A COOPERATIVE YES-NO QUERY SYSTEM FEATURING DISCOURSE PARTICLES 
Kjell Johan S/EB~ 
Seminar fiJr nattJrllch-sprochllche Systeme 
T iJblngen* 
and 
Norges allmennvltenskopellge forsknlngsr~d 
Abstract 
Cooperative dialog systems will offer extended answers 
to questions, that is, they will volunteer information 
not explicitly asked for. A complete response will be 
complex and the member sentences will evince an ex- 
tensive parallel, the indirect answer substituting an 
alternative for a focus in the question. Research on 
discourse particles has shown that they are necessary 
to ensure coherence between adjacent sentences evin- 
cing an extensive parallel, that is, that they reflect 
discourse relations as given in complex answers, so 
that such answers emerge as core contexts. Thus the 
proper mode of representation for discourse particles 
in a system coincides with the framework of coopera- 
tive question-answering. The PASSAT system centers 
on the r61e of particles in characterizing and reflec- 
ting relations such as underlie complex response. 
Discourse Particle Semantics 
Discourse particles are meta expressions in a natural 
language: They express discourse relations, which are 
necessary properties of (complex) discourses, that is, 
they refer to things in the language and not to things 
in the world. It follows that they do not influence 
truth conditions. Instead, they affect coherence: 
When occurring in a complex discourse such that 
the discourse relation "contradicts" its meaning, 
a discourse particle may cause incoherence. 
?Edison Invented the telegraph. 
Marconi also invented the phonograph. 
Conversely, discourse particles can cause coherence 
too: When one does not occur in a complex discourse 
such that the discourse relation "entails" its meaning, 
an incoherence comes about which the occurrence 
would prevent. What causes that incoherence is the 
occurrence of the empty, or zero, particle. 
So nonempty particles are sometimes necessary to 
reflect discourse relations, in other words, they 
substitute positive for negative presuppositions. 
?Edison Invented the telegraph. 
He invented the phonograph. 
And, those presuppositions do not refer to the world 
but to the environment of discourse. On a classical 
truth presupposition, enriched by a sensitivity to 
context, the empty particle in the example presupposes 
that Edison did not invent the telegraph. But then, he 
in fact did not, so the incoherence is not explained. 
Similarly, on classical terms the particle in the former 
example presupposes that Marconi invented the 
telegraph, but then, he in fact did. 
* The paper Is based on research done In the project 
L/LOG, financed and supervised by IBM Germany. 
So discourse particles, empty or not, react not to 
what is or is not the case but to what is or is =lot 
supposed to be the case: To the context. The proper 
context category for the discourse particle rategory 
is a complex sentence: A sentence pair. Any two 
sentences in sequence occasion the empty or some 
nonempty discourse particle in front of or withirl the 
second sentence. Thus a language in a model theory 
of discourse particles will consist in a pair: 
L=<SP, DP> 
SP is a set of sentence pairs sp, the individual con ~ 
stants, and DP is a set of discourse particles dp in- 
cluding the empty particle dpO, the predicates; and 
for any dp and sp, dp(sp), the application of dp on 
sp, is to represent the occurrence of dp in sp. An 
interpretation rule is to state a necessary condition 
for the coherence of any dp(sp) in terms of tile 
meaning of sp and the meaning of dp. Thus the 
model in the theory will consist in a triple: 
M = < DR , S , h > 
DR is a set of discourse relations, # is ,~n assignm~n¢; 
mapping constants, i.e. sentence pair~, onto member':, 
of DR, and I~ is an interpretation mapping predicate;:, 
i.e. discourse particles, onto subsets o~ DR. The, d~,- 
notation of any element of DP dp is defined as Lhe s~ 
of discourse relations b(dp). The interpre'~tion rtd.,:. 
states that for any sp, dp, dp(sp) is only coherent 
if S(sp) is not an element of b(dp') for any dr' 
different from dp. 
It is assumed that the denotations of the discou~'.,;e 
particles - the elements of the picture of DP under' 
h - are all disjunct, i.e. that for any dpl, dp2. in 
DP, the intersection of P(dpl) and P(dp2) is empty. 
Thus for any sp and dp such that ~e(sp) is in P(dp), 
dp'(sp) is incoherent for any dp' different from dp~ 
This means that a sentence pair instantiaUng a di~ r 
course relation that belongs to the denotation of a 
particle is only coherent if occasioning the parUcle, 
in other words, that a particle is necessary wi~h 
respect to every discourse relation in its denota-~ion. 
As DP contains the empty particle dpO, this principle 
corresponds to a negative presupposition. 
So there are discourses necessitating this particle 
or other in virtue of the relations they instantiate. 
A discourse relation is an abstraction on the v~ay the 
two members of the sentence pair compare to each 
other. Any two sentences can be segmented into a 
COl~rt, a depart, and a report, meaning the portion 
common to both sentences, the distinct portion of ~tt~ 
first sentence, and the distinct portion of tile second 
sentence. The common denominator of discourse 
relations in the denotation of any nonempty particle i~ 
that the de- and the report are minimal: That the two 
sentences differ in only one description. Discourses 
necessitating nonempty particles are characterized 
by an extensive parallel. Importantly, polarity 
549 
change does not count as a de- or a repart but 
serves as an extra parameter to differentiate 
specific discourse relation classes. 
Marconi Invented the telegraph, ?(but) he didn't Invent the phonograph. 
Discourse relations encode information on two other 
parameters: Whether the de- and the repart are in a 
semantic relation (scalarity or hyponymy) such that 
one sentence entails the other (negation exempt), and, 
if so, whether entailment is left-to-right or reversely, 
and if not, whether they are in a semantic relation 
(exclusion) such that one sentence entails the negation 
of the other, or the two sentences are compatible. 
These four features - polarity distribution accounts 
for two - give rise to sixteen cases, thirteen 
providing slots for German discourse particles: 
/ 
<C,D> 
/ 
÷ 
\ 
+<C,R> sogor/ > </ und 
zwar 
-<C,R> abet 
+<C,R> also 
-<C, R> contradiction 
/ *<C,R> contradiction 
v^ < v ~< -<C,R> also 
*<C,R> auch 
^ 
-<C,R> abet 
~.+<C,R> ober 
~-<C,R> ouch 
V^ +<C,R> sondern 
-<C,R> ouch 
\ > / +<C,R> contradiction >< / ~ -<C,R> erst recht 
\ +<C,R> aber 
-<C, R> elnmol/ Oberhaupt 
C = copart, D = depart, R =repart; 
*/- = positive/negative polarity; 
>< = depart and repart are ordered: 
> = R>D (*<C,R> entails *<C,D>), 
< = R<D (÷<C,D> entails +<C,R>); 
v^ = depart and repart are not ordered: 
v = D and R exclude each other 
(*<C,D> entails -<C,R>), 
^ = D and R do not exclude each other 
(*<C,D> and *<C,R> are compatible). 
Complex Response: The Basic DP Situation 
It is desirable to equip a computer system to execute 
the semantic theory sketched above. There are several 
possible approaches to this, but one embodies decisive 
advantages. That is a particular generative approach. 
It is a cornerstone of the theory that discourse 
particles not only introduce requirements on contexts 
but also have a communicative necessity; that contexts 
require them in that the nonoccurrence can be as 
damaging to coherence as can the occurrence. In 
representing the model in an automatic process it 
is especially desirable to capture this aspect. 
One way to go is to have a program test surface- 
language inputs and give notices of error whenever 
incoherence occurs. On this course, every piece of 
information is produced by the user, who must be 
acquainted with the theory in order to obtain an 
interesting reaction. Or one can have a program 
generate particles in accordance with the syntax 
and semantics of the contexts it generates. 
Again, there are more than one way to go: The 
input can be of a general nature, e.g. a description 
of a discourse relation, or it can be specific, e.g. 
a pair of discourse representation structures. Either 
way, the user has to specify the context unit to be 
generated, a complex discourse with a parallel, 
as long as she addresses it directly at some level. 
However, once such contexts are embedded in a 
communicative setting to motivate them independently, 
there is no need to prescribe anything. There is 
another, Indirect generative approach which promises 
spontaneous and systematic generation of proper 
contexts, and an interesting application: Dialog 
systems capable of cooperative question answering. 
This is an independently motivated field of research 
in Artificial Intelligence and in Computational 
Linguistics, seeking to simulate that crucial feature 
of human dialog behavior that answers are far from 
always formed in strict accordance with the semantic 
structure of the question. Frequently in actual 
conversation, answerers are expected to elaborate, in 
particular on a yes or a no. A realistic yes-no query 
system will be prepared to offer additional information 
in the form of extended answers, and several systems 
in this spirit have recently been devised. 
As it happens, complex responses are key contexts 
for discourse particles. A complete response consists 
in a sentence sequence. This means that the proper 
context category for particles is given a priori in 
this framework. Moreover, the context unit as such, 
a con~plex discourse with an extensive parallel, is 
given as well because a simple response is inadequate 
just in case corresponding information on a relevant alternative 
to a focus in the query is available to the 
responder. As far as wh- questions are concerned, the 
focus coincides with the wh- position. As regards 
yes-no questions, it may be any item suitable as a 
depart in a particle scheme. It is adequate, then, to 
supplement the simple response by the corresponding 
information. The sentence frame of the question car- 
ries over from the direct to the indirect answer as 
the focus (and possibly polarity too) is exchanged. 
(Occasionally, the focus (depart) is empty so that the 
alternative (report) adds a piece of information, 
typically an adjunct.) Thus complex response creates 
discourse relations such as necessitate discourse 
particles in a systematic way and on independent 
grounds. It may be considered the basic situation 
of utterance for discourse particles. 
550 
There is consensus that an extension to a response is 
appropriate if and only if the information it conveys 
is relevant. The challenge consists in defining what 
constitutes relevance in each single case. It is a 
prime desideratum to develop general guidelines for 
selecting alternatives. Relevant means relevant to the 
goal of the dialog, and any sensible approch will take 
Grice's Maxim of Quantity as a point of departure: 
Make your contribution as informative as is required 
for the current purpose of the exchange. There are 
various ways of exploiting this principle. One is to 
relate information to the assumed practical purpose of 
the query so as to suggest surrogate courses of action 
in case the direct answer is negative. Such a strategy 
is pursued by KAPLAN (1983). Alternatives will be 
identified on the basis of a functional synonymy: 
- Do you hove a rnatchl 
- No, but I hove a lighter. 
Another approach, adopted by JOSHI et al. (1984), 
is to focus on world knowledge so as to correct false 
default inferences licensed by the direct answer by 
stating exceptions to normal courses of events. 
Alternatives will be identified by way of stereotypes: 
- Is Tweety a blrd~ 
- Yes, but he cannot fly. 
In the theory of scalar conversational Impllcature, 
applied to question-answering by HIRSCHBERG (1985), 
the Maxim of Quantity is revised to refer to the 
strength of an utterance: Make the strongest relevant 
claim justifiable by evidence. The concept of relevance 
remains, but it is anchored to linguistic knowledge by 
a semantic relation: Strength as surfacing in scalarity. 
~;uch a strategy embodies two clear advantages: 
A stronger version of a question, whether positive or 
negative, cannot rationally be known to the questioner 
in advance; and, the search for a stronger version 
can be guided by rules which must be represented in 
a reasonably intelligent system anyhow, namely, 
lexical relations and meaning postulates. 
Semantic scales are defined by tuples of lexical items 
linearly ordered by entailment. Consider as an example 
the pair possible and probable and a query Is It 
possible or Is It probable for some proposition It, 
and assume the adjective to be the focus. The answer 
no to the former question will answer the other one 
too, as will the answer yes to the latter question. 
The answer yes to the former question will not, nor 
will the answer no to the latter, yet a responder is 
required to make the strongest relevant claim, and 
provided the other item counts as relevant, there 
is a straightforward way to do so: 
- Yes, It Is even probable. 
- No, It Is not even possible. 
In fact, if the maxim is revised to require the 
responder to assess the strongest relevant proposition, 
two more responses emerge as adequate, again on the 
condition that the other item counts as relevant: 
- Yes, but It Is not probable. 
- No, but it Is possible. 
In a wider sense, semantic scales are defined by 
tuples of lexical items arranged by entailment in a 
hierarchy of set inclusion and exclusion. Consider as 
an example the quadruple Scandinavian, Danish, Nor- 
wegian, and Swedish, and queries Is /t Scand/nav/an 
etc. for some referent It, and assume the adjective to 
be the focus. The answer no to the first question will 
answer all the other questions too, as will the answer 
yes to any subsequent question. The converse is not 
the case, yet a responder is required to make the 
strongest relevant claim, and provided the other items 
count as relevant, there is a straightforward way: 
- Yes, Cand In fact). Danish~Norwegian~Swedish. 
- No, (but~ Norwegian~Swedish~~ 
Danlsh/Swed/sh//Dan/sh/Norweg/on. 
- No, not Seandlnov/an at all. 
The proviso was made that the other items count as 
relevant, as the responses were given on the maxim 
"Assert/assess the strongest proposition relevant". 
Note, however, that a certain measure of relevance is 
secured by the circumstance that that proposition is 
not the strongest proposition as such, corresponding 
to a contradiction, or just any strong proposition, 
but one among a limited number stronger than another, 
in fact, exactly one as polarities go, corresponding 
to a (true) sentence entailing the question supplied 
with a sign and obtained by exchanging one item. 
So a link is established between the direct and the 
indirect answer prior to relevance considerations. 
Relevant alternative candidates are selected on the 
basis of independently accessible linguistic knowledge. 
The relevance question is reduced from What is rele- 
vant? to Is th/s relevant?; the discovery procedure 
is transformed into a decision procedure, and this 
process is low-level and domain-independent. 
Items that are interconnected by a semantic relation 
such as scalarity and hyponymy seem to tend to be 
mutually relevant so that irrelevance cases can be 
considered exceptions to the rule. It is not impossible 
that the assessment of a higher value is irrelevant 
once a value is confirmed or that the assessment of 
a lower value is uninteresting once a value is denied, 
but it is as improbable as it is that items arrived 
at on more pragmatic considerations are irrelevant. 
Likewise, one cannot exclude that the confirmation of, 
say, a subkind once a superkind is confirmed or the 
confirmation of a sisterkind or the denial of a super- 
kind once a natural kind is denied is uninteresting, 
but one can think it equally improbable. So regularly 
- by default - when there is a scale or a hierarchy 
around the item in question, all items in that scale 
or hierarchy will enter into the set of propositions 
at issue, then on meaning postulates, defining the 
interrelationships in terms of entailment, one al- 
ternative can be identified as the informative in de- 
pendence on the distribution of polarities in that set. 
With reference to the parametric discourse relations 
and particles paradigm presented above, the sl<etched 
cases of complex response cover five relation classes. 
Each of these is strongly motivated in tile framework 
of cooperative response insofar as any complex answer 
patterned on it is appropriate in principle. Given a 
query <C,D> where D is the focus, any complex 
response +/-,*/-<C,R> where R is an alternative 
such that the parameters are chosen accordingly 
- and truthfully - is basically adequate. 
When moving upward on a scale, in case the stronger 
statement is verified, the particle sogar applies; 
otherwise, if it is falsified, the particle aber. 
When moving downward on a scale, in case the weaker 
statement is verified, the particle abet applies again; 
otherwise, if it is falsified, the particle elnmel. 
When moving d6wnward in a hierarchy and the state- 
ment is verified, the specification particle und zwar 
applies. When moving upward and the statement is 
falsified, the particle iJberhaupt applies. Finally, 
moving sideways to verify, sondern is the particle. 
551 
The PASSAT System 
The tiny database query system PASSAT, consisting 
in one PROLOG II program comprising approximately 
100 clauses, is designed to demonstrate a register of 
rules regulating choices, of alternatives to lexical 
items and of particles of discourse, in accordance 
with semantic relations and in terms of polarity. It 
is devised to imitate a natural performance in three 
respects : 
- Quality, the search for the true response; 
- Quantity, the quest for the exhaustive response - 
information on a relevant alternative; 
- Coherence, the search for the discourse particle; 
and lexical entailments underlie all three aspects. 
So while the system is primarily intended as an 
illustration of a facet of centerpiece functions of 
German discourse particles, it is at the same time a 
small but systematic model of complex response 
principled on independently available knowledge. 
PASSAT exploits a sortal hierarchy of natural kinds 
and a scalar structure of ranked items to arrive at 
relevant alternative data and to select appropriate 
discourse particles to bridge the gaps, borrowing its 
terminology and database facts from shipping. Such 
computations rely on a variety of modules: 
- Lexicon. Here, semantic relations between and among 
lexical items, such as sortal sameness, "antonymy" 
(Disjunctivity: Difference and sortal sameness), 
hyponymy, intersectivity (cross concepts, uniting 
different-sort items), "synonymy" (comparability in a 
strict sense), and scalarity, are designated and 
defined in their own terms. 
- Meaning postulates, where semantic relations 
introduced in the lexicon are defined and interpreted 
by entailment, that is, in terms of (necessary) 
polarity in view of the sentence context. 
- Alternative relations, stating conditions on which 
one item constitutes an alternative to another in 
terms of lexical relations and (simple) polarity in 
view of the sentence context. 
- Particle relations, stating sufficient conditions 
for the output of a certain particle in terms of 
alternative relations and (simple) polarity in view 
of the sentence context. 
- Response rules, evaluating original queries and 
perceived and received substitute queries 
(alternatives in the sentence context) in terms of 
database facts or meaning postulates. 
- Knowledge bose, containing the minimal amount 
of primitive fact (no facts that are deducible from 
other facts on meaning postulates), representing 
(predominantly positive) polarity. 
{ dampfschlff, motorschlff, segelschlff } 
{ schraubendampfer, ra~damp~ 
{ bark, brlgg, schoner, vollschlff } 
This is one of the two sortal hierarchies PASSAT is 
acquainted with. Questions to be understood by 
PASSAT are of the form 
- Ist x P? 
where x is an individual name (that of a ship) and P 
is a predicate, e.g. a common noun, so throughout it 
is a question of a constant's membership in a set. 
First answers (ja or neln) are by a large measure 
calculated by way of meaning postulates defined on 
lexical relations like hyponymy or antonymy, and these 
same relations go in turn to compute second answers. 
Once a first response is found, PASSAT goes on to 
seek alternatives: Provided that answer was yes, it 
seeks to enhance the specificity of the predicate, 
e.g. to restrict the set denoted by the noun by 
moving downward in the sortal hierarchy: 
- Ist fortuna segelschlffl 
- Jo, und zwor bark. 
In case polarity is negative in the first run, the 
system seeks to increase informativity by e.g. 
searching for the set to which the individual does 
belong (moving sideways in the sortal hierarchy): 
- Ist preclosa bark? 
- nein, sondern brlgg. 
These two basic principles are enriched and extended 
by a recurslve mechanism: As soon as an alternative 
to the subject of interrogation has been determined, 
the search goes on for an alternative to that alter- 
native, entering into the rble of the subject of 
interrogation, and so on: 
- Ist concordla dampfschlffl 
- neln, sondern segelschiff, und zwar schoner. I II 
I 
- Ist prudentla schoner? 
- neln, I 
I ~berhaupt nlcht segelschlff, 
~J r 
sondern dampfschlff , 
I und zwar raddampfer. 
On the other hand, PASSAT is acquainted with an- 
other hierarchy too. The concept ship is subdivided 
on two equivalent points of view, the locomotion and 
the function : 
{ frachtschlff, possaglerschlff, spezlalschlff} 
\[ tankschlff, schiJttgutschlff, sttJckgutschlff } 
552 
And composites are introduced which combine these 
two hierarchies: For a constant to be a member of 
such a set, it must belong to both sets denoted by 
the two components: 
{ f rachtdompf er , I~ssoglerdompf er , motortonker I 
And here, a context sensitivity inside the complex 
answer has been installed (by means of an extra 
variable position in the alternative relations) to 
permit a second and a third alternative to the first 
answer to be stalled until the sequences of "lower- 
level" alternatives to the second and third answers 
(first and second alternatives) are exhausted, to be 
readdressed with bocktrocklng: 
- /st prudentio motortonker? 
- neln, 
soneer, ch,rr. / 
und zwor roddompfer, \ 
t 
und ouch nlcht tonkschlff, 
I iJberhoupt nlcht frochtschlff, 
I sondern possoglerschlff, 
also possoglerdompfer. 
- /st poseidon possoglerdompferl 
- neln, 
zwor dampfsc~iff, I 
und zwor schroubendompfer, \ 
ober nlcht possoglerschlff, 
I sondern frochtschlff, 
I und zwor stiJckgutfrachter , 
elso frochtdompfer. 
The conditions under which a cross-concept like 
frochtdompfer is an (ultimate) alternative to another 
are rather complex insofar as it requires numerous 
steps to come to a conclusion as to whether to draw 
a conclusion by use of Mso (approximating English 
so). It depends on the arrangement of both of the 
two intersected kinds, in casu, dompfschlff and 
frochtschlff, in relation to the other pair, say, 
motorschlff and tonkschlff uniting to motortonker. 
- /st pose/don motortonker? 
- neln, 
nlcht motorschlff , 
sondern dompfsch/ff, 
und zwm' schroubendompfer , 
und ouch nlcht tonkschlff, 
sondern stiJckgutfrachter. 
Thus frochtdampfer is no alternative to motortanker 
because the two corresponding component kinds 
tonkschlff and frochtschlff are downward specific (the 
former is more specific than the latter), whereas the 
converse is not the case - motortonker does form an 
alternative to frochtdompfer as frochtschlff and 
tonkschlff are upward specific; the latter is more 
specific than the former. 
\[ gest rondet , gescheltert \] 
This is one of two scales known to PASSAT. Again, 
comprehehsible queries are of the form Is? x P? where 
x is an individual name and P a predicate, but this 
time the predicate is not a common noun but a (per- 
fect participle) intransitive verb. Once a primary 
answer is given to a query, a search starts for" an 
alternative answer once over, and given a positive 
primary response, the system seeks, again, to en- 
hance the specificity of P, only now not by seeking 
to restrict the set by downward movement with re- 
spect to a hierarchy but by upward movement on 
the scale, to assess the next value irrespectively 
of whether it is valid or not: 
- /st precloso gestrondetl 
- /o, sogor gescheitert. 
- /st fortune gestrondet? 
- Jo, ober nlcht gescheitert. 
Given a primary answer with polarity negative, 
however, as before, PASSAT tries to increase the 
information value nevertheless through strengthening 
the statement, but not by searching for confirmation 
sideways or a more comprehensive denial upward in a 
hierarchy, 
- /st prudent/o schoner? 
- neln, fJberhoupt n/cht sege/sch/ff, sondern 
dompf schlff . . . 
but by assessing the next value in the downward 
direction on the scale induced by the verb: 
- Ist fortune gescheltert? 
- neln, abet gestrondet. 
- Ist fellclo gescheltert? 
- neln, nlcht elnmol gestrondet. 
Now there is another scale known to the system: 
\[gekentert , gesunken \] 
And the two scales are associated with one another in 
a structure which presents a pragmatic case of alter- 
nativity (the only one in the system). More precisely, 
the items gekentert and gestrandet are in a symmetric. 
relation termed syn as a pseudo-synonymy case, with 
the consequence (and purpose) that in case the - 
primary or secondary - answer to either one of the 
lower values - as a query or alternative - is 
negative - in the first or the second instance, and 
the answer to the other lower value is positive, then 
that other lower value is treated as an alternative, 
on the consideration that in view of the higher goal 
of the query, i.t will be of interest: 
- Ist concordla gestrondet? 
- neln, wohl ober gekentert, sogor gesunken. 
- Ist concordlo gescheltert? 
- nein, nlcht elnmol gestrondet, 
wohl ober gekentert, sogor gesunken. 
553 
Limitations 
The test of any natural-language system, whether 
generative or interpretative, is in its measure of 
generality or flexibility, in its aptness for expansion 
and extension in various dimensions. As far as the 
present program is concerned, these dimensions can 
be identified with a range of linguistic modules: 
Morphology (1), lexicon (2), syntax (3), semantics 
(4), and pragmatics (5). 
(I) Deliberately, no morphology has been built into 
the system. Relevant items would have been (a) the 
indefinite article (e/n), (b) gender variants (eln/-e) 
and (c) coherent forms (keln/-e). These refinements 
are omitted in order not to pay undue complexity to 
such inessentials, though the implementation would be 
feasible. 
(2) The lexical items and relations are not casuistic 
in the sense that they are unrepresentative of 
hierarchies and scales in German. Parallel structures 
can be added or substituted without difficulty. Only, 
real hierarchies and scales do not exist in isolation 
but in integration in a taxonomic superstructure. One 
problem is that a concept (e.g. sch/ff) may be open 
to specification in sequence (e.g. Ja, und zwor segel'- 
schlff, und zwar vo/Ischiff), another is that a con- 
cept (e.g. schiff) may be open to specification in 
two directions (e.g. \]o, und zwer possaglerschlff und 
motorschlff). Before the concept sch/ff enters into 
PASSAT in the obvious way, a method must be de- 
veloped to determine how far and which way relevance 
is to reach in each case in view of the user's interest. 
(3) Deliberately, only a minimal syntax has been built 
into the system. This is, again, to accentuate the 
central principles, but more to not create the 
impression that interesting syntax problems have been 
solved. Thus the rules of ellipsis have not been 
explored. PASSAT uses total ellipsis throughout = 
though not on deliberation, but by necessity. It could 
instead use partial ellipsis discriminately to put out 
answers like neln, zwer /st s/e eln x, und zwer eln y, 
abet s/e /st ke/n z, s/e /st iJberhaupt kein u, sondern 
e/n v = yet it would still do so not by first 
generating and then reducing complete structures but 
by producing those strings blindly. 
{4) The system suffers a serious shortcoming in not 
assessing the lexical relations in meaning postulates 
and alternative relations in terms of the sentence 
context semantic structure. By accident, predicates 
(common nouns or verbs) occur in predicative position 
with the copula throughout, so the semantic relation 
invariably comes to the surface. As soon as contexts 
are introduced where the noun e.g. serves to restrict 
quantification over a binary relation, or, as a 
simplification, it is an argument of a binary relation, 
as in hat fortunotus e/ne brigg (compatible with 
fortunetus hot elne bark), the relation ceases to 
carry semantic relevance, and the system must con- 
sider some semantic representation to judge whether 
meaning postulates and alternative relations apply. 
(5) The pragmatic open problem lies in where the 
focus lies, more exactly, to which component of the 
question any alternative can be considered relevant 
in particle terms, what part is the plausible depart 
in the first instance. At the current state of the 
system, the focus is located once and for all in the 
one-place predicate P, yet it is a commonplace that 
yes-no questions are systematically ambiguous insofar 
as their topic-focus structure has consequences for 
what constitutes a proper substitute, namely, a 
sentence where the topic stays the same and the focus 
changes. Thus it could be that a question like Ist 
concordla elne bark is intended to, in the event of a 
negative response, elicit not a continuation like 
sondern eln schoner but an extension like ober 
fortuno Ist elne - for example, in case the higher goal 
of the query is to ascertain that there is a bark 
available for a higher purpose still. In principle 
there are means in a language to posit focus, and 
those means may be syntactical, like word order or 
clefting. Such are not, however, available to 
PASSAT, being a German-interfaced system, so 
without the phonology Germans use it is necessary 
to exploit a memory of past exchange, so as to 
address the higher goal of the query directly. 
References 

HIRSCHBERG, J. (1985) A Theory of Scaler 
Impllcoture. Univ. of Pennsylvania diss. 

JOSHI, A./B.WEBBER/R.WEISCHEDEL. (1984) 
"Preventing False Inferences", in Proceedings 
of COLING 18tI. 

KAPLAN, J. (1983) "Cooperative Responses from a 
Portable Natural Language Database Query System". 
in Brady/Berwick (eds.): Computational Models of 
Discourse, Cambridge, Mass.: MIT Press, 167-208. 

SP~B~, K.J. (1988) A Model for Discourse Particles. 
LILOG Report n, Stuttgart: IBM. 
