LANGUAGE LEARNING AS PROBLEM SOLVING 
Modelling logical aspects of Inductive learning 
to generate sentences in French by ma n and machine 
Michael ZOCK Gil FRANCOPOULO 
Abdellatif LAROUI 
LIMSI, BP 30, 91406 Orsay - France 
Abstract: 
We present here a system under development, the present goals of which 
are to assist (a) students in inductively learning a set of rules to generate 
sentences in French, and (b) psychologists in gathering data on natural language learning. 
Instead of claimin~ an all-encompassing model or theory, we prefer to 
elaborate a tool, which is general and flexible enough to permit the testing 
of various theories. By controlling parameters such as initml knowledge, the 
nature and order of the data, we can empirically determine how each parameter affects the efficiency of learning. Our ultimate goal is the 
modelling of human learning by machine. 
Learning is viewed as problem-solving, i.e. as the creation and reduction of 
a search-space, t~y integrating the student into the process, that is, by 
encouraging him to ask an expert (the system) certain kinds of questions like: can one say x ? how does one say x ? why does one say x ? we can 
ennance not only the efficiency of the learning, but also our understanding 
of the underlying processes. By having a tra.~e of the whole dialogue (what 
questions have been asked at, what time), Ave should be able to infer the student's learning strategies. 
I THE PROBLEM OF LEARNING A LANGUAGE: 
Language learning can be viewed as a special case of problem solving in 
which tlae learner tries to build and intelligently explore a hypothetical 
search space. If this view is correct, then two sets of questions arise immediately. On one hand one may want to know: 
a) what the nature of this search space is (what are the variables 9) b) how it is built (incremental learning: local vg global view), ' ' 
c) how it is explored (strategies: intelligent opportunistics vs systematic search). 
On the other hand, one may want to investigate how (i) the knowledge at 
the outset and (ii) the ordering of the data will affect the building and the 
searching of the space. Typically one does not learn from scratch, nor is it 
likely that one encounters either well-ordered data, or a Complete set of examples: natural learning is incremental. 
Obviously, these.facts imply that: 
* initial knowledge, in particular, knowledge of other languages may 
bias the kind of variables (attributes or hypotheses) constdered, i.e., included in the search space; 
* the order of the data (the examp es encountered by the student) may 
determine what rules are likely to be inferred at what moment, and finally 
rues are referred from mcomplete data (incremental learning). 
Furthermore, the same data may be characterized in different ways. 
That is, several equivalent descriptions may be inferred from the same 
data set. Whieli of these descriptions turns out to be the most adequate generally cannot be established until one knows the 
complete data set. Thus, rules may have to be revised in the light of new evidence. Consequently, errors are not only unavoidable parts of 
the learning process,but also an indispensable source of information for the learner. 
2 THE PROBLEM OF TEACHING HOW TO LEARNt 
As we have shown, learning can be seen as searching. Actually, teaching, as 
well as learning, can be conceived of as problem solving or reasoning in an informatio.n-exchange environment. There is a sender, a goal, a message 
and a recewer. The SENDER may be a native speaker, a teacher, a parent, 
a book or a computer. The GOAL is the task or performance (output). In 
our case it is knowledge of how to produce sentences in French. The 
MESSAGE is the input to the learning component: examples from which the rules have to be inferred (1). The RECEIVER or learner can be any 
system, naturm or artlttcial, capable of perceiving, memorizing and 
analyzing a set of data and drawing the necessary conclusions: a child, a student, or a computer program (2). 
Learning occurs in various settings. Depending on the order of the 
examples and the control of the information flow we s eak of nator I ..... . p a 
experimental, or msmuttonal settings. Natural learning is characterized by 
the absence of a clearly defined learning objective (3), by noisy and 
heterogeneous material~ and by unordered examples. The underlying 
regumr:ties are thus multiple, diffuse, and hard to perceive. Experimentffl 
learning and teaching, on the other hand, have a \]earning objective, the 
material is error-free, homogeneous and coherently ordered according to 
some point of view (learner or teacher). Whereas experimental learning can 
be characterized by the following sequence: (i) encountering the data (ii) 
analysis, (iii) building and testing of h~,pothesis, (iv) feedback and (v) proof 
or aemonstration of the theo~, traditional teaching goes througfi the 
following stages: (i) exposition, 0i) practice, (iii) testing and (iv) evaluation. This can be schematized as follows: 
Teacher: sets the task and presents the learning material Student: analyzes the data; 
Teacher: provtdes a set of examples; Student: practices; 
Teacher: asks questions to test the gained knowledge; Student: answers the questions; 
Teacher: evaluates the answers, provides feedback (explanations) 
and organizes future data as a function of actual performance 
Student: integrates the feedback into the knowledge base and cor- rects misconceptions; 
As one can see, the information flow here is entirely teacher-controlled. He 
is the one who sets the task, and provides the examples and the feedback. 
Consequently, the teacher decides the nature and the order of the material to be learned. 
There are two major shortcomings in this approach. Not knowiug what 
information is needed by the learner, the teacher may present the wrong 
data. More importantly, the student is only loosely integrated in the 
learning process. Instead of being active, generating and testing plausible hypotheses (discovery learning), he reacts to questions, Thus, it may happen 
that the student perceives his task as the learning of the material rather than the learning of the underlying principles. 
I~norance of what or how to learn may result in (i) learning the unintended 0:) poor problem-solving skills or (iii) little transfer. As long as the learne~" 
does not go beyond the informaBon given (the concrete word level), he 
cannot transfer the gained knowledge to similar situations, because the perception of similarity presupposes abstraction. 
Given these criticisms, it would be useful to have a system which has the 
qualities mentioned above without having the drawbacks. A good learning environment should be both flexible and constraining enough: 
* to allow for simulation of real Communication, that is to say, to 
provide a setting where both participants can take the initiative and control the information-flow, 
* O " " t ensure the learmng of the appropriate material (i.e., what to learn) 
as well as the necessary problem-solving skills (the methods, i.e, how to learn). 
A computerprogram could provide such an environment. It would offer 
different k!nds oi' !nformation (see below: trace-function), while answering 
me stuuent's questions as ne goes aJong generating and testing different sorts of hypotheses. 
3 THE COGNITIVE ENGENEER'S TASK: 
to provide the user a friendly interface 
We will describe here a system under development, whose major goals are: 
* O ' " " t provide an environment whtch allows communication between a 
learner (student) and an expert (in our case the system); 
* O " ' " ' t s~mulate the mformatton-processmg aspect of natural learning, i.e., 
the inductive learning of grammatical rules to generate sentences in French. 
* to allow teachers and psychologists to test various theories. 
806 
The system we have in mind is designed to help the student build the search 
space (the set of all attribute-value pairs). The learner has to discover how 
to explore it, By applying a given set of operators and by watching the 
outcome, he t~n test (i) which information is relevant and (it) to what it is 
relevant (to .,~yntax or morphology), ltowever, in this kind of dialogue 
(controlled trial and error) the system not only answers the questions asEed 
by the learner, but also assists him in determining what questions are meaningful in this context. 
Learning, be it by man or by machine, implies exchange of information between two s~,stems, for example, a native speaker (expert) and a foreigner 
(learner). We will start by describing some ortbe features our s'cstem needs 
to have in order to allow for such an information exchange. We will then 
give a detailed example, showing what such a dialogue between a human 
learner and the machine might look like. Finally we will discuss whether machines can acquire linguistm competency in a humanlike way. 
Before showing how the system is designed to work, let us specify more clearly what the learning olJjective is. 
4 'I~IE STUDENT'S LEARNING OBJECTIVE: 
Tile learner's task consists of incrementally learning the morpho-syntactic 
rules of personal pronouns in French. More precisely, the student is 
expected to acquire the necessary knowledge n order togenerate sentences 
composed of several pronouns (-see examples (a) - (i)). l"n order to achieve this goal, he has to learn: 
- how to express a given concept (morphemes), 
- how to linemize these concepts (sentence patterns) and 
- under what conditions (rules) to use each of these words or sentence forms. 
MORPHOLOGY Example of rules to 
determine MORPHOLOGY 
SPEAKER: je, me, mot - nous if SYNT.FUNCTION: direct object LISTENER: tn, te, tot, - vous PERSON: third 
ELSE: il, ella, ils, elias REFLEXIVE: no 
It, la, les, lui, leur QUANTITY: definite on, an, st, sol, eux NUMBER: singular 
GENDER: female 
then DIRECT OBJECT -- > la 
SYNTAX: 
a) S--I)O-'IO'-V ~e la lul prdnente I Introduce her to him 
b) S-IO-DO-V Je te la prdnento ~ tlltroduce liar to you 
c) ~DO-V-pp-~O \]eta pr6sente ~ ella tlltroduce you to her 
d) S-iO-V-pp'-~:O ~o lUt parietal do tel I will toll her about you 
a) V-DO-IO pr*sente-la mol Introduce her to me 
f) neg-DO-IO.-V-neg no la lul prdsonta paa Dontt introduce her to him 
g) neg-IO-DO'-V-llog no me la prdeente pan Dealt Introduce her to me 
h) Ileq-DO-V-nl~g.'pp--ro 1|o Ine prdflento pail b ella Doltlt h~troduce me tO her 
i t n~g-Io~v-lmg-pp-Io lie lut parle pa~ de mol Don't tell her about me 
S: subject, DO: direct object, IO: indirect object, pp: preposition, 
nag: negation, V: verb 
As one can see from tile data, pronoun<onstructions in French can be fairly complex (4). This complexity is due to: 
* the number of features necessary to determine word order or morphology: 
PART OF SPEECH: (noun, pronoun) 
ie parle ~ (noun) je klJ, parle (pronoun) 
SYNTACTIC FUNt.7"ION (subject , direct object, indirect object) 
il 6erit h Pierre ~subject) 
Paul llli derit (redirect object) 
SENTENCE-TYPE: (declarative, interrogative, command) 
tu m~ le donnes? (interrogative) donne-It ~ ! (command) 
NEGATION: (,yes, no) 
oonnes-le alfi! (positive) 
ne ~ le donnes pasI (.negative) 
COMMUNICATIVE-ROLES: (I, you, tie) 
je te LE donne (IO = je LE llli donne (IO = ~o~) 
NUMBER: (singular, plural, indefinite) 
je te 1~ garde (singular) je te Its garde (plural) 
je t'gn garde (indefinite) 
GENDER: (male, female) 
je lg vois (male) je la vois 
(female) 
VERB CONSTRUCTION: {type of complement (DO vs IO), 
pJpe at preposition, reflexivity) 
je vois_Mane --> je la vois (direct object) 
je parle ~t Marie --> je 1~ parle (indirect object) 
SEMANTIC FEATURES: (animate, inanimate) 
il m'emm6ne ~ ~ --> il m' ~ emm6ne 
il me pr6sente ~t sa ~ --> il me pr6sente ~t 
* the structure of these features: if one compares (a) and (e), one will notice that the form of the indirect object (lui vs ella) depends on the 
value of the direct object (horizontal dependancy); 
* the inte~ependance of syntax and morphology: practically all variables, 
except NUMBER and GENDER are relevant both for ~ntax and 
morphology. Furthermore, the position of the direct object pronoun may depend on the value of the indirect object (compare (a) and (b) 
here above). In other words, changes in morphology often imply changes in syntactic structure. 
* the various knowledge sources: the determination of morphology and 
syntax requires information about the Le.fgLe.~ (number, gender, 
animacy).text fimctions (syntactic status of noun-phrase: noun vs 
pronoun, topicalisation, person), ~ (positive/negative), &tLe~:c_hh = 
aC.t,(st atement/question/corn m and), verb-construclio~ (type of 
complement: direct\]indirect, type of preposition: il, de), etc. 
Given these intricacies it is easy to understand why students so often fail to 
learn these rules. Modelling their learning is thus a challenging task. 
5 HOW CAN THE LEARNER BE INTEGRATED INTO THE PROCESS ? 
If one accepts this view of learning, then the problem of the student is to 
find out how to build and how to intelligently reduce the search space. The 
system will help the student in various ways. 
First of all, it will answer certain kinds of questions: 
/~ How does one say x ? 
What would hapl~en if...?, 
Can one say x ?, How should one say x ? 
Why does one say x ? 
All these questions occur in some form or another in natural settings. The following examples may illustrate these strategies or testing modes: 
(a) Question: ~~: "je lui pense" 
Answer : ie pense i~ ella 
e pease ,~ lui 
ele pense (5) 
(b) Question: ~~12tled~,if in the following sentence: 
Paulparle gl Made (Paul talks to Ma~ the object-noun was pronominalized . 
Answer: Paul lui parle 
(c) Question: .~: "je he pense"? 
Answer: no 
(d) Question: .ling "je lui pense", ~? 
Answer: je pense h ella 
le pense ?a lui je le pense (5) 
(e) Question: Why does one ~SKy: "Je le pense" 
Answer: explanation given by the system 
These strategies are complementary in that they correspond to different 
learning needs. They provide different kinds of feedback. The first two methods (the inductwe approach) seem useful if one does not have much 
knowledge yet. Tile third one allows to test the degree of generality or the 
extension of a given rule (deductive reasoning), the fourth method provides 
additional information in case of incorrect performance, while the last question may either confirm a hypothesis, or correct a misconception. 
Second, the system should show how to reach the solution (the demonstrative mode). This might be helpful if the student gets stranded, 
not knowing what to do. In this case the system takes over, showing how 
information may be processed. By watching the system, the student may learn how to explore, pc., how to generate and test a set of hypotheses. 
Third, the system keeps a record of the whole dialogue. Such a trace has 
many advantages: it allows the student to verify to explaiu and to 
remember. He may thus (i) check the consisten~ of the rules, (it) justify a 
given conclusion in the light of evidence and (iii) reorganize his knowledge 
base. This last possibility should enhance fis perception of underlying regularities. 
Psychologists could use this trace to infer the student's learning strategies. 
The rules a student Ires been testing at a given moment may be inferred on 
the basis of the nature and order of the questions being asked. 
007 
Finally, teachers could use the trace-function to gain feedback concerning 
the order of presentation of the data. By varyingthe nature and order ot 
information, they can determine experimentally t~e complexity of the data 
(examples, rules),-and thereby the relative efficiency of various teaching- 
strategLes. 
6 THE FUNCTIONING OF THE SYSTEM: 
The program works interactively. The user is given a set of options fi'om 
which he has to choose. The system converts this input into the adequate 
output, i.e., linguistic form. Input are meanings (what to say), output are 
sentences (how to say it). 
The process is started with a list of nouns and verbs. This list is a kind of 
knowledge base "i.e., a set of facts a potential user may want to talk about. 
This base is limited in scale, and arbitrary, in that it is given by the s},stem. 
However, this limitation is easily overcome. The base can be extenued by 
tile user at any moment. The important point is that, by feeding nouns and 
verbs into the knowledge base and by choosing among these entities, the student signals what he wants to say. In doing so, he builds propositions of 
various complexity (one-, two-, or three place predicates). 
The system will operate on these structures and build simple declarative 
sentences. In other words at this stage of interaction it is assumed that the 
student wants to know how the intended meaning translates into this 
canonical form. For example, the input (a) would yield the output (b). 
STUDENT SYSTEM 
input:(a) output:(b) 
regarder (Manuel, Christine) = = > Manuel regarde Christine watch (Manuel, Christine) = = > Manuelwatches Christine 
The student is queried a~ain to determine what he wants to say. Basically 
he has two possLbitities. Either he tries a complete new idea (proposition), or he modifies part of thepreceding one. In this latter case, tile system 
provides a list of options (-attribute-value pairs), inviting the student to 
discover what happens, i.e. how morphology and/or syntax are affected, as 
he changes the value of any of the attributes such as PART OF SPEECH, 
SENTENCE MODE, NEGATION, and so forth. Let us assume that the 
student had chosen to replace respectively Manuel and Christine by a 
pronoun. In this case the system would produce the following sentences: 
I1 regarde Christine 
Manuel la regarde 
How shoukl one say ? Why does (:no say ? 
\] / 
,-6 
~ explanation ~ il me le donne 
XPERT USER 
FIGURE 1 
By comparing these sentences with the base form, the student should notice 
certain differences and draw the necessary conclusions. For example, given 
the data he may conclude that: 
RI: if the direct object is pronominalized, . 
then it moves in front of the verb (syntax). 
R2: case (syntactic function) is morphologically relevant: 
if the subject is pronominalized then its surface form is "ir', 
113: if the direct object is pronominalized then its surface form is "la". 
Control is returned to the user. Actually, from now on we are in a loop, w th the dialogue having basically the same form. However, in each cycle 
the hypothesis to be tested is likely to be different and it is interesting to watch how a student proceeds in acquiring competency. What does he want 
to know 9 Is he systematic? What kind of strategy does he use (breadth first, depth first etc.)? Under what conditions does he change his m~thod? etc. 
The learner's problem is three-fold, he must find out: 
* which parameters (attributes) are relevant, 
* to what linguistic component they are relevant (syntax and/or 
morphology), and 
* to what extent they are relevant (6). 
A student may thus want to know: 
: whether the variable GENDER is morphologica!ly relevant!, 
whether this is the only relevant varmme, or It otlaer varmmes 
come into play' * whether it is relevant for all cases, irrespective of, for example, 
communicative role, negation or sentence mode (compare (e) and (g)); 
It should be noted, that every time the student is given control, he can 
choose two things: (i) the kindof information he wants to convey (what to 
say), and (ii) the dialogue-mode, i.e., HOW DOES ONE SAY?, CAN ONE 
SA~9 ete ) The following diagram illustrates the information flow. 
insert figure 1 here 
This kind of environment has three basic functions: 
b to answer different kinds of questions, to convert meaning into form and 
to help the student to discover how changes in meaning are reflected 
in changes in form. 
It should be noted that the student has most of the control. The following 
examples should give an idea of the dialogue. These hypothetical dialogues 
serve illustrative purposes. However, we believe that they are reasonably 
close to what might be encountered in an experimental session. 
6.1 EXAMPLE DIALOGUE NUMBER 1: 
Tile student's question (dialogue mode) is: HOW DOES ONE SAY? The 
figure below contains three columns which express respectively the 
student's intentions, i.e. what he wants to say, his observations, and his 
conclusions with respect to syntax and morphology. 
insert figure 2 here 
Having generated the following proposition: 
Max, voir/Max ' Paul) 
(see Paul)) 
he wants to know what would happen, if both arguments (Max, Paul) were 
pronominalized. The system generates the following answer: 
(1) il le volt 
The student analyzes tills sentence and draws as conclusions Rule 1 and 
Rule 2, mentionned here above. He goes then on to ask 2dl~IJY~l~\].~%~ 
if PAUL was replaced by MARY. The system answers: 
(2) il la voit 
The student concludes that GENDER is not relevant with regard to word 
order, but is a necessary condition to determine morphology (Rule 3). This 
latter kind of knowledge could be expressed as: 
R3: if PART OF SPEECH: pronoun & SYNTACTIC FUNGI'ION: direct object 
& GENDER: female 
then PRONOUN: la 
else if GENDER: male 
then PRONOUN: le 
In the next question he is concerned with the relevancy of NUMBER. He 
asks: what would happen if the direct object were CHILDREN (les 
enfants)? The system's answer 
(3) il les volt 
allows him to conclude that NUMBER is relevant for morphology but not 
for syntax, as there are no changes in word order, but there is a change in 
form. This fact is encoded in the following rule: 
808 
(INPUT) (~UTPUT) 
HOW DOES ONE SAY OBSERVATION CO~CLUSmN 
I) volr (Hox,Paur) 
Ha~ ~ l~ronuu. 
Paul = preheat ~> llJ~voit 
DO precedou V 
pl'o:lou:l ~ I0 
Ha~i~ = pronoun ~> il ~Ovolt 
posltlol) of 00 
cnnalat.~Itb R1 
change i. fot~ 
~) voir Ill,x. e~\]Fonts ) 
eat'ants = pronoun 
~dNBER = plural-> ii los veil 
rio ehor~tJo in 
posi finn 
change in form 
A) ~ (Hax, Mari~ ) 
Hox : pr o\[iotp~ 
Herin : pronoua -> II hll parle 
I0 preoedes V 
see R1 
SYNTACTIC CATEGORY o.d the SYNTACTIC 
FUNCTION of the rororont are ayntoc-. 
fleetly relevant 
RII If SYNTACTIC CATEGIJIIVI pronoun 
tf SYNTACTIC FUNCTION: dlr. obj. 
th~n pronoun In frer~k of the verb 
Subject ~ Direct O~oct - Verb 
SYNTACTIC F~CTtON io l~rphologlcally 
~BI~vB~tl 
R2: If SYNTACTIC FUNCTION: allPJ~t 
then PRONOUN~ ll 
If SYNTACTIC FUNCIIGN:(Uro obj. 
then PRONOUNI le 
S~i_~,±x_, 
GENDER Is syntactically not rel,lva,,t 
ogmmholo!gS~ 
GENDE8 is morphologically relevant 
R): 'f' SYNTACTIC CATEGORY: pro.sun 
tf SYN'IACIIG FUNCTION: dlr. obj. 
If GENDER 1 femal~ 
then PRONOUNt la 
if GENDER: mole 
than PRONOUN: le 
5y~t nx: 
N~NBER Is ~yntoctleally not relevant 
H_9.zo r~L~a.~ 
OAf if 5YNTAUTC FUNCTION: dlr. obj. 
if NUHUER: plural 
if GENDER: male (~) 
then PRONOUN: lab 
S~tn~£ 
Syntactle Tuner±o) (cane) la releva, t 
rsl if SYNTACTIC CAIEGORY: pronoun 
If SYNTACTIC FUNCTION: Ind.obj. 
then: Subject - Indlr.Ohj~et - Verb 
Generaltaation of RE & R5 = 116: 
R61 if ~n object is pronoainalized 
then: Subject - Object "- Verb 
ohange in form ~lnr~rl~hnl.~ 
SYNTACTIC FUNCTION of 10 morphological,- 
Iv relevant (sea R2) 
R71 If SYNTACTIC FUNCTIONI indil'.obj. 
if GENDER: f,~lale (**) 
Ellen PRONOUN: lui 
5) parlor (Max, \[au! ) 
Paul = pronoun -> ll ~ parle 
conalatent with 5_yntax: 
preceding rules GENDER is synt~ctically not relevant 
nn change in Mo~ 
morphology The gender af the I0 ~s morphologically 
noL relevant, consequently minx tile 
GENDER conatralnt of 117 
Correction of flY: 
Rill if SYNtACtiC FUNOIIDNI lndlr.obj. 
th=m PRONOUN* iul 
(*) since GENDER wan relevant for the singular th~ 8tudont a~SL~m~ that It 
it la oleo r.J.vant for, the plural 
(~*) 'Jlnc*~ GENDER wan ral,vant for tim DO (R3) Lho atud.,t aa~lum~l\] 
that it to a!tlo relevant far the IO 
FIGURE 2 
R4: if SYNTACTIC FUNCTION: direct object 
& GENDER: male 
& NUMBER: plural 
then PRONOUN: tes 
it is intm'esting to notice, that this rule is too specific, because GENDER is not a necessary condition, However, this conclusion is perfectly reasmmble 
given the data encountered so far. GENDER was a necessary condition for 
singular (see rule 3), and since then there has been no evidence to the 
contrary. Consequently, the student has no way to conclude from the data, that for direct objects GENDER isgenerally relevant only for the singular. 
(The onl~¢ reason we could think of that a student might consider this last 
hypothesis, would come from his knowledge of another language which has 
the very same property.) 
It is also noteworthy that for objects, GENDER is only relevant for the 
SINGUIAR. This has procedural implications; namely that NUMBER 
should be processed prior to GENDER. The former being nTore 
informative than the latter. 
In the following cycle (sentence 4) the student changes tile proposition 
altogether, asking the system how one would say: 
parler (Max, Patti) 
when both arguments are pronominalized. This would yield the following 
senteoce: 
(4) II lui parle 
From that he may conclude that the indirect objectprecedes the verb 
(Rule 5). Reco~aizing the similarity with rule 1, Le., rccognizi,lg the fact 
that the syntactic status of the object (direct vs indirect) does not affect word order, he may generalize these two rules and replace them by rnle 6: 
R6: if an object is pronominalized, it precedes the verb 
This rule is more general titan the former ones, in that the distinction 
between direct and indirect object has beeu dropped. It should be noted, however, that this rule, even though correct in the light of evidence, i.e., 
data encountered so far, is too general. For example, it does not apply to 
seutences composed of two objects (three place predicates), hi other wo'rds 
this rule needs refinement, .e., addit onal constraints. 
With respect to morphology, the student concludes that the attribute CASE 
(syntactic function) ts relevant, which yields the following rule: 
R7: if SYNTACI'IC FUNCTION: indirect object 
& GENDER: female 
then PRONOUN: lui 
Again, the morpheme is overspecified, because GENDER is not a 
necessary condition. Having noticed that GENDER was relevant for direct 
obiects (-rule 3) the student has overgeneralized, assuming that it was also 
relevant for the indirect object. It is noteworthy, however, that this 
particular overgeneralization does not produce incorrect results. 
Finally, the student asks the system to replace MARY b~/PAUL. Getting 
the same answer as in 4 he concludes that for indirect objects the 
GENDER is irrelevant for syntax as well as for morphology. Consequently, he relaxes the gender-constraint of rnlc 7. Once again, this conclusion is 
valid only with respect to the set of examples he saw. 
6.2 EXAMPLE DIALOGUE 2: 
This time the dialogue-mode is CAN ONE SAY. The three colmnns 
correspond to the student's questions, his hypotheses, and his conclusions. 
The controlled variable (a change of attribute or a change of its value) is 
underlined. 
insert figure 3 here 
The figure being rather self explauatory, We will make only some short 
comments. At stage 3 the student wants to know whether the 
comnmnicative role of the indirect object, the attribute PERSON, is 
syntactically relevant. From the data he has seen, he conch, des that this was 
not the case. However, this conclnsion, even though correct with respect to 
the data, has to be revised in the light of new evidence (next sentence, i.e., 
sentence 4). 
It is interesting to note, that the student would probably never have drawn 
tbis conclusion if sentence 4 had preceded sentence 3. hT other words, he 
would have noticed the relevancy of the attribute PERSON right away, and 
never have drawn conclusion 5. 
l le LUI donne 
l TE le donne 
This shows how the order of the data is a critical variable determining the 
efficiency of rule-inference, i.e., what conclusions are drawn at what 
moment. 
7 CAN MACHINES ACQUIRE LINGUISTIC COMPETENCY IN A 
'HUMAN" WAY ? 
Actually there are three questions: 
- Can machines learn? 
- Can they learn in an intelligent or "humaW' way? 
- What kind of knowledge would a computer program need to have in order 
to learn the rules I have been talking about? 
The answer to the first question is clearly yes (see Michalski, Carbonell & 
MitcheU 1983). The ,latter two questions are more controversial. Let us 
begin with the last one. 
Inductive learning basically consists of drawing conclusions li'om the 
similarities and differences of abstract data descriptions (contrastive 
analysis). The crucial points are thus data description and analysis; 
- in what terms should we characterize the data? 
-what additional kiud of knowledge is needed to infer the rules ? 
13 \[)9 
CAN ONE SAY ? HYPOTHE51S: CONCLUSION: 
1 il male donne . ql: Do both objects 
yes precede the verb ? 
q2: Which one precedes 
the other ? 
Is II me LAdonne Q\]: is the GENDER of the 
yea direct object syntacti- 
cally relevant? 
2a ~I me~r~_do nne Q~: la the NUMBER of the 
yes direct object syntacti- 
cally relevant? 
3 IIT\[ le donne QS| la PERSON of the 
yes indirect object syntac- 
tically relevant ? 
4 il LUL le donne 
no: ille lui donne 
5 ille S E garde 
no= il ae le garde 
6 il S' ~ mOdUS 
yes 
7 il enLUIdonne Q6t if the DO : en & 
no: il lui en donne if the IO = lui, 
Wbich one of them 
precedes the other ? 
8 il m~' an donne 
yea 
Answer to Oft 
yeet both objects precede 
the verb 
Annwer to Q2: 
the indirect object pre- 
cedes th~ direct oriel 
s-Ig-OO-V 
Answer to Q): 
the GENDER is syntnctical- 
ly not relevant 
Answer to Q4: 
the variable NUMOER ia not 
relevant for syntax 
Answer to Q5: 
no, PERSON is not relevant 
for syntax 
Correctian and refinement 
of Conclusions 2 and 5: 
the variable PERSON ia syn- 
tactlcally relevant= 
Concluoion 6: 
if PERSON-IO: 3d 
than= S-DO-IO-V 
Conclusion 7: 
if PERSON-IO: fat or 2nd 
them S-\[O-DO-V 
refinement of conclusion 6 
Canclusion 8: 
if PERSON-IO: 3d 
if V-CONSTRUCT.= reflex. 
then= S-\[O-OO-V 
confirms conclusion 8 
conclusion 9= 
if NUHBER~DO: indefinite 
then= S-IO-DO-V 
Answer to q6: 
the indirect object, conse- 
quently conclusion 6 has to 
be Pefined. 
Conclusion lO= 
if PERSON-IO: 3d 
if NUHBER-DO: indefinite 
then: 5-IO-DO-V 
confirms conclusions 7 & 9 
with regard to the examples 
given in 7 and 9 we may re- 
lax the, PERSON-canstraint 
of conclusion 10 
FIGURE 3 
Obviously, a system capable of performing the kind of learning we have 
been talking about would have to be able to parse the sentences; that is, it 
would have to produce as output an adequate description of the input 
sentences described above. 
This raises a terminological problem. Data can be described in various ways. Different descriptions can be functionally equivalent (7). Clearly, the chorea 
of metalinguistm terminology differs depending on whether the goal is 
machine learning or modelling "human'rlearnmg. In the first case, the 
problem is descriptive adequacy, whereas in the second case we deal with an additional constraint, that ot the universal status of the terminology. Do 
all humans, irrespective of culture and education, use the kind of terms 
linguists use to analyze sentences ? Is there a universally shared subset of 
metalinguistic vocabulary ? In the absence of answers to these empirical 
~tuestions we will stick with the terminology currently used in computational 
hnguistics. 
• A differeut, but related problem is the guestion of how a system may be enabled to draw conclusions from a set of data (infering general rules). 
As we have said above, generalizations are made on the basis of contrastive analysis. In order to allow for such generalizations, the learning component 
needs a hierarchically structured metalanguage, that is, a vocabulary whose 
low level concepts (primitives) are subsumed by more highly ordered, 
abstract forms of knowledge. For example: 
masculine & feminine = = > GENDER; 
singular & plural = = > NUMBER; 
subject, direct object = = > CASE 
We will now turn to the question of whether computers can learn in an intelligent or "human" way ? Obviously this question raises the problem of 
what intelligence is. Instead of answering this question, we will focus on two 
aspects of intelligent learning, namely economy and flexibility of methods. 
Exhaustive search is neither natural nor economical. Since memory is 
associative, we find it hard to be consistently systematic. Like gamblers, we 
tend to use search methods which are more or less risky. 
People (learners) generally have a set of methods and a separate 
component (critique) for evaluating these strategies with respect to their 
relative efficiency. As different proSlems require different problem-solving 
methods, it is very unlikely that there is a umque, universal problem-solving 
method. People tend to be opportunistic in their approach rather than 
systematic or scientific. Both tile nature of strategies and the depth of processing will vary with the needs of the learner. Corrolarily, it is equally 
unlikely that one fmds the optimal method immediately, since one operates 
on incomplete data. Inductive learning is typically incremental. Hence 
methods have to be adapted or gradually refined in the light of new 
evidence. 
Intelligent learning is thus intimately linked to strategic knowledge (8) and 
to (more or less) general information-processing principles. These principles may be expressed in terms of simplidty, informativeness, 
generality, andso forth. 
For example, the notion of simplicity may be used to choose among 
different options. In fact, a learner could hypothesize that two-place 
predicates (to see) are simpler to process than three-place predicates (to 
give). 
The notion of information is related to efficiency. It can be used to reduce 
the search space. This claim is substantiated by the fact that rules governing 
morphology of first and second persons (I, you) are generally learned faster 
than those which determine the form of the third person (he). 
In conclusion, we believe that, in principle, certain aspects of intelligent 
learning could be modelled by a computer. However, before trying to model human learning, it may be worthwhile to start gathering data on ~aow 
humans learn. This is precisely one of our goals. By watching people asthey 
use this tool, i.e. by keeping a trace of the dialogue, one should be able to 
infer the strategies they use. 
8 CONCLUSION : 
We have described a system under development that is meant to be a tool 
for theory builders (cognitive psychologists), application designers 
(language teachers) and end users (students). The system is meant to assist psychologists teachers and students in their respective tasks: model 
learning, optimize teaching and learning strategies. 
The emphasis in this paper has been on learning rather than teaching. For 
the time being the task of learning is to be performed by a human, however, 
in principle it is possible to extend the system so as to allow for automatic 
learning, the ultimate goal being to model human-like behavior. 
Computers, with their large, indelible memories, are powerful tools. They 
allow us to control virtually any number of parameters. Consequently, one 
¢'an trace a reasoning process or test a given theory, i.e., determine ",~mpirically how different variables affect the efficiency of learning, and so 
forth. 
This has an interesting consequence with respect to theoretical 
commitments. Instead of claiming an all-encompassing model or theory, one can write a program general and flexible enough to permit the testing 
of various theories. That is what we are trying to do. 
Watching how people use the tool, we may gain insights about the way 
humans learn (strategies), and thus eventually move from artificial to natural intelligence. 
NOTES : 
(1) This message has to be interpreted. Thus the learning task is not 
the surface form of the message, i.e., words and sentences, but the 
underlying principles (abstractions: rules and sentence patterns) allowing tqaetr generation. While some forms (e.g., words) have to be 
learneR, they generally serve for illustrative purposes. Rote learning of 
the entire set of surface forms (words and word combinations) is not 
only inefficient, but in fact impossible, because of time constraints: 
there are more possible combinations than we have time to learn. 
Learning is thus more than a quantitative change of performance 
(speed, number of errors). It generally implies a restructuring of the knowledge base. 
(2) It should be noted, however, that we are not dealing here with 
children learning a first language. Instead we would like to model some aspects of the sctentific-minded foreign language learner. 
(3) One may object that there is a global goal, namely learning the 
language. However, it seems to me that the primary goal is communication rather than attaining a local objective like, let us say, 
learning the pronoun system in French. 
(4) For a more detailed discussion, in particular with respect to the procedural implications, see Zock, et al. (1986). 
810 
5) In art ambiguous situation the system will either produce all cases see here above), or ask for clarification. For example: 
Student: How does one say: il te me pr6sente 
System: This depends on what you want to say. 
Do you mean (a) or (bl? 
(a) il te pr6sente ~ moi " " 
(b) il me pr6sente .~ toi 
(6) This last problem, which consists in finding the right degree of 
generality (underspecification vs overgeneralization), is partteularly 
delicate in that conclusions have to be reached on the basis of incomplete data (incremental learning). 
(7) This fact is illustrated by the variety of parsers. Parsers analyze 
sentences and assign them descriptions on various levels such as: part 
of speech, syntacti~ function, case-roles and so forth. For a review of 
the state of the art See King (1983) or Winograd (1983). For a French 
parser see Francopoulo (1986). 
(8) These strategies could either be part of the system, in which case 
they must be explicit (one needs a model), or they could be part of the 
learning process, in which case the system learns not only domain 
specific knowledge but also methods of how to learn (metaknowiedge). 

REFERENCES

Francopoulo,G. 1986 Machine Learning as a Tool for building a Deterministic Parser 
in: Rollinger C. & Horn,W, Second Austrian Congress on 
Artificial Intelligence, GWAI 86, Springer Verlag 

King,M. 1983 Parsing Natural Languages, Academic Press, New York, 1983 

Michalski,R., Carbonell,J. & MitchelI,T. 
1983 Machine Learning: An Artificial Intelligence Approach Tioga 
Pub ishing Company, Palo Alto, Cal fornia 

Winograd,T. 
1983 Language as a Cognitive Process, Addison-Wesley, Reading, 
Mass. 

Zock,M., G.Sabah, C.Alviset 
1986 From Structure to Process: Computer Assisted Teaching of 
various Strategies for Generating Pronoun-Constructions in 
French, in: Proceedings of the llth COLING, Bonn 
