COLING 82, J. Horeck~ lad.) 
North-Holland Publishing Company 
© Academia, 1982 
FORMALIZATION OF ~ATION 
IN NEWSPAPER TEXTS 
Dietmar F.Roesner 
Math. Institut A 
Joachim H.Laubsch 
Institut f. Informatik 
Universitaet Stuttgart 
D-7ooo Stuttgart 
West Germany 
The paper discusses the role of argumentation schemata and 
their interaction with other knowledge sources within a 
computer model for in-depth understanding of newspaper 
texts about jobmarket developments. Some of these schemata 
are presented; the "ALTHOUGH ..."-argumentation (German: 
"TROTZ ...") and its formalization are discussed in 
detail. 
1. Introduction 
Reading newspapers may be seen as an "everyday cognitive task". Therefore it is 
not astonishing that some work in Artificial Intelligence aims at the simulation 
of aspects of the understanding processes of newspaper readers \[Eisenstadt 1977; 
Rosenberg 1977\]. There have even been pioneering system~: the Script ~0plier 
Mechanism \[Cullingford 1978\] demonstrated careful reading of "event~--oriented news 
stories about earthquakes, vehicle accidents, plane crashes and so on, whereas 
FRUMP \[DeJong 1979\] si~Jlated skimming abilities using texts about the same 
themes. 
We are working with texts about jobmarket developments taken from West German 
newspapers \[Laubsch&Roesner 1980\]. An important aspect of our texts is that they 
deal with argtm~ntations about reported data and their respective changes. In this 
paper we describe our (not yet implemented) conceptual approach for processing 
such structures. 
2. Understanding ar~nentstion 
In our view, argtm~ntation schemata in general work on a kind of "theory" of the 
domain of discourse, i.e. on a structure that encodes knowledge about domain- 
specific dependencies. We use a dependency-network in order to represent the 
average reader's expertise and qualitative reasonin@ ability about the jobmarket. 
Our representation is influenced by ideas from \[Sussman&Steele 1980\]. 
In some sense, one may interpret an argtm~ntation schema as a function that takes 
propositions (<PROP>s) as arguments. In case of coherent or correct use, these 
propositions have to fulfill certain constraints that are defined with reference 
_ to the "theory" of the specific domain of discourse. 
Thus, wundersta~ir~" ~tati(m~ in terms of these schemata can be seen as 
establishing the constraint relations between given propositions by an inference 
mechanism that operates on the dependency net of the domain. This always involves 
testing on discourse coherence (if it is assumed to have a correct "theory") 
and/or checking the "theory" (if it is assumed to have correct texts). 
325 
326 D.F. ROESNER and J.H. LAUBSCH 
3. Explanation schemata 
Explanation schemata deal with reasons - in cur domain these are mostly reasons 
for the jobmarket fluctuations under discussion. Although we use German surface- 
oriented labels (which we transcribe into ~glish), we are discussing all of the 
following examples as deep structures that may be underlying to diverse surface 
constructs. 
~i> <PR~a-2>': (~-OF <PROP-l> <PROP-2>) 
This is a general explanation schema. If used correctly, then <PRUP-I> gives 
reason(s) for <PROP-2> : If <PROP-l> is a "simple" fact in our theory, then there 
must be an inference path in the dependency net along cause-effect links leading 
from <PROP-l> to <PROP-2>. If <PROP-I> is a conjunction of two (or more) "simple" 
facts, then inference paths starting from those points must interfere in such a 
way that they finally lead to <PRDP-2>. 
Among German surface triggers for the BECAUSE-OF-schema are: "WEGEN ...", "WL " 
...", "DA ...", "DURCH ...", "AUFCAqD%D %~N ...", "INFDLGE DES ...", but also more 
elaborate ones like " ... DARAUF ~EN DASS ..." or "... DAMIT ~EN 
DASS " 
<PROP>': (MOREOVER <PROP>) 
This schema may be used if we are discussing a complex situation where 
consequences of several factors interfere and where the already mentioned (or 
derived) propositions are not sufficient to explain a given result. <PROP> must 
satisfy the constraint that it gives additional supportive reason for an 
unexplained effect, i.e. <PROP> opens a new inference-path in the dependency net 
such that interference with previous i~lete paths is possible in a way which 
finally produces the explanation that is searched for. 
<PR~*-I> <PR~*-2>': (ALTH0\[K~ <PROP-l> <PROP-2>) 
This sch~na deals with expectations and their non-fulfillment. Other surface 
manifestations are constructs like "O~90HL <PROP-l> <PROP-2>", "UNGEACHTET <PROP- 
i> <PR0P-2>" or "ZWAR <PROP-l> D~k%X3CR <PROP-2>". 
The relation between <PROP-l> and <PROP-2> in an ALTB0t~H-schema involves a third 
proposition <PROP-2'>, whose constraints are: 
a) <PROP-2'> is in a contradictive relation (e.g. negation) to <PROP-2> 
b) <PROP-2'> could be expected as a (default) conm ~uence of <PROP-l> (or in 
other words: (BECAUSE-OF <PROP-l> <PROP-2'>) could be verified). 
In coherent texts the contradiction between expected and actual development given 
with an ALTH0t~R-schema will demand further explanation. ~0~@exsta~ing" 
~lamat/cm thus involves answering the following questions: 
AI: ~lat was the unfulfilled expectation <PROP-2'> contrary to <PROP-2>? 
I%2: Why did the expectstion~s) <PROP-2'> fail? 
A3: ~at caused the fact(s) of <PROP-2> to happen? 
Until we cannot sufficiently re=.x)Ive these explanation tasks, the ALTH0t~B-sche~a 
will keep active and guide the processing of further input. 
4. A detailed example 
In order to clarify our _approach, let us trace the processing of the flow of 
argumentation in an actual newspaper article (taken from "Stbttgarter 
Nachrichten", March 7, 1979). 
ARGUMENTATION STRUCTURES IN NEWSPAPER TEXTS 327 
S~_= TRDTZ DES ANHALT~-NDEN ~ IST IN DER BUNDESREPUBLI K DIE 
AK~EITSLOS~ZAHL IM FEBRUAR LEICHT Z~KG~GANGEN. (Englo : In spite of continuing 
winter weather the numbe~ of une~@loyed in the FRG decreased slightly in 
february. ) 
Input to the inference machine is vmitten in a surface-oriented frame notation 
(~-~hich could e.g. be produced by a semantic ATN-grammar). The representation of S1 
is: 
(ALTHOUSH-1 
PI~OP-I: (WEATHER TYP~ ~!Ik~I~R 
MOD: * (AIL~IALT~k~)) ~) 
PIK)P-2" (CHANGE-I 
QUANTITY: (NR-UN~LOYED 
AREA: BRD) 
TIME: * (IM FEBRHAR) 
VALUE: (DECREASE 
MOD: ~ (LEICHT) ~) ) ) 
(Slot-fillers with ,,en have been taken literally from the given sentence and are 
processed by "specialists"~ e°g~ IH FEBRUAR is interpreted - aocording to a 
default text convention - as publication year's february.) 
The first step in processing AL~THOL~JH-I is to construct PROP-2' as a negation of 
PROP-2. 
In this case, the "subject" of PIRDP-2 is "CHANGE-I NR-~ (IN ~ FIRG IN 
FEBRUARY ..)", whereas the "predicate" is the filler of the VALUE-slot, i°eo 
SLIGHT-DECREASE. The procedure for generating a candidate PROP-2' preserves the 
"subject", but negates the "predicate". 
What is the negation of a SLIGHT-DBCREASE? For pur~s of qualitative reasoning, 
we take VALUEs for CHANGEs frcm a five point scale from "++" (~-INCRFASE) to 
"~" (~-DECRFASE), i.e. "-~ corresponds to SLIGHT-DECREASE. Interpreting 
negation of SLIGHT-DECREASE as simply taking the complement of {-} with respect to 
t/~e set of all VALUES = {++, +, 0, -, --} would yield I~-DECREASE L%ICHANGI~) 
SLIO}Ff-~ ~-INCREASE\], but a constraint for surface constructs allows us 
to exclu61e IARGE-DECI%EASE in this case. 
If a larger value had been expected (and not the opposite direction of change), 
then this would have been indicated by a modifier like "NUR" (EngI."ONLY") as in: 
TIIOTZ DER SAI~E IM OK"rOBER STIES DIE ARBEIT~LOSEN~ IN DEM MZX~IAT NUR IIM 
0oi AUF 3.3 PROZENT AN (from "Stuttgarter Nachrichten", Nov. 7, 1979). (\]~gl.: In 
spite of change of season in October the unemployment rate only increased by 0oi 
to 3.3 percent during this month.) 
Thus: the structural analysis yields 
PROP-2 ' : (CHANGE-2 
QUANTITY: NR-~LOYED 
VALUE: (O~F {0 + ++})) 
The next step in order to answer A1 is: Can PROP-2' be expected given PROP-l? 
Indeed: we find a (generic) default rule in our dependency net, that relates 
WINTER WEATHER with a (~IANGE of NR~~. 
328 D.F. ROESNER and .LH. LAUBSCH 
I i=.=...= i I TYP: WINTER) d>fault e~tati>n 
path 
QUANTITY: NR~ I wu~: C~ {+ ~}) 
The next t~o sentences only give further details of the change and are no~. 
interesting for the present discussion° 
$2 ,5 $3~ SIE VERRINGEI{r SIC}{ L~I 37°300 AUF 1.134.100o DAS F~TSPRICHT Eik~F~ 
ARBEITSI~3Sm~Q\[KTfE ~ON RDND ~JENF PI~OZENT~ (Englo: It dec~eases by 37o3~0 ~ 
1.134o100o T~nis corresponds to an unemployment rate of ~bout five percento~ 
Argumentation is continued in $4: 
DER PRAESID\]~T DER ~ BUNDESANb~LT F~ ARBEIT~ JOSEF STINGL~ FOEHR/:~ 
DIESE ENTWI~ AM DIJKNSTAG DARAUF ZURUSCK~ DASS DIE ZAHL DER ARBEITSLOS~-N IN 
D~--N AUSS~%~3E~JF~N NICHT L~7~R ~NSTIEG~ (~gio ~ ~rne president of the Federal Labor 
Agency0 JoStingl, attributed this developn~rt to the fact that the number of 
unemployed in outdoor jobs did no longer increase)o 
"FUEHRTE .o DARAUF Z~K~ DASS o0;' indicates a BECAUSE-OF-schema as ~a,.t of 
someones D~"LARATION: 
(DECLARATIfX~-I 
SPE~ER: STINGL 
STATEMENT: ( BECAUSE-OF-I 
PROP-l: (CHANGE-3 
QUA~iTY: (NR-UNEMPLOYED GIKTJP: OOTDOO~-JOBS) 
~/ALU~\]: NON-INCREASE) 
PROP-2: (DEV~TuOI~ENT-I ~-~i~{IS -~) ) ) 
The first step in processing BECAUSE-OF-I is looking fo~ the referent of PROP-2o 
(D~I ~4IS*) rr~y match all preceding frames denoting ~y kind of 
development° Since a CHANGE is a kind of DEVELOI~K~T, the definite phrase '~DIESE 
~NTWICKLUNG" (this development) is interpreted as referring to the already 
mentioned changes: the actually, happened SLI~D-DECRF~%SE (CHAh~-I) and the 
expected, but unfulfilled h-qCRFASE of the (global) NR-\[rN~v~LOYED (C~I~N~E-2)o This 
"reference by abstraction" is often found in r~e~,~spaper texts \[l~)senberg 1977\] o 
Next step: Can we infer any of these changes from BECAUSE-OF-I~s PROP-l? \[,bre 
specifically~ How may bKIN-INCRFASE of NR-\[rNEMPLOYED in OUTG~'~DR-JOBS e~,-plain 
CHANGE-I or CHANGE-2? 
Since $4 gave no contrary info_~aation, CHANGE-3'S TIb~-siot is filled with 
FEBRUARY which is the context default established by the preceeding sentences. 
If taken in isolation, NfLN-INCREASE for (XS'IDGOR-JOBS gives no direct ,~y to infer 
the Overall SLIGhT-DECREASE stated with CH~NGE-Io But: Since these two changes are 
given as facts and since outdoor-jobs are a subset of all jobs~ ~ co~clude~ that 
there must have been an interfering DECREASE in (an)other part(s) of the 
jobmarket, and therefore c~eate an expectation E1 for subsequent information of 
this kind, by the rule 
IF (91obal result is: DECRFASE) 
& (local change is- INCREASE or kAqN-DECRF~E) 
(expect: local DECRFI~E in other \[~rts),, 
With regards to CHA_NGE-2, we take a ~blo,:~-up" (using a shorthand notation) of our 
default expectation path betweec, winter weather and global increase: 
ARGUMENTATION STRUCTURES 1N NEWSPAPER TEXTS 329 
\[ 
!~ d.exp. 
~°°=> 
L 
INCREASE OF I d.exp. I I~ OF 
NR-UN~4P~ = °.. => NR-~ 
FOR OUTDbOR-JOBS GLOBAL 
G~>~e the dependency net is built in various levels of detail~ any "non-primitive" 
%aticn can be ~blown up~'~ i°e° be locked at in more detail° 
?~:~,chi~@ '/hrough possible paths relating WINTER ~ to expected INCRFASE NR- 
,. ~.~C:?LO~_ED Gh0BAL, the bl~e-up procedure selects the one with an intermediate node 
.",~~cerning C%HANGE-3'S QUANTITY, i.e. NR~LOYED for OU~IDOOR~° 
i '~ ;:~ in D~X3~SE-OF-I a N~'~-INCREASE for OUTDOOR-JOBS ~ NR-UN\]KMP~ is given as 
f~:~to %~his has two consequences: 
~.) The expectation path based o a ~NCREASE of this number is made invalid° This 
,in.ers question A2 still pending from ALTHOL~-I: Diny did PROP-2' fail? 
2) S.~-ce a,~ ,.~CR~SE of this number ~s expected by default~ we create a n£w 
,~?'~:~,,ah\[on task (that affects processing of subsequent input): K~nat are the 
._u. u~ for. C~GE~-3 (k~J~-~CREASE of NR-UN~PLOYED (DTDOOR-JOBS in FEBRUARY)? 
~; r.c\[E P.~URDkN ~;~q Dk~ 0t-~-~Y/~qD~qNLICH ~ D/IN'I~RS BEREITS IM JANUAR 
'~,~A~S~sq,"(Ph~gl:"Due to /he unusually strong winter they had been layed off 
$,1' eo~y in Ja,luary".) 
i?, ,.::J~qE-OF-2 
PI,;2-2: (!~Y-OY~-I G~UP: ~I~IS* 
'l'!".'~\]': * ('1"IV\] JAk~TAR) ~:) ) 
"~n O--C~e,~ to a-~-cept BEC~USE-OF-2~ we have to sh(~% that its PRDP-I leads to the 
e~cpectation of its P~P-2~ Searching through our d<mmin knowledge for relations 
betwaen ~%,~ ~nd LAY-O~'F of a GF~JP of persons gives: 
i• ~'~ I I £AY-OFF i MOD: ST~I~G 
==> o o. ==-> , GROUP: (IOTL~OR-WORKERS 
• L 
I** order to use t/,is genetic rule ~ should be able to resolve the missing refe- 
~er~ce from IAY-O~F-I's GROJP-slot (c.T~LIS*) with ~R-WORKERS. This choice is 
indeed supported bY ti~e fect that CFOTDOOR-JOBS are under discussion in CHANGE-3o 
Since CHANGE-3 (t/-~e NO~,~-ISq~R~I~E of k~-UNEMPLOYED of OUTDOOR-JOBS) is not yet 
e~.plained, we u~e an appropriate blo~-;,-up of dependencies again: 
--- ~ 0b~I~O:q ~ --==> OV±I~3OR- 
1 I CONSTRAI~C: 
I there are still 
~ outdoo~ -worker s 
f Ik~REA~E 
~=> NR-HNEMPL ° 
no new hiring~ 
of outdoor I 
~orkers _~ 
330 D.F. ROF.SN£R and J.H. LAUBSCh 
For JANOARY %~e have an instantiateL ~L~AUSE-OF-relaticn beh-~een ~rlT~TER and a 
stated IAY-OFF of ~A-WOR~So ~n inference rule for iAY-O~-~J is~ 
IF '-~.ere are layoffs (and no in~o~.~ation ab~J~= interfering nsw hirings) 
the number of ~rkers in the \[espzctive re~.:1on, branch etCo 6ecreaseSo 
As a consequence of the O~;~t7 I::~Y-~::~ ~ thus conclude, that the ~T}~E ARE 
STILL OOIDC~R-~?0~"-~?~NST£'RA\]:~ \[~ay no ionc3er hold i~ ~PCfo This invalidates 
the inference path above~ th~s su~ficienkiy ~=xplainin 9 C~iANGE-3o This in turn 
all(x.~s us to finish proc~.ssing of Bh~U~SE-O.?-!~ Recall: Question A3 from ALT~IOU~I- 
1 an~\] its subtask expectation E1 are still pending° 
S~: DARL~BER Hi~AUS \[(O~X%~E M.~tq I~i Z~',~IT~k ~ f;~ DES QUARTALS U-~LI~fE~'~EISE EiTqF/q 
RI/~f(C~%I~ DER ~RBEITSLOSIGXEIT I%\] DE~ ANGE~TZLL~BE~JF~N BEOBAC~TE~Wo (Englo : 
Mmreover one usually observes a decrease in une~toloyment of ~hite collar v~rkers 
in 'dhe second month of the quarter. ) 
(MOREOgE~-I 
PROP: ( CHAk~E- 4 
QUAk~ITY: (NR-U~,~LO~D 
SUB~-%~JP: ~ITE-C0!L~R-JOBS ) 
TIME:e(IM ZWEI%~X! M\[XNAT DES QUAKq'ALS):; 
VALUE: DECREAS~ 
M~O: *UEBLI~SE*) ) 
If a proposition is modified wit\]\] "D~BLiCb~:'~iS~ := (engi: by default) we process 
it as stating a fact. The abstract description for the TIME period matches 
FEBirUARY, ~hich is also the filler of the ~i~\[E-slot of the still unexplained 
~-io 
~he function of the PROP of a MOREOVER is to give additicnal information that 
helps answering open questions. 
In fact, CHAk~E-4'S information answers ALTHOUGH-I's question A3~ DECREASE of NR- 
UNEMPLOYED for ~X~I~-fDLIAR-JOBS in FEBRUARY matches expectation E1 (D~CREASE in 
l~3N~R-parts of the job-market). E1 on the other hand has been set up when 
processing BECAUSE-OF-I in locking for reasons for ALTHOUGH-I~s PNOP-2: the (up to 
this point unexplained) SLIGHT-DECRFASE of NR-UNEMPIK)YED GLOBAL. 

References

C~llir~ford,R.Eo Script Application: Computer Understanding of Newspaper Stories Jan. 1978, Yale University~ Dep. of Computer Sciencet Research Report 116 

De Jonq,G.F° Skir~ning Stories in Real Time: An Experiment in Integrated 
Understanding, May 1979, Yale University, Dep. of Compo Sc., Research Report 158 

Eise~stadt,M. Some Criteria for the Design of a Robust Newspaper Comprehender 
In: Int. Sem. on Intelligent Question-Answering and Data Base Systems, Bonas, 1977 

La~m~h,J.H. & l~mer,D.F. Active Schemata and their Role in Semantic Parsing 
In: OOLING '80, Proceedings, Tokyo 1980 

I~rg,St.T. Frame-based Text Processing, MIT-AI Memo 431, Cambridge, 1977 

S~.~m~an,G.J. & Steele,G.Lo CONSTRAINTS - A Language for Expressing Almost- 
Hierarchical Descriptions, ~xtlflcial Intelliger~e 14 (1980), pp. 1-39 
