COLING 82, J. Horeclo) (ed.) 
North.Holland Publishing Company 
@ Aeodcm~ 1982 
NATURAL LANGUAGE UNDERSTANDING AND THE PERSPECTIVES 
OF QUESTION ANSWERING 
Petr Sgall 
Department of Applied Mathematics 
Faculty of Mathematics and Physics 
Charles University 
Prague 
Czechoslovakia 
A method of automatic answering of questions in 
natural language, based only on input texts and 
a set of rules of inference, is described. A first 
experimental system including a grammatico-seman- 
tic analysis of the input texts and questions, a 
procedure of inferencing, a search for appropriate 
answers to individual questions and a synthesis 
of the answers are being implemented, mainly in 
the language Q and PL/I. The output of the ana- 
lysis, the underlying representations of the utter- 
ances of the input text, serves as a base of the 
knowledge representation scheme, on which the 
inference rules (mapping dependency trees into 
dependency trees\] operate. 
The important, though partial possibilities of automatic under- 
standing of natural language gave rise to different kinds of experi- 
mental systems, ranging from sophisticated systems of machine trans- 
lation through various kinds of modelling of dialogue (with robots, 
data bases, etc.\] to question answering. 1 From a linguistic viewpoint 
the main challenge consists in attempting to transfer the burden of 
the communication between humans and computers to the latter, who 
should be able to react in an appropriate way to the user s input 
texts formulated in her or his native language, without serious re- 
strictions. The necessity of thousands of human beings preparing data 
"for computers" (not only encoding messages, but also compiling data 
bases\] should be removed. 
This challenge constitutes one of the central tasks of modern 
linguistics; an explicit description of the main features of the lan- 
guage system, which is necessary for these purposes, must be based on 
a sound theoretical framework suitable for the description of grammar 
as well as of the linguistically patterned aspects of semantics and 
pragmatics. A close cooperation of linguistics with logic, computer 
science and cognitive science has become urgent. This task presents 
also an effective way of checking the results of theoretical linguis- 
tics in various important fields. 
These considerations have led the group of algebraic linguistics 
in Prague (now belonging to the department of applied mathematics, 
faculty of mathematics and physics, Charles University) to start work- 
ing on an experimental system based on the approach called TIBAQ 
(Text-and-~nference Based Answering of Questions).2 Its four main pro- 
cedures are (I\] grammatico-semantic analysisi (2\] rules of inference, 
(3) identification of a full (direct\] Qr partial answer, and (4\] syn- 
thesis; see the overall scheme in Fig. I. 
357 
358 P. SGALL 
input text 
+ 
LOOK-UP FOR 
RELEVANT STATEMENTS 
INFERENCE 
LOOK-UP FOR ANSWER 
SYNTHESIS 
~- ANALYSIS m~eaning 
is it a question? 
set of statements 
set of relevant 
statements 
enriched set of 
relevant statements 
~- answers 
Fig. 1 
An overall scheme of a system based on the method TIBAQ 
(I) The automatic grammatico-semantic a n a 1 y s i s3 is be- 
ing prepare~ in such a form that it can handle Czech and English 
polytechnical texts (papers, reports, monographs) in their usual 
shape, and also questions formulated in Czech. Thus there will be no 
need for the user to "cope with the needs of the computer system". 
The procedure of analysis has the following characteristic properties 
distinguishing it from a mere parsing procedure: 
(i) The analysis procedure is based on a systematic theoretical 
account of the structure of natural language, the functional generat- 
ive description; this linguistic approach, elaborated in the Prague 
group of algebraic linguistics, 4 makes use of the results of the em- 
pirical research carried out in the frame of European structural lin- 
guistics, and also of the methodological requirements formulated by 
Chomsky and the different wings that developed from his school. The 
resulting linguistic approach is used as a general base ensuring that 
the particular practical solutions (in ambiguity removal, etc.) chos- 
en for a restricted area can be replaced by more generally valid sets 
of rules, whenever it appears as necessary to cross the boundaries of 
this narrow area Ce.g. when applying the method to a new kind of 
texts, to a new polytechnical domain, etc.). This is ensured thanks 
to the universal character of natural language and to the fact that 
the linguistic framework (if appropriately chosen) provides means for 
an adequate description of all its subdomains Ccf. Haji~ov~ and Sgall, 
198Oa). 
(ii) In connection with this requirement the analysis procedure 
is designed to transfer the input sentences from their outer form to 
a disambiguated notation of their meanings <which can be identified 
with their underlying structures, in the framework of functional 
LANGUAGE UNDERSTANDING AND QUESTION ANSWERING 359 
generative description). The level of meaning of sentences includes 
such syntactic units as Actor, Objective, Addressee and other partic- 
ipants or cases, Manner, Instrument, Place, Direction and other free 
adverbial modifications, as well as lexical and morphological meanings 
(the latter including e.g. number, tense, modalities>. This level is 
formulated as a linguistic counterpart of intensional structure,which 
makes it possible to define the concept of strict synonymy of expres- 
sions and to ensure an algorithmic transition'to a postulated univer- 
sal formal language of intensional logic~<among the trends that start- 
ed with Montague, our account of meaning- stands close to that by Da- 
vid Lewis, though the form of formal language we prefer has much in 
common with Tich~'s framework>. The representations of the meanings 
of sentences serve as the main components of knowledge representation 
in the semantic networks of the systems based on the method TIBAQ. 
They can be illustrated by the representation in Fig. 2. 
<iii> As can be seen from this representation, our approach 
works with dependency trees as the form of meanings of sentences. 
This allows us to work with relatively simple underlying structures 
in which such notions as "head" and "modifier", or "noun" phrase vs. 
"verb" phrase, as well as the relations described by Fillmore as cases 
find an economical treatment. 
<iv> Not only the roles of the elements Of syntactic relations, 
but also the topic-focus articulation of sentences finds its proper 
place in the representations yielded by this procedure of analysis. 
Also the whole pragmatically based interplay of topic, fccus, contex- 
tual boundness and communicative dynamism, as combined with the recur- 
sive properties of sentence structure can in principle be rendered in 
the chosen form of representations of the meanings of sentences. 6 
Analysis of written texts does not allow for a complete identification 
of all the items relevant for the topic-focus articulation, and the 
present form of our algorithms gives results which are not fully re- 
liable, but the errors appear to be neither too numerous nor too grave 
for the given purpose. The main rules consiat in understanding the 
parts of a sentence standing to the left of the finite verb as belong- 
ing to the topic, while the verb itself <if it is not semantically 
void, as the copula, or become, carry out, etc.) and the elements 
following it are classed as belonging to the focus in the Czech poly- 
techpical texts. 7 Such a treatment appears as sufficient for ensuring 
that those cases in which the topic-focus articulation is semantically 
relevant will be handled appropriately. This concerns the relative 
scopes of quantifiers in such sentences as Every car has several 
wheel_______ss and the "holistic ~ understanding of the topic e.g. in Smokin~ 
is dangerous, as well as Kuno s "exhaustive listing" and the difference 
between thetic and categorical judgements; even more important is the 
relevance of the boundary between topic and focus for the determina- 
tion of the scope Qf negation, and thus also for the identification 
of presuppositions in some cases: Many arrowsd~dn't hit the target 
does not imply that the target wasn-t hit by many arrows, and The 
king of France didn't come to COLING 82 does not presuppose th~-~x- 
istence of a king of France. The relevance of topic and focus for 
natural lanuuage understandinq is most clearly recognized in connect-. 
ion with the assignment of reference to definite noun phrases <and 
other expressions>. 
Cv~ The procedure of analysis provides also for a treatment 
of the interconnections between the individual assertions <which are 
Stored in the shape of the meanings of sentences). This is done by 
means of two main devices: first, in the representation of each lex- 
ical meaning in the lexicon there is an indication of the relations 
360 P. SGALL 
of synonymy and hyponymy (subordination, su\[erordination> of the given 
item to others, and also semantic features are used (for a partial 
modelling of the object domain pertinent to the treated area of poly- 
technical texts>; 8 second, the relation between an object and the oc- 
currences of expressions referring to it in the texts is handled by 
means of a register or concordance, supplying addresses of all the 
occurrences of a given unit in the whole set of knowledge represen- 
tation. 
After having examined different means of implementation of the 
analysis procedure, esp. Kay's parser, Wood's ATN, the Grenoble system 
and others, we decided that among the systems actually available to 
us the framework elaborated in the T.A.U.M. group, based on Colme- 
rauer s Q-systems, can serve best our aims. Thanks to the Canadian 
colleagues we got the possibility to implement Q-systems <through For- 
tran> on such computers as IBM 360, EC 1040 (Robotron>and others (by 
means of a procedure given at our disposal by B. Thouin who together 
with R.Kittredge introduced us to the intricacies of their systems>. 
It appeared that Q-systems are a means flexible enough to be used for 
our purposes, in spite of the fact that several major differences can 
be found between the original goals Q-systems were designed for and 
between our goals: after a couple of years of experience our program- 
mers <first of all Z.Kirschner and K.Oliva> are able to use Q-systems 
for a dependency-based analysis attempting to penetrate into the under- 
lying structures of sentences <which is necessary also for translation 
between typologically different languages>. The trees Q-systems were 
designed to operate on can be readily interpreted as standing close 
to our dependency trees <though instead of each of the nodes exempli- 
fied in Fig.2 it is necessary to have a whole subtree composed of sev- 
eral nodes, since Q-language works only with elementary node labels>. 
Moreover, it became also clear that Q-systems are a suitable means to 
handle inflectional languages exhibiting complicated systems of mor- 
phemic ambiguity and synonymy, 9 as well as the so-called free word 
order <which is not free at all, but determined by the topic-focus 
articulation, esp. by communicative dynamism, in a much more straight- 
forward way than is the case in English>. It is not necessary to work 
with individual rules for the different permutations of the elements 
of a sentence, since an approach working - roughly speaking - with an 
elementary dependency tree for every tentative clause <a finite verb 
and its neighbours on both sides> is possible, including the use of 
list variables for the irrelevant parts of the tree. I0 
The strong combinatoric power of Q-systems, as well as its re- 
lative transparency, made it possible to formulate a procedure of 
analysis, which is by far not yet complete, but which accounts already 
for hundreds of kinds of phenomena from the syntax of Czech. These in- 
clude a relatively complete analysis of the structure of noun phrases, 
achieved by means of checking the agreement of an adjective with its 
governing noun, and preferring a noun in the genitive case to be under- 
stood as an adjunct of an immediately preceding noun, whenever this 
is possible, while with the other oblique cases <simple and preposi- 
tional> there is a complex scale, elaborated by J.Panevov~, deciding 
whether the given noun functions as an adjunct of this or that preced- 
ing noun or as a modifier of the verb <the indices of the given nouns, 
verbs and morphemic means are used to determine the specific dependency 
relation>. The participants modifying the verb are identified with the 
help of lexical data concerning valency <obligatory and optional mod- 
ifications and their usual morphemic forms>. ~fe mentioned alreaJy the 
identif~catfon of topic and focus, achieved precisely on the base of 
the "fzee" word order. 
LANGUAGE'UNDERSTANDING AND QUESTION ANSWERING 361 
Thus it seems that a syntactico-semantic analysis of the texts 
of a limited polytechnical domain (we started with texts on operation- 
al amplifiers> is feasible. In other words, it is possible to obtain 
in an automatic way an image of the input text having the shape of s 
set of disambiguated underlying representations of sentences <called 
statements in the sequel>,interconnected by means of pointers based 
on the lexicon and on the paradigmatic relations registered there 
(hyponymy, etc.>. 
Whenever a user's input question is analyzed <by the same anal- 
ysis procedure as the statements are>, the system goes over to other 
procedures, which operate on the set of statements gained by the 
grammatico-semantic analysis. 
(2} First of all, the whole set of statements is searched through 
(by means of the concordance we mentioned in (v> above>, to identify 
the subset of statements possibly relevant to the given question <in 
the first experiments, a non-empty intersection of the two sets of 
autosemantic lexical units being treated as a sufficient condition 
for these statements>. The rules of i n f e r e n c e , which are 
then applied to this restricted set of statements, are described (to- 
gether with the procedure of identification of appropriate answers> 
in the short communication presented by P.Jirk~ and J.Haji~, who are 
the main authors of the respective programmes; we can limit ourselves 
here to a few illustrations of these two procedures. In the rules of 
inference such modifications of the statements are included as the 
deletion of an adverbial unde~ certain conditions <e.g. from "It is 
possible to maintain X without employing Y" it fellows that it is pos- 
sible to maintain X), or several shifts of verbal modalities, a shift 
of Actor and Instrument in some cases, and also a conjuction or a sim- 
ilar connection of two statements; e.g. "X is a device with the prop- 
erty Y" and "X can be applied to handle Z" are combined to yiel d "X 
is a device that has the property Y and can be applied to handle Z". 
In the first experiment the inference rules are applied only 
during the handling of a given question. In case a procedure checking 
all newly analyzed statements forcompatibility with the already giv- 
en pieces of information is formulated at a later stage of the re- 
search, then it will also be necessary to decide which inference rules 
should be applied already during that procedure <i.e. independently 
on questions asked by the user>, and which types of consequences should 
be included permanently in the stock of data. It will also be necesn 
sary, in further experiments, to use heuristic strategies for the 
choice of the inference rules to be applied at a given time point. The 
growth of the enriched set of statements must be controlled and lim- 
ited. 
(3) The enriched set of relevant statements is then searched 
through by means of a procedure of the choice of an a n s w e r . The 
representation of the question is compared wSth the statements belong- 
ing to the enriched set, with three kinds of possible results: 
(a) the Statement is found to give a full answer to the question, 
if the two representations differ only in that the answer includes 
specific lexical units (perhaps a whole subtree> in the position oc- 
cupied by the question word in the question <this position being shift- 
ed to the end of our representation of the question>; 
(b) the statement contains information which probably can be o9 
interest to the user, though either some of the parts of the two re- 
presentations are not identical, e.g. these representations differ in 
what concerns hyponymy, or in semantically relevant aspeCts of their 
word order (com/nunicative dynamism>; 
362 P. SGALL 
<c) the statement is not relevant for the given question, if 
either the sequence of edges of the tree going from the root to the 
question word does not have a corresponding counterpart there, or if 
the two representations are radically different in their other parts. 
In case (b) the representation of the answer is assigned the 
prefix "I <only> know that ..." to point out that the answer is not 
complete. 
<43-An answer undergoes then the procedure of s y n t h e s i s, 
transducing the underlying representation to the graphemic shape of 
a Czech sentence. This procedure has been implemented in PL/I on the 
computer EC 1040 and is being checked within a rather broad system 
of random generation of Czech sentences, which encompasses several 
hundreds of rules covering most different grammatical phenomena of 
all levels (cf. Panevov~'s paper presented at this conference). 
The system prepared for the first experiments with the method 
TIBAQ is limited in several respects. An enrichment concerning the 
linguistic aspects <broadening of the lexicon, inclusion of yes/no 
questions> does not seem to be too difficult, since the grammatical 
patterning has already been included in the algorithms to a rather 
large extent. Thus the two main problems that have to be solved in 
adapting the system to handle open texts from a chosen branch of pol- 
ytechnics or science in an appropriate way consist in 
(i) the relation of instantiation <or of different objects 
bearing tha same lexical denomination), i.e. of the assignment of 
reference to definite noun phrasesand other expressions has to be 
solved (in the texts processed in the first experiments only general 
concepts are present, so that up to now this step was not necessary); 
at least three kinds of means should be used here, namely the degrees 
of salience of the images of individual objects in the stock of know- 
ledge shared by the speaker and the hearer (see Haji~ov~ and Vrbov~, 
this volume, about preparatory studies in this direction), further an 
evaluation of the known tendency to keep the topic of an utterance in 
its function also in the next utterance of a connected text, and, 
thirdly, rules concerning the role of factual konwledge in the deter- 
mination of reference; this last point, which goes beyond the lin- 
guistic structuring, is probably restricted to a rather narrow domain 
in well-formulated technical t@xts; 
(ii) an enlargement of the rules of inference <from about thirty 
that were already formulated to hundreds of them); it may be necessary 
to add rules of new shapes and to have a procedure for checking what 
effect a specific rule of inference will have in connection with the 
individual lexical and grammatical phenomena; in this respect only 
the first steps have been done in the empirical research, so that 
when enriching the lexicon we may face new problems of checking all 
the already formulated rules Of inference. Only when more experience 
in these new fields in gained will it be possible to formulate regular 
patterns and general procedures which could be adequate for these new 
areas of artificial intelligence. Such an inquiry certainly belongs 
to most promising directions leading to a deeper insight into the re- 
lationships between communication and cognition. 
FOOTNOTES: 
1 We do not have in mind here the systems including only an 
elementary or marginal linguistic equipment, though some of them can 
well serve the purposes of text information retrieval (cf. e.g. the 
method MOSAIC, intended for automatic indexing and for extracting, 
LANGUAGE UNDERSTANDING AND QUESTION ANSWERING 363 
prepared by Z. Xirschner in the Prague group>, or of natural language 
front-end contact with data bases. Systems belonging to the domain of 
artificial intelligence and serving for man-machine communication in 
natural language need a much more complete linguistic elaboration. 
This concerns the systems intended for open set~ of instructions for 
a robot and for a dialogue with it (Winograd's SHRDLU>, with which 
the robot's reactions can serve as a criterion for checking whether 
the input was "understood" by the system. The investigations Of KRL 
by Bobrow, Winograd, Kay and others, the task oriented dialogue sys- 
tem prepared at SRI <Robinson, Hendrix, Hobbs, Grosz and others>, as 
well as e.g. the models of dialogues constructed by the group of 
W. yon Hahn in Hamburg may be classed with the systems of natural 
language understanding. However, these systems (and also those design- 
ed to analyze or generate narrative and other texts on the base of 
scripts, scenarios and similarly9 differ in the level and complete- 
ness of the linguistic approaches underlying them. As for machine 
translation, it is interesting that most of the linguistically well 
equipped systems Cthose of Vauquois and his group, of Kulagina~ of 
Apresjan and of T.A.U.M.> concern French. 
2 The first characterization of a question-answering system of 
this kind was presented at the 6th International conference on com- 
putational linguistics, Ottawa; see Haji~ov~ (1976). 
3 A preliminary characteristics of this procedure can be found 
in Panevov~ and Sgali(1979); as for an account including illustrat- 
ions of its technical aspects, see Panevov~ and Oliva (in press>. 
4 See Sgall et al. (1969>; Haji~ov~ and Sgall (198Oa>; Sgall, 
Haji~ov~ and Panevov~ (in prep.>. 
5 Sgall, Haji~ov~ and Proch~zka (1977>; Sgall (i980>. 
6 For a short empirical and formal characterization of this 
interplay see Haji~ov~ and Sgall (1980b>; more details are given in 
Sgali, Haji~ov~ and Panevov4 (in prep.>. 
7 In English the situation is more difficult, since even in 
printed texts it is usual here that the intonation pattern of a sen- 
tence is marked, esp. with adverbials of time and place following 
the intonation centre (which cannot be readily recognized by an auto- 
matic analysis of the written sentence>; these adverbials in such a 
position belong to the topic: W e came to PRAGUE yesterday differs 
from We came to Pra ue YESTERDAY; see Haji~ovl and Sgall (1975;1980b>, 
where some"rules of thumb" for the identification of topic and 
focus in such sentences were given. 
8 For our example in Fig. 2 with the lexical unit device there 
are pointers to such subordinated units as operational amplifier , 
filter, bandpass filter, stopband filter, etc.; all these units are 
assigned the semantic feature "device"; apply has a pointer to its 
synonym use; design has a semantic feature of an action noun and a 
pointer to its synonym ~, while system is assigned a semantic 
feature of "intellectual category". 
9 The morphemic analysis of Czech was implemented in PL/I in 
the seventies, see Kr~llkov~, Weisheitelov~ and Sgall (1982). 
I0 Cf. Panevov~ and Oliva (1982); a German translation of 
Colmerauer's definition of Q-systems will appear in Prague Bull. of 
Mathematical Linguistics 38, 1982. 
364 P. SGALL 
BE 
AMPLIFIER-b-Ag DEVICE-Pred 
OPERATIONAL-b-Det VERSATILE-Det SPAN-Det 
APPLY-b-Ag BC "CONDITION-Pat AND DESIGN-Pat 
GE.N-b-Ag DEVICE-b-Pat GEN-b-Ag SIGNAL-Pat GEN-b-Ag SYSTEM-Pat 
Fig. 2: A TR of "Operational amplifier is a versatile device with 
applications spanning signal conditioning and systems design" 

REFERENCES

\[i\] Haji~ovl, E., Question and answer in linguistics and in man-mach- 
ine communication, Statistical Methods in Linguistics (SMIL) 
(1976), No. i, 36-46. 

~2\] Haji~ov~, E. and Sgall, P., Topic and focus in transformational 
grammar, Papers in Linguistics 8 <1975) 3-58. 

\[3\] Haji~ov~, E. and Sgall, P., Linguisti~ j meaning and knowledge re- 
presentation in automatic understanding of natural language, 
in: COLING 80 - Proceedings of the 8th Int. Conference ok Comput- 
ational Linguistics (Tokio) 67-75; reprinted in Pragu~ Bull. of 
Mathematical Linguistics 34 (1980a) 5-21. i 

\[4\] Haji~ov~, E. and Sgall, P., Dependency-based specif~ation of 
topic and focus , Statistical Methods in LinguistiCs (SMIL) 
(198Ob), No.l-2, 93-140~ 

\[5\] Kr~ifkov~, K., Weisheitelov~, J. and Sgall, P., Automatic morph- 
emic analysis of Czech, Explizite Beschreibung der Sprache und 
automatische Textbearbeitung VII (P~ague, 1982). 

\[6\] Panevov~, J. and Sgall, p., Towards an automatic parser for Czech, 
Int. Review of Slavic Linguistics 4 <1979) 433-445. 

\[7\] Panevov~, J. and Oliva, K., On the use of Q-language for syntactic 
analysis of Czech, in: Explizite Beschreibung der Sprache und auto- 
matische Textbearbeitung VIII <Prague, 1982). 

\[8\] Sgall,P., Towards a pragmatically based theory of meaning, in: 
Searle, J.R., Kiefer, F. and Bierwisch, M. (eds.), Speech act 
theory and pragmatics (D. Reidel, Dordrecht, 1980, 233-246). 

\[9\] Sgall,P., Haji~ov~,E. and Panevov~, J., The meaning of the senten- 
ce in its semantic and pragmatic aspects <Academia, Prague, in 
prep.). 

Sgall, P., Haji~ov~, E. and Proch~zka, 0., On the role of linguist- 
ic semantics, Theoretical linguistics 4 <1977) 31-59. 

Sgall, P., Nebesk~, L., Goral~fkov~, A. and Haji~ov~, E., A funct- 
ional approach to syntax in the generative description of language 
<American Elsevier, New York, 1969). 
