage ~ne~ Acti°ns 
\View ~ ~ Pedagogical ~ws~Vi~ Processor /i~ ns 
Schemas 
Tutorial Schema Tool Y 
Figure 2. FLUENT Architecture 
This section describes first the overall architecture of 
FLUENT and then turns to the specific NLP 
components, noting their roles and requirements in the 
context of FLUENT'S pedagogical goals. Flexibility, 
extendibility and teacher involvement are achieved in 
FLUENT through its layered architecture, shown in 
Figure 2. In this framework, independent processes 
work on declarative structures that have been acquired by 
the system through tools that do not demand a 
knowledge of programming or linguistic theory 
(Schoelles and Hamburger, 1996). This work shares the 
interface orientation of Caldwell and Korelsky (1994), 
but is more abstract and flexible in that user- 
specifications are independent of, and combinable with, 
domain plans. On the other hand we do not follow 
work like Traum and Allen (1994), in pushing toward 
computation about more and more discourse phenomena 
The outer ring of Figure 2 depicts the knowledge 
acquisition level. It identifies existing tools for 
building tutorial schemas, language usage structures and 
the graphics of objects. The tutorial schema tool lets a 
teacher express pedagogical expertise in the form of 
lessons. The view tool allows the teacher some degree 
of control over the language generated by the system. 
The teacher inputs the language specifications for a 
lesson by manipulating a graphical user interface. The 
tools process these specifications to produce instantiated 
objects for the declarative level. 
The knowledge base, in the second ring, consists of: 
• Tutorial Schemas - structures, created by a teacher 
using tools, that coordinate the activities, 
conversational style and conversational variety of 
FLUENT when the system is in tutoring mode. 
• Views - structured objects that provide language 
specifications and constraints to achieve 
conversational variety, control over the difficulty of 
input presented to the student, and situational 
flexibility. 
• Lexicon - currently 550 words entered in base form, 
with information for finding inflected forms and 
with irregular forms are subentries. Contains both 
subcategorizafion and theta grids. 
• Grammar- context-free productions plus reduction 
functions that restrict the structure, perform 
agreement and other checks, build the parse tree and 
reference the case frame interpreter. See Felshin 
(1995) for more on the lexicon and grammar. 
• Plans and actions - rules, expressed in terms of 
classes of objects, that specify how the tutor 
organizes activity in the microworld, how the 
student initiates actions and how the microworld 
state is altered. 
• Microworld - a hierarchy of objects with their 
situational and graphical properties, as well as the 
associated plans and actions 
1 29 
The procedural level in the third ring shows the 
processes that operate on the knowledge base to provide 
language tutoring or a learning environment. The main 
processes that achieve these functions are: 
• Pedagogical Processing - to decide what the tutor 
will do and say next, by selecting a tutorial schema 
from the knowledge base and executing it. 
• NLP Processing - a generative system which 
interprets language specifications in declarative 
form. These specifications are derived from the 
tutorial schema (incorporating views and 
microworld information) to construct an abstract 
syntactic structure, which in turn undergoes 
grammatical and morphological processing to 
produce a sentence. 
• Situation Reasoning - a plan executor that 
instantiates the plan and action rules based on the 
current situation. 
• Microworld Processing maintains object 
information and performs the graphical actions in 
the microwodd. 
The innermost ring represents the system executive 
which controls the interaction between the student and 
the system and coordinates the main processes based on 
this interaction. The student's interaction with the 
system takes two forms, learning and control. In a 
learning interaction, the student causes linguistic and 
graphical events in the microworld. In a control 
interaction the student can take control of choices about 
the plans and actions, the conversational style and the 
degree of linguistic variety. Further details on the 
operation of FLUENT can be found in Hamburger (1995) 
and in Schoelles and Hamburger (1996). 
The remainder of this section describes the functions 
of NLP within the FLUENT architecture. We point out 
their importance in this framework to FLUENT'S 
pedagogical goals. As seen in Figure 2, NLP is 
involved in three levels of the architecture: the 
knowledge acquisition level, the declarative level and the 
procedural level. 
FLUENT, like many other natural language generation 
systems distinguishes between an early or deep phase 
that determines content and organization (what to say) 
and a later or surface phase for lexical content, 
morphology, and syntactic structure (how to say it). 
The deep phase is driven by the current tutorial schema, 
which, as noted earlier, coordinates the activities, 
conversational style and conversational variety. 
A tutorial schema is a list of triples, each consisting 
of a plan or action, an interaction type and a list of 
views, as in the example in Figure 3. The resulting 
conversation depends in part on student responses and 
some randomized choices, but would be something like: 
I am picking up a cup. \[Tutor puts cup on table.\] 
The cup is on the table. There is another cup on the 
sink. What is on the table? \[Student: The cup.\] 
Good! Pick up a cup. \[Student turns on the water(!)\] 
You did not pick up a cup, but you turned on the faucet. 
Turn off the faucet. \[Student does so.\] Great! You 
turned off the faucet. Pick up a cup... 
Plan/Action Interaction 
Pick-up Tourguide 
Put-down Tourguide 
None Quizmaster 
Pick-up Commander 
View 
Present-Progresv 
(Location 
Location-other) 
Wh-on-top 
(Command-indef 
Present-action 
Command-fail) 
Figure 3. Part of a Tutorial Schema (Lesson) 
A plan groups together subplans and individual 
actions to achieve a goal. The plans and actions are 
flexible in that their arguments are variables which can 
be bound when the schema is created, or at the time of 
execution. The sequence of plans and actions 
determines what to talk about by specifying predicates 
and arguments. 
Another component of a tutorial schema is the 
interaction type. There are three interaction types that 
can be specified in a tutorial schema. In Tourguide 
mode, the system performs the actions as well as 
generating language about the action or state. 
Commander mode is similar to Tourguide mode, except 
that for each action in the plan the system tells the 
student what to do (i.e. gives the student a command), 
the student then uses the mouse to perform the action 
command. The system will respond based on whether 
the student successfully or unsuccessfully carried out the 
command. In Quizmaster mode, the system performs 
an action and then asks a question about the action 
performed. The student responds by typing in an 
answer (i.e. a noun phrase) which is checked by the 
system. 
How to talk about the predicates and arguments or 
what type of question to ask is specified in the view. 
This data structure, whether selected or created using the 
view tool, becomes part of the tutorial schema and 
serves as the input to the surface generator. As an 
example, a portion of the Command-fail view used in 
Figure 3 is spelled out in Figure 4. It gives rise to the 
final portion of the conversation example accompanying 
Figure 3. 
130 
View-Type: Complex 
Relation: Contrast 
Subview-I : View-Type: 
Polarity: 
Subtype: 
Tense: 
Subview-2: View-Type: 
Polarity: 
Simple 
Negative 
Action 
Past 
Simple 
Positive 
Figure 4 - A Specific View 
The view is the interface to the syntactic/semantic 
component. In a tutorial schema, an interaction type 
and view can by specified for entire plans, actions 
within a plan, or individual actions. Thus, at each step 
in the schema, how to talk about the 
predicate/argument structure can be specified. Views are 
also specified to handle error conditions. Some of the 
parameters specified in a view are the tense, aspect, 
modality, type of question (Wh or Yes/No), minimum 
or maximum number of arguments, and definiteness of 
a noun phrase. Views also specify whether to talk 
about the action itself or the state of the world before or 
after the action. 
Views mediate between the language teaching 
functions of the system and the NLP component that 
performs the actual text generation. This component is 
the NLG system developed for the Athena Project at 
M.I.T. by Sue Felshin. It is responsible for 
morphology, syntax, and some semantic functions. It 
was chosen because it does provide some semantics and 
a well-defined language independent interface called an 
Interlingua Structure (ILS). The ILS is a specification 
of data structures containing syntactic and semantic 
information. These structures are built by the 
knowledge module and passed to the NLG. The specific 
semantic features that this framework provides is an 
extensive hierarchical system of theta roles. 
Generation of linguistic responses by the system 
translates from the interlingua structure to Case 
Frame(CF)-structure to D-structure and to S-structure in 
separate steps. First, focusing chooses how to refer to 
objects. CF-structure is then built. Next, D-structure is 
built by applying the parameters of the English to 
principles of universal grammar. S-structure is 
produced by applying code transformation to D- 
structure. Finally, the S-structure is flattened to 
produce printable text (Felshin 1995). 
A limitation of the Athena framework is the lack of 
an interface tool to the lexicon. Lexical entries can 
only be added or changed by editing the source files and 
recompiling the lexicon. A task that can not be easily 
performed by a language teacher or exercise designer. 
Another limitation is that in some instances when 
variation in surface structure is possible the choice is 
made randomly thereby diminishing the application's 
control over the language generated. 
4 FLUENT Pedagogy and NLP 
Even a flexible, extendible and teacher-accessible system 
may be of little value unless it is also engaging and 
beneficial to students. Our initial results, presented in 
the next section, suggest that it is engaging to the 
students who have used it. Here we argue that one 
should expect it to be beneficial as well, in view of its 
relationship to several pedagogical issues. 
Language learning pedagogy is made complex by the 
great complexity and variety of both languages and 
learners. It is therefore out of the question to make a 
universally useful pedagogy-based requirements 
statement and proceed to the ultimate, correct system 
design. Our aim in designing FLUENT has thus never 
been to create a panacea, but we do claim to be 
developing an important form of language learning 
interaction. Besides supporting this claim, we also 
point out the crucial role of NLP in enabling our 
ambitious approach. 
We consider five pedagogical issues. The first is 
specific to the domain of language: the explicit teaching 
of language facts versus implicit learning from ordinary 
conversation about non-language topics. The next two 
issues - situatedness and error-handling - are of general 
concern but play out in unique ways in the domain of 
language. Finally, active learning and the choice of 
difficulty level are issues that relate to general cognitive 
considerations and so are relevant for the learning of 
other domains as well as for language. 
The explicitness issue arises from a curious 
observation: all children succeed at implicit learning of 
their first language, while many adults fail in attempts 
to learn a second one. Is it that children spend all day at 
it, or, lacking a language, are more motivated? Do 
adults know too much, including another language, or 
have they lost some language capacity in maturation? 
Or are they just not properly taught? If it's the 
teaching, can we adapt the child's immersion experience 
to adults, taking care to make the situation simple and 
clear at all times as we do with toddlers? 
Success with immersion is not confined to first- 
language learning. Many an adult who has experienced 
131 
second-language immersion will testify to its benefits, 
and Krashen reports classroom benefits with his 
method. FLUENT provides a conversational setting 
where a linguistic fact is not presented explicitly but 
rather can be indirectly acquired by the student through 
exposure to it in a meaningful context. To learn to 
communicate, one must sooner or later practice doing 
so, and FLUENT permits practice in an ongoing and 
completely understandable situation. 
NLP plays the crucial role of providing the variability 
required by the conversation. Even fairly simple 
microworlds lead to a combinatorial explosion of 
possible events and situations, along with a 
combinatorial explosion of sentences that must occur in 
appropriate circumstances. 
Next, consider the pedagogical proposal that situating 
a realistic activity in an authentic environment can 
promote learning both cognitively and motivationally. 
Without pursuing this issue in general, we note that it 
has a special significance for language, because the 
choice of a linguistic construct often hinges on aspects 
of the current situation, including the conversation 
itself. These pragmatic and discourse phenomena can be 
made to arise in FLUENT, since it conducts conversation 
in an ongoing situation. The resulting practice on 
these aspects of language can lead to the kind of 
implicit learning described above. 
An example of a situation-based aspect of language 
that FLUENT presents to students is that the choice of a 
definite article may depend on there being only one 
object of some type. Another is that different verb 
forms are used in descriptions that occur during versus 
after the action described. An example that relates to 
the conversation itself, as opposed to the situation, is 
that the discourse center is a prime contender for 
pronominal referent. Another is the correlation of 
imperative, interrogative and declarative forms with the 
type of conversational interaction. 
Pragmatics and discourse are very difficult for NLP. 
Although FLUENT does not engage in sophisticated 
NLP-style discourse reasoning, its tutorial schema tool 
and view tool make it easy for us or a teacher to 
prescribe the right discourse results without losing 
computational flexibility at the semantic and syntactic 
levels. 
A third pedagogical issue is the handling of syntactic 
and semantic errors made by the student. This issue is 
also related to the explicitness issue. For people and 
systems that teach linguistic form explicitly, error 
explanations are paramount. In the communicative 
approach, errors are tolerated in order to achieve the 
primary goal of communication. It is argued that 
explanations are often not well understood and in any 
case disrupt communication. 
132 
A fully communicative system needs robust NLP on 
the language input side to make some sense of what the 
student is trying to communicate even in the presence 
of errors. We have had very positive response from 
teachers to a different use of NLP, on the generation 
side. In FLUENT, a misconception on the student's part 
is often indicated by an incorrect action performed by 
the student in relation to a command given by the 
system. In such a case, the system can (if the teacher 
has so specified) comment on the contrast between the 
instruction and the actual student action. To do so 
requires a two-clause sentence, of the form "I asked you 
to <do\[present\] action-x>, but you <do\[past\] action- 
y>." Generating such responses is clearly a job for 
NLP. Before getting back to a repeat of the original 
command, the NLP module generates an instruction to 
perform an action to undo the original erroneous action. 
A fourth pedagogical issue is active versus passive 
learning. Active involvement by the student has proven 
successful in many domains. For language learning in 
particular, TPR and the natural approach (see section 2) 
let students respond to a human teacher both via 
language, in the case of questions, and with actions, for 
commands. In the computational realm, ITLEs use 
problem-solving environments to afford students the 
opportunity to act. In its tutorial mode, FLUENT is an 
1TLE for TPR. Beyond that, in its student mode (and 
student-driven aspects of tutorial mode too) the student 
can actively participate by initiating and controlling 
actions though the mouse. 
The NLP component can express these student 
actions as verbs. The benefit of a NLP approach over a 
direct mapping between mouse actions and stored text, 
or even templates with fillers, is that a wider coverage 
of language can be achieved in an efficient manner. For 
example, a click on a graphical object could represent 
the action of picking up that object. To express this in 
past tense versus present tense only requires a change of 
a single parameter to the NLP module, whereas with a 
direct mapping two separate sentences must be stored. 
Clicking on a different object for a NLP based system 
again requires only the changing of a single parameter. 
Moreover, the same object can be described in terms of 
different properties at various times, according to the 
situation and discourse. The combinatorial nature of 
this design clearly permits the system to expose the 
student to a greater variety of language than would be 
practical for a non-NLP system. 
Finally, a NLP component provides the flexibility to 
give the student material at the right level of difficulty. 
In FLUENT, simply by varying a parameter the system 
can generate a command, a declarative statement or a 
question. Commands require understanding, but 
questions are more demanding, in that they require 
language output from the student as well. Another way 
to vary difficulty involves verbs with optional 
arguments. Use of NLP lets us choose whether a 
sentence is to contain all the arguments that a verb can 
subcategorize for or only the required ones. Again this 
is done by changing a single parameter input to the 
NLP module. 
5 FLUENT in the ESL Classroom 
FLUENT is currently being evaluated in an English as a 
Second Language (ESL) environment. The Arlington 
Education and Employment Program is using FLUENT 
in its language laboratory. The students in this 
program are non-English speaking adults. The 
evaluation results reported in this paper are from two 
groups of students at level 2 of the program, with 
limited English and little or no exposure to computer- 
based systems. What exposure they have had has been 
with text-based systems mainly for vocabulary. 
FLUENT is their first exposure to an interactive 
conversational language learning system. The first 
group of 14 students generally worked in groups of two 
or three in a cooperative effort. In the second group, 
students worked individually. This time, students were 
so engrossed that they were allowed to work longer, so 
there was time for only six of them. The most 
difficult aspect of the interface for the students to master 
was the use of the mouse. However, the students who 
used the system a second or third time showed increased 
proficiency with the mouse. 
The evaluation procedure required each student to 
complete three lessons that had been prepared using the 
tutorial schema tool, with the advice of their teacher. 
This was followed by approximately 5 minutes in 
Movecaster mode to explore other facets of the 
microworld. All three lessons used the same plan, but 
progressively harder interaction types: first Tourguide, 
then Commander, and finally Quizmaster. In 
Commander mode, a student who did y when told to do 
x was told, in effect, "You did not do x. Undo y." In 
Quizmaster, each WH-question required a noun phrase. 
If the student's answer was not correct, the system 
displayed a dialog box with a list of possible answers. 
The student was then to click on the correct choice. If 
the student made another error at this point - something 
that happened only once - the system displayed the 
correct answer. 
The students were asked to complete the questionnaire 
shown in Figure 5. The results are shown in Figure 6. 
Ten of the 14 students in the first group and all of the 
second group responded. 
1. How well did you like it? It was 
VERY GOOD GOOD __POOR 
2. How good was it for learning? It was 
__VERY GOOD GOOD POOR 
3. Was it too hard? Was it too easy? 
__TO0 HARD GOOD TO0 EASY 
4. Was the voice OK? YES NO 
5. Do you want to say more? 
Figure 5. The questionnaire 
Ill ooo  2. 5 VERY GOOD 1 TOO HARD 8 YES 1 NO 5 GOOD 0 POOR 5 GOOD 0 POOR 9 GOOD 0 TOO EASY 1 <no response> 
Figure 6a. Questionnaire Results - Group 1 
lill 6 VERY GOOD 
5 VERY GOOD 
1 TOO HARD 
6 YES 0 NO 
0 GOOD 0 POOR 
1 GOOD 0 POOR 
4 GOOD 1 TOO EASY 
0 <no response> 
Figure 6b. Questionnaire Results - Group 2 
Observations made during the sessions point to some 
strong points. The verbal protocol shows that the 
students liked the positive responses in Commander 
mode. When the student successfully completes the 
action, the system responds with a positive comment 
such as "Great, you picked up the cup." or "Good, you 
turned on the faucet." The students also liked being 
able to pick the correct answer in Quizmaster mode. 
They often knew the answer but didn't know how to 
spell it. When they were able to choose the answer 
they often carefully noted the spelling and were able to 
spell the word correctly the next time. 
Observations made during the first session indicated 
some weak points. These led us to make changes 
which apparently led to an even better response in the 
second session. Several students were frustrated in the 
first session when the system was doing the action and 
generating language, because the sequence of actions 
was occurring too rapidly for them to understand the 
utterances. They would be concentrating on what was 
said and their train of thought would be broken when 
the sentence was overwritten with the text of the next 
action. This was corrected for the second group by 
providing a button that allowed the student to advance 
the system to the next action. 
Another annoyance for the student was when a 
response was given that was correct but not exactly 
133 
what the lesson required. For example, one action is to 
"turn on the faucet". When asked "What did I turn on?", 
several students responded "the water". Although not 
what the system was expecting (i.e. "the faucet") this 
response should be accepted. For the second session the 
system was able to accept various correct answers. 
The speech generated is a female voice using the text- 
to-speech capabilities of the Macintosh. A point of 
concern to us was how the students would react to this 
somewhat automated sounding voice. This decision to 
use text-to-speech was made after consulting with 
teachers of ESL They feel the quality is good enough 
for the level of students we are targeting. The student 
responses show that in general they were not distracted 
by the quality of the voice. They seem to be 
concentrating on the syntax and meaning of the sentence 
more than the pronunciation. 
6 Conclusions 
NLP has an important role to play in foreign language 
intelligent tutoring and learning environments. The 
generativity and variability that NLP brings to *,he 
conversational tutor FLUENT are crucial to its meeting 
its pedagogical goals. In particular, the layered design 
of the NLP, from teacher specification to actual text is 
key to allowing teacher control over the language input 
to the student, while at the same time producing 
language that is meaningful to the student in the 
evolving situation. 
7 Acknowledgements 
Our thanks to the Arlington Education and Employment 
Program, and especially to Daniel Norton for his 
support and suggestions. 

References 

Asher, A. 1977. Learning Another Language 
Through Actions: The Complete Teacher's 
Guidebook. Sky Oaks Productions, Los Gatos, 
California. 

Caldwell, D. and Korelsky, T. 1994 Bilingual 
Generation of Job Descriptions from Quasi-Conceptual Forms. 4th Conference on Applied 
Natural Language Processing, Stuttgart, Germany. 

Chanier, T. 1994. Special Issue on Language Learning: 
Editor's Introduction Journal of Artificial 
Intelligence in Education. Association for the 
Advancement of Computing in Education, 
Charlottesville, Virginia. 

Chanier, T., Pengelly, M., Twidale, M. and Self, J. 
1992 Conceptual Modelling in Error Analysis in 
Computer-Assisted Language Learning Systems. In 
M. L. Swartz and M. Yazdani, editors, Intelligent 
Tutoring Systems for Foreign Language Learning. 
Springer-Verlag, Berlin, Germany. 

Clancey, W. J. 1987. Knowledge-Based Tutoring 
The GUIDON Program. The MIT Press, Cambridge, 
Massachusetts. 

Felshin, S. 1995. The Athena Language Learning 
Project NLP System: A Multilingual System for 
Conversation-Based Language Learning. In V. M. 
Holland, J. D. Kaplan, M. R. Sams, editors, 
Intelligent Language Tutors: Theory Shaping 
Technology. Lawrence Erlbaum Associates, 
Mahwah, New Jersey. 

Hamburger, H. 1995. Tutorial Tools for Language 
Learning by Two-Medium Dialogue. In V. M. 
Holland, J. D. Kaplan, M. R. Sams, editors, 
Intelligent Language Tutors: Theory Shaping 
Technology. Lawrence Erlbaum and Associates, 
Mahwah, New Jersey. 

Holland, V.M. 1995. Introduction: The Case for 
Intelligent CALL. In V. M. Holland, J. D. Kaplan, 
M. R. Sams, editors, Intelligent Language Tutors: 
Theory Shaping Technology. Lawrence Erlbaum 
Associates, Mahwah, New Jersey. 

Krashen, S. and Terrell, T. 1983. The Natural 
Approach: Language Acquisition in the Classroom. 
Pergamon, Oxford, England. 

Schoelles, M. and Hamburger, H. 1996 Teacher-Usable Exercise Design Tools. In C. Frasson, G. 
Gauthier and A. Lesgold, editors, Intelligent Tutoring 
Systems: Proceedings of Third International 
Conference, Springer-Verlag, Berlin, Germany. 

Traum, D. R. and Allen, J. 1994 Discourse 
Obligations in Dialogue Processing. Proceedings of 
the 32nd Annual Meeting of the Association for 
Computational Linguistics, Las Cruces, New 
Mexico. 

Zock, M. 1996. Computational Linguistics and its 
Use in Real World: the Case of Computer Assisted- 
Language Learning. In COLING-96 Proceedings. 
Center for Sprogteknologi, Copenhagen, Denmark. 
